tcg/mips: Split out tcg_out_setcond_int

Return the temp and a set of flags, to be used as a
primitive for setcond, brcond, movcond.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>
Message-Id: <20231026041404.1229328-2-richard.henderson@linaro.org>
This commit is contained in:
Richard Henderson 2023-10-25 21:13:59 -07:00
parent 58b797130c
commit 42221a64da

View File

@ -871,71 +871,83 @@ static void tcg_out_addsub2(TCGContext *s, TCGReg rl, TCGReg rh, TCGReg al,
} }
} }
/* Bit 0 set if inversion required; bit 1 set if swapping required. */ #define SETCOND_INV TCG_TARGET_NB_REGS
#define MIPS_CMP_INV 1 #define SETCOND_NEZ (SETCOND_INV << 1)
#define MIPS_CMP_SWAP 2 #define SETCOND_FLAGS (SETCOND_INV | SETCOND_NEZ)
static const uint8_t mips_cmp_map[16] = { static int tcg_out_setcond_int(TCGContext *s, TCGCond cond, TCGReg ret,
[TCG_COND_LT] = 0, TCGReg arg1, TCGReg arg2)
[TCG_COND_LTU] = 0, {
[TCG_COND_GE] = MIPS_CMP_INV, int flags = 0;
[TCG_COND_GEU] = MIPS_CMP_INV,
[TCG_COND_LE] = MIPS_CMP_INV | MIPS_CMP_SWAP, switch (cond) {
[TCG_COND_LEU] = MIPS_CMP_INV | MIPS_CMP_SWAP, case TCG_COND_EQ: /* -> NE */
[TCG_COND_GT] = MIPS_CMP_SWAP, case TCG_COND_GE: /* -> LT */
[TCG_COND_GTU] = MIPS_CMP_SWAP, case TCG_COND_GEU: /* -> LTU */
}; case TCG_COND_LE: /* -> GT */
case TCG_COND_LEU: /* -> GTU */
cond = tcg_invert_cond(cond);
flags ^= SETCOND_INV;
break;
default:
break;
}
switch (cond) {
case TCG_COND_NE:
flags |= SETCOND_NEZ;
if (arg2 == 0) {
return arg1 | flags;
}
tcg_out_opc_reg(s, OPC_XOR, ret, arg1, arg2);
break;
case TCG_COND_LT:
tcg_out_opc_reg(s, OPC_SLT, ret, arg1, arg2);
break;
case TCG_COND_LTU:
tcg_out_opc_reg(s, OPC_SLTU, ret, arg1, arg2);
break;
case TCG_COND_GT:
tcg_out_opc_reg(s, OPC_SLT, ret, arg2, arg1);
break;
case TCG_COND_GTU:
tcg_out_opc_reg(s, OPC_SLTU, ret, arg2, arg1);
break;
default:
g_assert_not_reached();
}
return ret | flags;
}
static void tcg_out_setcond_end(TCGContext *s, TCGReg ret, int tmpflags)
{
if (tmpflags != ret) {
TCGReg tmp = tmpflags & ~SETCOND_FLAGS;
switch (tmpflags & SETCOND_FLAGS) {
case SETCOND_INV:
/* Intermediate result is boolean: simply invert. */
tcg_out_opc_imm(s, OPC_XORI, ret, tmp, 1);
break;
case SETCOND_NEZ:
/* Intermediate result is zero/non-zero: test != 0. */
tcg_out_opc_reg(s, OPC_SLTU, ret, TCG_REG_ZERO, tmp);
break;
case SETCOND_NEZ | SETCOND_INV:
/* Intermediate result is zero/non-zero: test == 0. */
tcg_out_opc_imm(s, OPC_SLTIU, ret, tmp, 1);
break;
default:
g_assert_not_reached();
}
}
}
static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGReg ret, static void tcg_out_setcond(TCGContext *s, TCGCond cond, TCGReg ret,
TCGReg arg1, TCGReg arg2) TCGReg arg1, TCGReg arg2)
{ {
MIPSInsn s_opc = OPC_SLTU; int tmpflags = tcg_out_setcond_int(s, cond, ret, arg1, arg2);
int cmp_map; tcg_out_setcond_end(s, ret, tmpflags);
switch (cond) {
case TCG_COND_EQ:
if (arg2 != 0) {
tcg_out_opc_reg(s, OPC_XOR, ret, arg1, arg2);
arg1 = ret;
}
tcg_out_opc_imm(s, OPC_SLTIU, ret, arg1, 1);
break;
case TCG_COND_NE:
if (arg2 != 0) {
tcg_out_opc_reg(s, OPC_XOR, ret, arg1, arg2);
arg1 = ret;
}
tcg_out_opc_reg(s, OPC_SLTU, ret, TCG_REG_ZERO, arg1);
break;
case TCG_COND_LT:
case TCG_COND_GE:
case TCG_COND_LE:
case TCG_COND_GT:
s_opc = OPC_SLT;
/* FALLTHRU */
case TCG_COND_LTU:
case TCG_COND_GEU:
case TCG_COND_LEU:
case TCG_COND_GTU:
cmp_map = mips_cmp_map[cond];
if (cmp_map & MIPS_CMP_SWAP) {
TCGReg t = arg1;
arg1 = arg2;
arg2 = t;
}
tcg_out_opc_reg(s, s_opc, ret, arg1, arg2);
if (cmp_map & MIPS_CMP_INV) {
tcg_out_opc_imm(s, OPC_XORI, ret, ret, 1);
}
break;
default:
g_assert_not_reached();
break;
}
} }
static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1, static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
@ -948,9 +960,7 @@ static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
[TCG_COND_GE] = OPC_BGEZ, [TCG_COND_GE] = OPC_BGEZ,
}; };
MIPSInsn s_opc = OPC_SLTU; MIPSInsn b_opc = 0;
MIPSInsn b_opc;
int cmp_map;
switch (cond) { switch (cond) {
case TCG_COND_EQ: case TCG_COND_EQ:
@ -959,7 +969,6 @@ static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
case TCG_COND_NE: case TCG_COND_NE:
b_opc = OPC_BNE; b_opc = OPC_BNE;
break; break;
case TCG_COND_LT: case TCG_COND_LT:
case TCG_COND_GT: case TCG_COND_GT:
case TCG_COND_LE: case TCG_COND_LE:
@ -968,133 +977,76 @@ static void tcg_out_brcond(TCGContext *s, TCGCond cond, TCGReg arg1,
b_opc = b_zero[cond]; b_opc = b_zero[cond];
arg2 = arg1; arg2 = arg1;
arg1 = 0; arg1 = 0;
break;
} }
s_opc = OPC_SLT;
/* FALLTHRU */
case TCG_COND_LTU:
case TCG_COND_GTU:
case TCG_COND_LEU:
case TCG_COND_GEU:
cmp_map = mips_cmp_map[cond];
if (cmp_map & MIPS_CMP_SWAP) {
TCGReg t = arg1;
arg1 = arg2;
arg2 = t;
}
tcg_out_opc_reg(s, s_opc, TCG_TMP0, arg1, arg2);
b_opc = (cmp_map & MIPS_CMP_INV ? OPC_BEQ : OPC_BNE);
arg1 = TCG_TMP0;
arg2 = TCG_REG_ZERO;
break; break;
default: default:
g_assert_not_reached();
break; break;
} }
if (b_opc == 0) {
int tmpflags = tcg_out_setcond_int(s, cond, TCG_TMP0, arg1, arg2);
arg2 = TCG_REG_ZERO;
arg1 = tmpflags & ~SETCOND_FLAGS;
b_opc = tmpflags & SETCOND_INV ? OPC_BEQ : OPC_BNE;
}
tcg_out_reloc(s, s->code_ptr, R_MIPS_PC16, l, 0);
tcg_out_opc_br(s, b_opc, arg1, arg2); tcg_out_opc_br(s, b_opc, arg1, arg2);
tcg_out_reloc(s, s->code_ptr - 1, R_MIPS_PC16, l, 0);
tcg_out_nop(s); tcg_out_nop(s);
} }
static TCGReg tcg_out_reduce_eq2(TCGContext *s, TCGReg tmp0, TCGReg tmp1, static int tcg_out_setcond2_int(TCGContext *s, TCGCond cond, TCGReg ret,
TCGReg al, TCGReg ah, TCGReg al, TCGReg ah, TCGReg bl, TCGReg bh)
TCGReg bl, TCGReg bh)
{ {
/* Merge highpart comparison into AH. */ int flags = 0;
if (bh != 0) {
if (ah != 0) { switch (cond) {
tcg_out_opc_reg(s, OPC_XOR, tmp0, ah, bh); case TCG_COND_EQ:
ah = tmp0; flags |= SETCOND_INV;
} else { /* fall through */
ah = bh; case TCG_COND_NE:
} flags |= SETCOND_NEZ;
tcg_out_opc_reg(s, OPC_XOR, TCG_TMP0, al, bl);
tcg_out_opc_reg(s, OPC_XOR, TCG_TMP1, ah, bh);
tcg_out_opc_reg(s, OPC_OR, ret, TCG_TMP0, TCG_TMP1);
break;
default:
tcg_out_setcond(s, TCG_COND_EQ, TCG_TMP0, ah, bh);
tcg_out_setcond(s, tcg_unsigned_cond(cond), TCG_TMP1, al, bl);
tcg_out_opc_reg(s, OPC_AND, TCG_TMP1, TCG_TMP1, TCG_TMP0);
tcg_out_setcond(s, tcg_high_cond(cond), TCG_TMP0, ah, bh);
tcg_out_opc_reg(s, OPC_OR, ret, TCG_TMP0, TCG_TMP1);
break;
} }
/* Merge lowpart comparison into AL. */ return ret | flags;
if (bl != 0) {
if (al != 0) {
tcg_out_opc_reg(s, OPC_XOR, tmp1, al, bl);
al = tmp1;
} else {
al = bl;
}
}
/* Merge high and low part comparisons into AL. */
if (ah != 0) {
if (al != 0) {
tcg_out_opc_reg(s, OPC_OR, tmp0, ah, al);
al = tmp0;
} else {
al = ah;
}
}
return al;
} }
static void tcg_out_setcond2(TCGContext *s, TCGCond cond, TCGReg ret, static void tcg_out_setcond2(TCGContext *s, TCGCond cond, TCGReg ret,
TCGReg al, TCGReg ah, TCGReg bl, TCGReg bh) TCGReg al, TCGReg ah, TCGReg bl, TCGReg bh)
{ {
TCGReg tmp0 = TCG_TMP0; int tmpflags = tcg_out_setcond2_int(s, cond, ret, al, ah, bl, bh);
TCGReg tmp1 = ret; tcg_out_setcond_end(s, ret, tmpflags);
tcg_debug_assert(ret != TCG_TMP0);
if (ret == ah || ret == bh) {
tcg_debug_assert(ret != TCG_TMP1);
tmp1 = TCG_TMP1;
}
switch (cond) {
case TCG_COND_EQ:
case TCG_COND_NE:
tmp1 = tcg_out_reduce_eq2(s, tmp0, tmp1, al, ah, bl, bh);
tcg_out_setcond(s, cond, ret, tmp1, TCG_REG_ZERO);
break;
default:
tcg_out_setcond(s, TCG_COND_EQ, tmp0, ah, bh);
tcg_out_setcond(s, tcg_unsigned_cond(cond), tmp1, al, bl);
tcg_out_opc_reg(s, OPC_AND, tmp1, tmp1, tmp0);
tcg_out_setcond(s, tcg_high_cond(cond), tmp0, ah, bh);
tcg_out_opc_reg(s, OPC_OR, ret, tmp1, tmp0);
break;
}
} }
static void tcg_out_brcond2(TCGContext *s, TCGCond cond, TCGReg al, TCGReg ah, static void tcg_out_brcond2(TCGContext *s, TCGCond cond, TCGReg al, TCGReg ah,
TCGReg bl, TCGReg bh, TCGLabel *l) TCGReg bl, TCGReg bh, TCGLabel *l)
{ {
TCGCond b_cond = TCG_COND_NE; int tmpflags = tcg_out_setcond2_int(s, cond, TCG_TMP0, al, ah, bl, bh);
TCGReg tmp = TCG_TMP1; TCGReg tmp = tmpflags & ~SETCOND_FLAGS;
MIPSInsn b_opc = tmpflags & SETCOND_INV ? OPC_BEQ : OPC_BNE;
/* With branches, we emit between 4 and 9 insns with 2 or 3 branches. tcg_out_reloc(s, s->code_ptr, R_MIPS_PC16, l, 0);
With setcond, we emit between 3 and 10 insns and only 1 branch, tcg_out_opc_br(s, b_opc, tmp, TCG_REG_ZERO);
which ought to get better branch prediction. */ tcg_out_nop(s);
switch (cond) {
case TCG_COND_EQ:
case TCG_COND_NE:
b_cond = cond;
tmp = tcg_out_reduce_eq2(s, TCG_TMP0, TCG_TMP1, al, ah, bl, bh);
break;
default:
/* Minimize code size by preferring a compare not requiring INV. */
if (mips_cmp_map[cond] & MIPS_CMP_INV) {
cond = tcg_invert_cond(cond);
b_cond = TCG_COND_EQ;
}
tcg_out_setcond2(s, cond, tmp, al, ah, bl, bh);
break;
}
tcg_out_brcond(s, b_cond, tmp, TCG_REG_ZERO, l);
} }
static void tcg_out_movcond(TCGContext *s, TCGCond cond, TCGReg ret, static void tcg_out_movcond(TCGContext *s, TCGCond cond, TCGReg ret,
TCGReg c1, TCGReg c2, TCGReg v1, TCGReg v2) TCGReg c1, TCGReg c2, TCGReg v1, TCGReg v2)
{ {
bool eqz = false; int tmpflags;
bool eqz;
/* If one of the values is zero, put it last to match SEL*Z instructions */ /* If one of the values is zero, put it last to match SEL*Z instructions */
if (use_mips32r6_instructions && v1 == 0) { if (use_mips32r6_instructions && v1 == 0) {
@ -1103,27 +1055,9 @@ static void tcg_out_movcond(TCGContext *s, TCGCond cond, TCGReg ret,
cond = tcg_invert_cond(cond); cond = tcg_invert_cond(cond);
} }
switch (cond) { tmpflags = tcg_out_setcond_int(s, cond, TCG_TMP0, c1, c2);
case TCG_COND_EQ: c1 = tmpflags & ~SETCOND_FLAGS;
eqz = true; eqz = tmpflags & SETCOND_INV;
/* FALLTHRU */
case TCG_COND_NE:
if (c2 != 0) {
tcg_out_opc_reg(s, OPC_XOR, TCG_TMP0, c1, c2);
c1 = TCG_TMP0;
}
break;
default:
/* Minimize code size by preferring a compare not requiring INV. */
if (mips_cmp_map[cond] & MIPS_CMP_INV) {
cond = tcg_invert_cond(cond);
eqz = true;
}
tcg_out_setcond(s, cond, TCG_TMP0, c1, c2);
c1 = TCG_TMP0;
break;
}
if (use_mips32r6_instructions) { if (use_mips32r6_instructions) {
MIPSInsn m_opc_t = eqz ? OPC_SELEQZ : OPC_SELNEZ; MIPSInsn m_opc_t = eqz ? OPC_SELEQZ : OPC_SELNEZ;