6049 lines
225 KiB
LLVM
6049 lines
225 KiB
LLVM
; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
|
|
; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+sse2 -O3 | FileCheck %s --check-prefixes=SSE-32
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+sse2 -O3 | FileCheck %s --check-prefixes=SSE-64
|
|
; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx -O3 | FileCheck %s --check-prefixes=AVX-32
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx -O3 | FileCheck %s --check-prefixes=AVX-64
|
|
; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512f -mattr=+avx512vl -O3 | FileCheck %s --check-prefixes=AVX512-32
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f -mattr=+avx512vl -O3 | FileCheck %s --check-prefixes=AVX512-64
|
|
; RUN: llc < %s -mtriple=i686-unknown-unknown -mattr=+avx512f -O3 | FileCheck %s --check-prefixes=AVX512F-32
|
|
; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f -O3 | FileCheck %s --check-prefixes=AVX512F-64
|
|
|
|
define <4 x i32> @test_v4f32_oeq_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_oeq_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpeqps 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_oeq_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpeqps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_oeq_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpeqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_oeq_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpeqps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_oeq_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpeqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_oeq_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpeqps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_oeq_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpeqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_oeq_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpeqps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"oeq",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ogt_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ogt_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm3[3,3]
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm2[3,3]
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmoval %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm3[1]
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm6
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm2[1]
|
|
; SSE-32-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmoval %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm5
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-32-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmoval %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-32-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-32-NEXT: cmoval %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ogt_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm3[3,3]
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm2[3,3]
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-64-NEXT: movl $-1, %ecx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmoval %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm3[1]
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm6
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm2[1]
|
|
; SSE-64-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmoval %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm5
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-64-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmoval %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-64-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-64-NEXT: cmoval %ecx, %eax
|
|
; SSE-64-NEXT: movd %eax, %xmm2
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ogt_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmplt_oqps %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ogt_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmplt_oqps %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ogt_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpgt_oqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ogt_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmplt_oqps %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ogt_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmplt_oqps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ogt_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmplt_oqps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ogt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_oge_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_oge_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm3[3,3]
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm2[3,3]
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovael %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm3[1]
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm6
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm2[1]
|
|
; SSE-32-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovael %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm5
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-32-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovael %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-32-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-32-NEXT: cmovael %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_oge_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm3[3,3]
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm2[3,3]
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-64-NEXT: movl $-1, %ecx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovael %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm3[1]
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm6
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm2[1]
|
|
; SSE-64-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovael %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm5
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-64-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovael %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-64-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-64-NEXT: cmovael %ecx, %eax
|
|
; SSE-64-NEXT: movd %eax, %xmm2
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_oge_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmple_oqps %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_oge_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmple_oqps %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_oge_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpge_oqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_oge_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmple_oqps %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_oge_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmple_oqps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_oge_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmple_oqps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"oge",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_olt_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_olt_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm2[3,3]
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm3[3,3]
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmoval %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm2[1]
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm6
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm3[1]
|
|
; SSE-32-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmoval %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm5
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-32-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmoval %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-32-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-32-NEXT: cmoval %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_olt_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm2[3,3]
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm3[3,3]
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-64-NEXT: movl $-1, %ecx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmoval %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm2[1]
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm6
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm3[1]
|
|
; SSE-64-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmoval %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm5
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-64-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmoval %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-64-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-64-NEXT: cmoval %ecx, %eax
|
|
; SSE-64-NEXT: movd %eax, %xmm2
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_olt_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmplt_oqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_olt_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmplt_oqps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_olt_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmplt_oqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_olt_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmplt_oqps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_olt_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmplt_oqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_olt_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmplt_oqps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"olt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ole_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ole_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm2[3,3]
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm3[3,3]
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovael %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm2[1]
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm6
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm3[1]
|
|
; SSE-32-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovael %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm5
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-32-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovael %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-32-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-32-NEXT: cmovael %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ole_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm2[3,3]
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm3[3,3]
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-64-NEXT: movl $-1, %ecx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovael %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm2[1]
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm6
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm3[1]
|
|
; SSE-64-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovael %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm5
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-64-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovael %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-64-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-64-NEXT: cmovael %ecx, %eax
|
|
; SSE-64-NEXT: movd %eax, %xmm2
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ole_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmple_oqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ole_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmple_oqps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ole_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmple_oqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ole_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmple_oqps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ole_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmple_oqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ole_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmple_oqps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ole",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_one_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_one_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpneqps %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpordps %xmm3, %xmm2
|
|
; SSE-32-NEXT: andps %xmm4, %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_one_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpneqps %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpordps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm4, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_one_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpneq_oqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_one_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpneq_oqps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_one_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpneq_oqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_one_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpneq_oqps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_one_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpneq_oqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_one_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpneq_oqps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"one",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ord_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ord_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpordps 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ord_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpordps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ord_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpordps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ord_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpordps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ord_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpordps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ord_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpordps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ord_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpordps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ord_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpordps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ord",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ueq_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ueq_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpeqps %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpunordps %xmm3, %xmm2
|
|
; SSE-32-NEXT: orps %xmm4, %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ueq_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpeqps %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpunordps %xmm3, %xmm2
|
|
; SSE-64-NEXT: orps %xmm4, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ueq_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpeq_uqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ueq_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpeq_uqps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ueq_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpeq_uqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ueq_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpeq_uqps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ueq_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpeq_uqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ueq_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpeq_uqps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ueq",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ugt_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ugt_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm2[3,3]
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm3[3,3]
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbl %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm2[1]
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm6
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm3[1]
|
|
; SSE-32-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbl %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm5
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-32-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbl %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-32-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-32-NEXT: cmovbl %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ugt_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm2[3,3]
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm3[3,3]
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-64-NEXT: movl $-1, %ecx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbl %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm2[1]
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm6
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm3[1]
|
|
; SSE-64-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbl %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm5
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-64-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbl %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-64-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-64-NEXT: cmovbl %ecx, %eax
|
|
; SSE-64-NEXT: movd %eax, %xmm2
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ugt_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpnle_uqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ugt_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnle_uqps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ugt_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpnle_uqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ugt_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnle_uqps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ugt_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpnle_uqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ugt_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnle_uqps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ugt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_uge_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_uge_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm2[3,3]
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm3[3,3]
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbel %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm2[1]
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm6
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm3[1]
|
|
; SSE-32-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbel %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm5
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-32-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbel %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-32-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-32-NEXT: cmovbel %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_uge_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm2[3,3]
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm3[3,3]
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-64-NEXT: movl $-1, %ecx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbel %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm2[1]
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm6
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm3[1]
|
|
; SSE-64-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbel %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm5
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-64-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbel %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-64-NEXT: ucomiss %xmm2, %xmm3
|
|
; SSE-64-NEXT: cmovbel %ecx, %eax
|
|
; SSE-64-NEXT: movd %eax, %xmm2
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_uge_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpnlt_uqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_uge_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnlt_uqps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_uge_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpnlt_uqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_uge_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnlt_uqps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_uge_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpnlt_uqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_uge_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnlt_uqps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"uge",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ult_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ult_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm3[3,3]
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm2[3,3]
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbl %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm3[1]
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm6
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm2[1]
|
|
; SSE-32-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbl %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm5
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-32-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbl %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-32-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-32-NEXT: cmovbl %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ult_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm3[3,3]
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm2[3,3]
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-64-NEXT: movl $-1, %ecx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbl %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm3[1]
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm6
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm2[1]
|
|
; SSE-64-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbl %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm5
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-64-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbl %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-64-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-64-NEXT: cmovbl %ecx, %eax
|
|
; SSE-64-NEXT: movd %eax, %xmm2
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ult_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmpnle_uqps %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ult_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnle_uqps %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ult_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpnge_uqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ult_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnle_uqps %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ult_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmpnle_uqps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ult_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnle_uqps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ult",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ule_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ule_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm3[3,3]
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm2[3,3]
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbel %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm3[1]
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm6
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm2[1]
|
|
; SSE-32-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbel %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm5
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-32-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbel %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-32-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-32-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-32-NEXT: cmovbel %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ule_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm4 = xmm4[3,3],xmm3[3,3]
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm5
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm5 = xmm5[3,3],xmm2[3,3]
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomiss %xmm4, %xmm5
|
|
; SSE-64-NEXT: movl $-1, %ecx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbel %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: movaps %xmm3, %xmm5
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm5 = xmm5[1],xmm3[1]
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm6
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm6 = xmm6[1],xmm2[1]
|
|
; SSE-64-NEXT: ucomiss %xmm5, %xmm6
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbel %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm5
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm5 = xmm5[0],xmm4[0],xmm5[1],xmm4[1]
|
|
; SSE-64-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbel %ecx, %edx
|
|
; SSE-64-NEXT: movd %edx, %xmm4
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm3 = xmm3[1,1,1,1]
|
|
; SSE-64-NEXT: shufps {{.*#+}} xmm2 = xmm2[1,1,1,1]
|
|
; SSE-64-NEXT: ucomiss %xmm3, %xmm2
|
|
; SSE-64-NEXT: cmovbel %ecx, %eax
|
|
; SSE-64-NEXT: movd %eax, %xmm2
|
|
; SSE-64-NEXT: punpckldq {{.*#+}} xmm4 = xmm4[0],xmm2[0],xmm4[1],xmm2[1]
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm5[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ule_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmpnlt_uqps %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ule_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnlt_uqps %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ule_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpngt_uqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ule_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnlt_uqps %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ule_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmpnlt_uqps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ule_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnlt_uqps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ule",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_une_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_une_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpneqps 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_une_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpneqps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_une_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpneqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_une_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpneqps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_une_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpneqps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_une_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpneqps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_une_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpneqps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_une_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpneqps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"une",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_uno_q(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_uno_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpunordps 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_uno_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpunordps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_uno_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpunordps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_uno_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpunordps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_uno_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpunordps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_uno_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpunordps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_uno_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpunordps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_uno_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpunordps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"uno",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_oeq_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_oeq_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpeqpd 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_oeq_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpeqpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_oeq_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpeqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_oeq_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpeqpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_oeq_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpeqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_oeq_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpeqpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_oeq_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpeqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_oeq_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpeqpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"oeq",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ogt_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ogt_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmoval %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-32-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-32-NEXT: cmoval %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ogt_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-64-NEXT: movq $-1, %rcx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovaq %rcx, %rdx
|
|
; SSE-64-NEXT: movq %rdx, %xmm4
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-64-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-64-NEXT: cmovaq %rcx, %rax
|
|
; SSE-64-NEXT: movq %rax, %xmm2
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ogt_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmplt_oqpd %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ogt_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmplt_oqpd %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ogt_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpgt_oqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ogt_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmplt_oqpd %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ogt_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmplt_oqpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ogt_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmplt_oqpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ogt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_oge_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_oge_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovael %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-32-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-32-NEXT: cmovael %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_oge_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-64-NEXT: movq $-1, %rcx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovaeq %rcx, %rdx
|
|
; SSE-64-NEXT: movq %rdx, %xmm4
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-64-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-64-NEXT: cmovaeq %rcx, %rax
|
|
; SSE-64-NEXT: movq %rax, %xmm2
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_oge_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmple_oqpd %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_oge_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmple_oqpd %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_oge_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpge_oqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_oge_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmple_oqpd %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_oge_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmple_oqpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_oge_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmple_oqpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"oge",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_olt_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_olt_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmoval %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-32-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-32-NEXT: cmoval %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_olt_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-64-NEXT: movq $-1, %rcx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovaq %rcx, %rdx
|
|
; SSE-64-NEXT: movq %rdx, %xmm4
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-64-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-64-NEXT: cmovaq %rcx, %rax
|
|
; SSE-64-NEXT: movq %rax, %xmm2
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_olt_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmplt_oqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_olt_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmplt_oqpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_olt_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmplt_oqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_olt_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmplt_oqpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_olt_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmplt_oqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_olt_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmplt_oqpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"olt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ole_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ole_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovael %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-32-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-32-NEXT: cmovael %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ole_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-64-NEXT: movq $-1, %rcx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovaeq %rcx, %rdx
|
|
; SSE-64-NEXT: movq %rdx, %xmm4
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-64-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-64-NEXT: cmovaeq %rcx, %rax
|
|
; SSE-64-NEXT: movq %rax, %xmm2
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ole_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmple_oqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ole_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmple_oqpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ole_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmple_oqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ole_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmple_oqpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ole_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmple_oqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ole_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmple_oqpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ole",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_one_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_one_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpneqpd %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpordpd %xmm3, %xmm2
|
|
; SSE-32-NEXT: andpd %xmm4, %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_one_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpneqpd %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpordpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm4, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_one_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpneq_oqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_one_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpneq_oqpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_one_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpneq_oqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_one_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpneq_oqpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_one_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpneq_oqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_one_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpneq_oqpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"one",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ord_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ord_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpordpd 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ord_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpordpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ord_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpordpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ord_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpordpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ord_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpordpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ord_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpordpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ord_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpordpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ord_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpordpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ord",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ueq_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ueq_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpeqpd %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpunordpd %xmm3, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm4, %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ueq_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpeqpd %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpunordpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm4, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ueq_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpeq_uqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ueq_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpeq_uqpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ueq_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpeq_uqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ueq_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpeq_uqpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ueq_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpeq_uqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ueq_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpeq_uqpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ueq",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ugt_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ugt_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbl %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-32-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-32-NEXT: cmovbl %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ugt_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-64-NEXT: movq $-1, %rcx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbq %rcx, %rdx
|
|
; SSE-64-NEXT: movq %rdx, %xmm4
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-64-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-64-NEXT: cmovbq %rcx, %rax
|
|
; SSE-64-NEXT: movq %rax, %xmm2
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ugt_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpnle_uqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ugt_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnle_uqpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ugt_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpnle_uqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ugt_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnle_uqpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ugt_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpnle_uqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ugt_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnle_uqpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ugt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_uge_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_uge_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbel %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-32-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-32-NEXT: cmovbel %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_uge_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-64-NEXT: movq $-1, %rcx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbeq %rcx, %rdx
|
|
; SSE-64-NEXT: movq %rdx, %xmm4
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-64-NEXT: ucomisd %xmm2, %xmm3
|
|
; SSE-64-NEXT: cmovbeq %rcx, %rax
|
|
; SSE-64-NEXT: movq %rax, %xmm2
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_uge_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpnlt_uqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_uge_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnlt_uqpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_uge_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpnlt_uqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_uge_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnlt_uqpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_uge_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpnlt_uqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_uge_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnlt_uqpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"uge",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ult_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ult_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbl %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-32-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-32-NEXT: cmovbl %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ult_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-64-NEXT: movq $-1, %rcx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbq %rcx, %rdx
|
|
; SSE-64-NEXT: movq %rdx, %xmm4
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-64-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-64-NEXT: cmovbq %rcx, %rax
|
|
; SSE-64-NEXT: movq %rax, %xmm2
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ult_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmpnle_uqpd %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ult_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnle_uqpd %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ult_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpnge_uqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ult_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnle_uqpd %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ult_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmpnle_uqpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ult_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnle_uqpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ult",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ule_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ule_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: xorl %eax, %eax
|
|
; SSE-32-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-32-NEXT: movl $-1, %ecx
|
|
; SSE-32-NEXT: movl $0, %edx
|
|
; SSE-32-NEXT: cmovbel %ecx, %edx
|
|
; SSE-32-NEXT: movd %edx, %xmm4
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm4 = xmm4[0,0,1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-32-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-32-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-32-NEXT: cmovbel %ecx, %eax
|
|
; SSE-32-NEXT: movd %eax, %xmm2
|
|
; SSE-32-NEXT: pshufd {{.*#+}} xmm2 = xmm2[0,0,1,1]
|
|
; SSE-32-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-32-NEXT: pand %xmm4, %xmm0
|
|
; SSE-32-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-32-NEXT: por %xmm4, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ule_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: xorl %eax, %eax
|
|
; SSE-64-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-64-NEXT: movq $-1, %rcx
|
|
; SSE-64-NEXT: movl $0, %edx
|
|
; SSE-64-NEXT: cmovbeq %rcx, %rdx
|
|
; SSE-64-NEXT: movq %rdx, %xmm4
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm3 = xmm3[1,1]
|
|
; SSE-64-NEXT: unpckhpd {{.*#+}} xmm2 = xmm2[1,1]
|
|
; SSE-64-NEXT: ucomisd %xmm3, %xmm2
|
|
; SSE-64-NEXT: cmovbeq %rcx, %rax
|
|
; SSE-64-NEXT: movq %rax, %xmm2
|
|
; SSE-64-NEXT: punpcklqdq {{.*#+}} xmm4 = xmm4[0],xmm2[0]
|
|
; SSE-64-NEXT: pand %xmm4, %xmm0
|
|
; SSE-64-NEXT: pandn %xmm1, %xmm4
|
|
; SSE-64-NEXT: por %xmm4, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ule_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmpnlt_uqpd %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ule_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnlt_uqpd %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ule_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpngt_uqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ule_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnlt_uqpd %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ule_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmpnlt_uqpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ule_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnlt_uqpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ule",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_une_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_une_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpneqpd 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_une_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpneqpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_une_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpneqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_une_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpneqpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_une_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpneqpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_une_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpneqpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_une_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpneqpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_une_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpneqpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"une",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_uno_q(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_uno_q:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpunordpd 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_uno_q:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpunordpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_uno_q:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpunordpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_uno_q:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpunordpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_uno_q:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpunordpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_uno_q:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpunordpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_uno_q:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpunordpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_uno_q:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpunordpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"uno",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_oeq_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_oeq_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpeqps %xmm3, %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_oeq_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpeqps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_oeq_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpeq_osps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_oeq_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpeq_osps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_oeq_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpeq_osps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_oeq_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpeq_osps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_oeq_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpeq_osps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_oeq_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpeq_osps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"oeq",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ogt_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ogt_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: cmpltps %xmm2, %xmm3
|
|
; SSE-32-NEXT: andps %xmm3, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm3
|
|
; SSE-32-NEXT: orps %xmm3, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ogt_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpltps %xmm2, %xmm3
|
|
; SSE-64-NEXT: andps %xmm3, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm3
|
|
; SSE-64-NEXT: orps %xmm3, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ogt_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmpltps %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ogt_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpltps %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ogt_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpgtps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ogt_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpltps %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ogt_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmpltps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ogt_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpltps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ogt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_oge_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_oge_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: cmpleps %xmm2, %xmm3
|
|
; SSE-32-NEXT: andps %xmm3, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm3
|
|
; SSE-32-NEXT: orps %xmm3, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_oge_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpleps %xmm2, %xmm3
|
|
; SSE-64-NEXT: andps %xmm3, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm3
|
|
; SSE-64-NEXT: orps %xmm3, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_oge_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmpleps %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_oge_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpleps %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_oge_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpgeps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_oge_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpleps %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_oge_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmpleps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_oge_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpleps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"oge",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_olt_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_olt_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpltps 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_olt_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpltps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_olt_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpltps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_olt_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpltps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_olt_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpltps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_olt_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpltps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_olt_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpltps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_olt_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpltps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"olt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ole_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ole_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpleps 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ole_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpleps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ole_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpleps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ole_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpleps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ole_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpleps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ole_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpleps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ole_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpleps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ole_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpleps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ole",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_one_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_one_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpneqps %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpordps %xmm3, %xmm2
|
|
; SSE-32-NEXT: andps %xmm4, %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_one_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpneqps %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpordps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm4, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_one_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpneq_osps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_one_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpneq_osps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_one_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpneq_osps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_one_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpneq_osps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_one_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpneq_osps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_one_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpneq_osps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"one",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ord_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ord_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpordps %xmm3, %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ord_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpordps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ord_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpord_sps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ord_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpord_sps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ord_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpord_sps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ord_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpord_sps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ord_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpord_sps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ord_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpord_sps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ord",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ueq_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ueq_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpeqps %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpunordps %xmm3, %xmm2
|
|
; SSE-32-NEXT: orps %xmm4, %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ueq_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpeqps %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpunordps %xmm3, %xmm2
|
|
; SSE-64-NEXT: orps %xmm4, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ueq_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpeq_usps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ueq_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpeq_usps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ueq_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpeq_usps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ueq_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpeq_usps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ueq_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpeq_usps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ueq_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpeq_usps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ueq",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ugt_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ugt_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpnleps 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ugt_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpnleps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ugt_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpnleps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ugt_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnleps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ugt_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpnleps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ugt_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnleps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ugt_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpnleps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ugt_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnleps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ugt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_uge_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_uge_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpnltps 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_uge_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpnltps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_uge_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpnltps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_uge_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnltps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_uge_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpnltps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_uge_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnltps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_uge_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpnltps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_uge_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnltps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"uge",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ult_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ult_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: cmpnleps %xmm2, %xmm3
|
|
; SSE-32-NEXT: andps %xmm3, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm3
|
|
; SSE-32-NEXT: orps %xmm3, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ult_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpnleps %xmm2, %xmm3
|
|
; SSE-64-NEXT: andps %xmm3, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm3
|
|
; SSE-64-NEXT: orps %xmm3, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ult_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmpnleps %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ult_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnleps %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ult_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpngeps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ult_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnleps %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ult_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmpnleps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ult_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnleps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ult",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_ule_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_ule_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: cmpnltps %xmm2, %xmm3
|
|
; SSE-32-NEXT: andps %xmm3, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm3
|
|
; SSE-32-NEXT: orps %xmm3, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_ule_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpnltps %xmm2, %xmm3
|
|
; SSE-64-NEXT: andps %xmm3, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm3
|
|
; SSE-64-NEXT: orps %xmm3, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_ule_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmpnltps %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_ule_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnltps %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_ule_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpngtps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_ule_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnltps %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_ule_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovaps 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmpnltps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_ule_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnltps %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"ule",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_une_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_une_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpneqps %xmm3, %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_une_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpneqps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_une_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpneq_usps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_une_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpneq_usps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_une_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpneq_usps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_une_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpneq_usps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_une_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpneq_usps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_une_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpneq_usps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"une",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <4 x i32> @test_v4f32_uno_s(<4 x i32> %a, <4 x i32> %b, <4 x float> %f1, <4 x float> %f2) #0 {
|
|
; SSE-32-LABEL: test_v4f32_uno_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movaps 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpunordps %xmm3, %xmm2
|
|
; SSE-32-NEXT: andps %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-32-NEXT: orps %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v4f32_uno_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movaps %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltps %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpunordps %xmm3, %xmm2
|
|
; SSE-64-NEXT: andps %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnps %xmm1, %xmm2
|
|
; SSE-64-NEXT: orps %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v4f32_uno_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpunord_sps 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v4f32_uno_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpunord_sps %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvps %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v4f32_uno_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpunord_sps 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v4f32_uno_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpunord_sps %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmd %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v4f32_uno_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpunord_sps 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v4f32_uno_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpunord_sps %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmd %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmd %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(
|
|
<4 x float> %f1, <4 x float> %f2, metadata !"uno",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <4 x i1> %cond, <4 x i32> %a, <4 x i32> %b
|
|
ret <4 x i32> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_oeq_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_oeq_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpeqpd %xmm3, %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_oeq_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpeqpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_oeq_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpeq_ospd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_oeq_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpeq_ospd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_oeq_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpeq_ospd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_oeq_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpeq_ospd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_oeq_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpeq_ospd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_oeq_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpeq_ospd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"oeq",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ogt_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ogt_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: cmpltpd %xmm2, %xmm3
|
|
; SSE-32-NEXT: andpd %xmm3, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm3
|
|
; SSE-32-NEXT: orpd %xmm3, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ogt_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpltpd %xmm2, %xmm3
|
|
; SSE-64-NEXT: andpd %xmm3, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm3
|
|
; SSE-64-NEXT: orpd %xmm3, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ogt_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmpltpd %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ogt_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpltpd %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ogt_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpgtpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ogt_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpltpd %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ogt_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmpltpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ogt_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpltpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ogt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_oge_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_oge_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: cmplepd %xmm2, %xmm3
|
|
; SSE-32-NEXT: andpd %xmm3, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm3
|
|
; SSE-32-NEXT: orpd %xmm3, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_oge_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmplepd %xmm2, %xmm3
|
|
; SSE-64-NEXT: andpd %xmm3, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm3
|
|
; SSE-64-NEXT: orpd %xmm3, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_oge_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmplepd %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_oge_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmplepd %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_oge_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpgepd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_oge_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmplepd %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_oge_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmplepd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_oge_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmplepd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"oge",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_olt_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_olt_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpltpd 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_olt_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpltpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_olt_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpltpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_olt_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpltpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_olt_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpltpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_olt_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpltpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_olt_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpltpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_olt_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpltpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"olt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ole_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ole_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmplepd 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ole_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmplepd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ole_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmplepd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ole_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmplepd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ole_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmplepd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ole_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmplepd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ole_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmplepd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ole_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmplepd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ole",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_one_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_one_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-32-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpneqpd %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpordpd %xmm3, %xmm2
|
|
; SSE-32-NEXT: andpd %xmm4, %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_one_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-64-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpneqpd %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpordpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm4, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_one_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpneq_ospd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_one_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpneq_ospd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_one_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpneq_ospd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_one_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpneq_ospd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_one_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpneq_ospd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_one_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpneq_ospd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"one",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ord_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ord_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpordpd %xmm3, %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ord_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpordpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ord_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpord_spd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ord_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpord_spd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ord_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpord_spd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ord_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpord_spd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ord_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpord_spd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ord_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpord_spd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ord",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ueq_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ueq_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-32-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpeqpd %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpunordpd %xmm3, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm4, %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ueq_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-64-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpeqpd %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpunordpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm4, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ueq_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpeq_uspd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ueq_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpeq_uspd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ueq_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpeq_uspd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ueq_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpeq_uspd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ueq_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpeq_uspd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ueq_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpeq_uspd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ueq",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ugt_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ugt_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpnlepd 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ugt_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpnlepd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ugt_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpnlepd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ugt_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnlepd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ugt_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpnlepd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ugt_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnlepd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ugt_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpnlepd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ugt_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnlepd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ugt",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_uge_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_uge_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: cmpnltpd 8(%ebp), %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_uge_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpnltpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_uge_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpnltpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_uge_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnltpd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_uge_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpnltpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_uge_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnltpd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_uge_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpnltpd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_uge_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnltpd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"uge",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ult_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ult_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: cmpnlepd %xmm2, %xmm3
|
|
; SSE-32-NEXT: andpd %xmm3, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm3
|
|
; SSE-32-NEXT: orpd %xmm3, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ult_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpnlepd %xmm2, %xmm3
|
|
; SSE-64-NEXT: andpd %xmm3, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm3
|
|
; SSE-64-NEXT: orpd %xmm3, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ult_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmpnlepd %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ult_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnlepd %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ult_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpngepd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ult_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnlepd %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ult_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmpnlepd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ult_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnlepd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ult",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_ule_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_ule_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: cmpnltpd %xmm2, %xmm3
|
|
; SSE-32-NEXT: andpd %xmm3, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm3
|
|
; SSE-32-NEXT: orpd %xmm3, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_ule_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: cmpnltpd %xmm2, %xmm3
|
|
; SSE-64-NEXT: andpd %xmm3, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm3
|
|
; SSE-64-NEXT: orpd %xmm3, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_ule_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX-32-NEXT: vcmpnltpd %xmm2, %xmm3, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_ule_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpnltpd %xmm2, %xmm3, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_ule_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpngtpd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_ule_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpnltpd %xmm2, %xmm3, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_ule_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vmovapd 8(%ebp), %xmm3
|
|
; AVX512F-32-NEXT: vcmpnltpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_ule_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpnltpd %xmm2, %xmm3, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"ule",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_une_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_une_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpneqpd %xmm3, %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_une_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpneqpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_une_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpneq_uspd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_une_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpneq_uspd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_une_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpneq_uspd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_une_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpneq_uspd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_une_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpneq_uspd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_une_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpneq_uspd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"une",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
define <2 x i64> @test_v2f64_uno_s(<2 x i64> %a, <2 x i64> %b, <2 x double> %f1, <2 x double> %f2) #0 {
|
|
; SSE-32-LABEL: test_v2f64_uno_s:
|
|
; SSE-32: # %bb.0:
|
|
; SSE-32-NEXT: pushl %ebp
|
|
; SSE-32-NEXT: movl %esp, %ebp
|
|
; SSE-32-NEXT: andl $-16, %esp
|
|
; SSE-32-NEXT: subl $16, %esp
|
|
; SSE-32-NEXT: movapd 8(%ebp), %xmm3
|
|
; SSE-32-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-32-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-32-NEXT: cmpunordpd %xmm3, %xmm2
|
|
; SSE-32-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-32-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-32-NEXT: movl %ebp, %esp
|
|
; SSE-32-NEXT: popl %ebp
|
|
; SSE-32-NEXT: retl
|
|
;
|
|
; SSE-64-LABEL: test_v2f64_uno_s:
|
|
; SSE-64: # %bb.0:
|
|
; SSE-64-NEXT: movapd %xmm2, %xmm4
|
|
; SSE-64-NEXT: cmpltpd %xmm3, %xmm4
|
|
; SSE-64-NEXT: cmpunordpd %xmm3, %xmm2
|
|
; SSE-64-NEXT: andpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: andnpd %xmm1, %xmm2
|
|
; SSE-64-NEXT: orpd %xmm2, %xmm0
|
|
; SSE-64-NEXT: retq
|
|
;
|
|
; AVX-32-LABEL: test_v2f64_uno_s:
|
|
; AVX-32: # %bb.0:
|
|
; AVX-32-NEXT: pushl %ebp
|
|
; AVX-32-NEXT: movl %esp, %ebp
|
|
; AVX-32-NEXT: andl $-16, %esp
|
|
; AVX-32-NEXT: subl $16, %esp
|
|
; AVX-32-NEXT: vcmpunord_spd 8(%ebp), %xmm2, %xmm2
|
|
; AVX-32-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-32-NEXT: movl %ebp, %esp
|
|
; AVX-32-NEXT: popl %ebp
|
|
; AVX-32-NEXT: retl
|
|
;
|
|
; AVX-64-LABEL: test_v2f64_uno_s:
|
|
; AVX-64: # %bb.0:
|
|
; AVX-64-NEXT: vcmpunord_spd %xmm3, %xmm2, %xmm2
|
|
; AVX-64-NEXT: vblendvpd %xmm2, %xmm0, %xmm1, %xmm0
|
|
; AVX-64-NEXT: retq
|
|
;
|
|
; AVX512-32-LABEL: test_v2f64_uno_s:
|
|
; AVX512-32: # %bb.0:
|
|
; AVX512-32-NEXT: pushl %ebp
|
|
; AVX512-32-NEXT: movl %esp, %ebp
|
|
; AVX512-32-NEXT: andl $-16, %esp
|
|
; AVX512-32-NEXT: subl $16, %esp
|
|
; AVX512-32-NEXT: vcmpunord_spd 8(%ebp), %xmm2, %k1
|
|
; AVX512-32-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-32-NEXT: movl %ebp, %esp
|
|
; AVX512-32-NEXT: popl %ebp
|
|
; AVX512-32-NEXT: retl
|
|
;
|
|
; AVX512-64-LABEL: test_v2f64_uno_s:
|
|
; AVX512-64: # %bb.0:
|
|
; AVX512-64-NEXT: vcmpunord_spd %xmm3, %xmm2, %k1
|
|
; AVX512-64-NEXT: vpblendmq %xmm0, %xmm1, %xmm0 {%k1}
|
|
; AVX512-64-NEXT: retq
|
|
;
|
|
; AVX512F-32-LABEL: test_v2f64_uno_s:
|
|
; AVX512F-32: # %bb.0:
|
|
; AVX512F-32-NEXT: pushl %ebp
|
|
; AVX512F-32-NEXT: movl %esp, %ebp
|
|
; AVX512F-32-NEXT: andl $-16, %esp
|
|
; AVX512F-32-NEXT: subl $16, %esp
|
|
; AVX512F-32-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-32-NEXT: vcmpunord_spd 8(%ebp), %xmm2, %xmm2
|
|
; AVX512F-32-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-32-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-32-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-32-NEXT: movl %ebp, %esp
|
|
; AVX512F-32-NEXT: popl %ebp
|
|
; AVX512F-32-NEXT: vzeroupper
|
|
; AVX512F-32-NEXT: retl
|
|
;
|
|
; AVX512F-64-LABEL: test_v2f64_uno_s:
|
|
; AVX512F-64: # %bb.0:
|
|
; AVX512F-64-NEXT: # kill: def $xmm1 killed $xmm1 def $zmm1
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 def $zmm0
|
|
; AVX512F-64-NEXT: vcmpunord_spd %xmm3, %xmm2, %xmm2
|
|
; AVX512F-64-NEXT: vptestmq %zmm2, %zmm2, %k1
|
|
; AVX512F-64-NEXT: vpblendmq %zmm0, %zmm1, %zmm0 {%k1}
|
|
; AVX512F-64-NEXT: # kill: def $xmm0 killed $xmm0 killed $zmm0
|
|
; AVX512F-64-NEXT: vzeroupper
|
|
; AVX512F-64-NEXT: retq
|
|
%cond = call <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(
|
|
<2 x double> %f1, <2 x double> %f2, metadata !"uno",
|
|
metadata !"fpexcept.strict") #0
|
|
%res = select <2 x i1> %cond, <2 x i64> %a, <2 x i64> %b
|
|
ret <2 x i64> %res
|
|
}
|
|
|
|
attributes #0 = { strictfp nounwind }
|
|
|
|
declare <4 x i1> @llvm.experimental.constrained.fcmp.v4f32(<4 x float>, <4 x float>, metadata, metadata)
|
|
declare <2 x i1> @llvm.experimental.constrained.fcmp.v2f64(<2 x double>, <2 x double>, metadata, metadata)
|
|
declare <4 x i1> @llvm.experimental.constrained.fcmps.v4f32(<4 x float>, <4 x float>, metadata, metadata)
|
|
declare <2 x i1> @llvm.experimental.constrained.fcmps.v2f64(<2 x double>, <2 x double>, metadata, metadata)
|