; RUN: llc -O3 -march=hexagon -hexagon-small-data-threshold=0 -disable-packetizer < %s | FileCheck %s ; This test was orignally written to test that we don't save an entire double ; register if only one of the integer registers needs to be saved. The problem ; occurs in exception handling, which only emit information for the registers ; in the callee saved list (and not complete double registers unless both ; parts of the double registers are used). ; Overtime, we evolved in to saving the double register and updating the debug ; information to cover the entire double register. ; Disable the packetizer to avoid complications caused by potentially ; packetizing one of the stores with allocframe, which would change the ; relative order of the stores with the CFI instructions. ; CHECK: cfi_startproc ; CHECK-DAG: cfi_offset r16 ; CHECK-DAG: cfi_offset r17 ; CHECK-DAG: cfi_offset r18 ; CHECK-DAG: cfi_offset r19 ; CHECK: memd(r29+{{.*}}) = r17:16 ; CHECK: memd(r29+{{.*}}) = r19:18 %s.0 = type { i32 } @g0 = global i32 0, align 4 @g1 = external constant i8* ; Function Attrs: noreturn define void @f0(i64 %a0) #0 personality i8* bitcast (i32 (...)* @f2 to i8*) { b0: %v0 = alloca %s.0, align 4 %v1 = trunc i64 %a0 to i32 %v2 = lshr i64 %a0, 32 %v3 = trunc i64 %v2 to i32 %v4 = getelementptr inbounds %s.0, %s.0* %v0, i32 0, i32 0 store i32 0, i32* %v4, align 4, !tbaa !0 %v5 = load i32, i32* @g0, align 4, !tbaa !5 %v6 = or i32 %v5, 1 store i32 %v6, i32* @g0, align 4, !tbaa !5 %v7 = call i8* @f1(i32 4) #1 %v8 = bitcast i8* %v7 to i32* %v9 = bitcast %s.0* %v0 to i8* %v10 = getelementptr inbounds i8, i8* %v9, i32 %v3 %v11 = bitcast i8* %v10 to %s.0* %v12 = and i32 %v1, 1 %v13 = icmp eq i32 %v12, 0 br i1 %v13, label %b2, label %b1 b1: ; preds = %b0 %v14 = bitcast i8* %v10 to i8** %v15 = load i8*, i8** %v14, align 4 %v16 = add i32 %v1, -1 %v17 = getelementptr i8, i8* %v15, i32 %v16 %v18 = bitcast i8* %v17 to i32 (%s.0*)** %v19 = load i32 (%s.0*)*, i32 (%s.0*)** %v18, align 4 br label %b3 b2: ; preds = %b0 %v20 = inttoptr i32 %v1 to i32 (%s.0*)* br label %b3 b3: ; preds = %b2, %b1 %v21 = phi i32 (%s.0*)* [ %v19, %b1 ], [ %v20, %b2 ] %v22 = invoke i32 %v21(%s.0* %v11) to label %b4 unwind label %b5 b4: ; preds = %b3 store i32 %v22, i32* %v8, align 4, !tbaa !5 call void @f4(i8* %v7, i8* bitcast (i8** @g1 to i8*), i8* null) #2 unreachable b5: ; preds = %b3 %v23 = landingpad { i8*, i32 } cleanup call void @f3(i8* %v7) #1 resume { i8*, i32 } %v23 } declare i8* @f1(i32) declare i32 @f2(...) declare void @f3(i8*) declare void @f4(i8*, i8*, i8*) attributes #0 = { noreturn "target-cpu"="hexagonv55" } attributes #1 = { nounwind } attributes #2 = { noreturn } !0 = !{!1, !2, i64 0} !1 = !{!"_ZTS1A", !2, i64 0} !2 = !{!"int", !3, i64 0} !3 = !{!"omnipotent char", !4, i64 0} !4 = !{!"Simple C/C++ TBAA"} !5 = !{!2, !2, i64 0}