Skip to content

Commit fcf5ac8

Browse files
urnathanurnathan
authored
[llvm][NFC] Autoupdater x86 intrinsic selection (#73046)
Sort x86 intrinsics and use prefix checking.
1 parent 494c9e5 commit fcf5ac8

File tree

1 file changed

+98
-101
lines changed

1 file changed

+98
-101
lines changed

‎llvm/lib/IR/AutoUpgrade.cpp

Lines changed: 98 additions & 101 deletions
Original file line numberDiff line numberDiff line change
@@ -486,10 +486,8 @@ static bool ShouldUpgradeX86Intrinsic(Function *F, StringRef Name) {
486486
static bool UpgradeX86IntrinsicFunction(Function *F, StringRef Name,
487487
Function *&NewFn) {
488488
// Only handle intrinsics that start with "x86.".
489-
if (!Name.starts_with("x86."))
489+
if (!Name.consume_front("x86."))
490490
return false;
491-
// Remove "x86." prefix.
492-
Name = Name.substr(4);
493491

494492
if (ShouldUpgradeX86Intrinsic(F, Name)) {
495493
NewFn = nullptr;
@@ -507,113 +505,112 @@ static bool UpgradeX86IntrinsicFunction(Function *F, StringRef Name,
507505
return true;
508506
}
509507

508+
Intrinsic::ID ID;
509+
510510
// SSE4.1 ptest functions may have an old signature.
511-
if (Name.starts_with("sse41.ptest")) { // Added in 3.2
512-
if (Name.substr(11) == "c")
513-
return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestc, NewFn);
514-
if (Name.substr(11) == "z")
515-
return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestz, NewFn);
516-
if (Name.substr(11) == "nzc")
517-
return UpgradePTESTIntrinsic(F, Intrinsic::x86_sse41_ptestnzc, NewFn);
511+
if (Name.consume_front("sse41.ptest")) { // Added in 3.2
512+
ID = StringSwitch<Intrinsic::ID>(Name)
513+
.Case("c", Intrinsic::x86_sse41_ptestc)
514+
.Case("z", Intrinsic::x86_sse41_ptestz)
515+
.Case("nzc", Intrinsic::x86_sse41_ptestnzc)
516+
.Default(Intrinsic::not_intrinsic);
517+
if (ID != Intrinsic::not_intrinsic)
518+
return UpgradePTESTIntrinsic(F, ID, NewFn);
519+
520+
return false;
518521
}
522+
519523
// Several blend and other instructions with masks used the wrong number of
520524
// bits.
521-
if (Name == "sse41.insertps") // Added in 3.6
522-
return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_insertps,
523-
NewFn);
524-
if (Name == "sse41.dppd") // Added in 3.6
525-
return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dppd,
526-
NewFn);
527-
if (Name == "sse41.dpps") // Added in 3.6
528-
return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_dpps,
529-
NewFn);
530-
if (Name == "sse41.mpsadbw") // Added in 3.6
531-
return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_sse41_mpsadbw,
532-
NewFn);
533-
if (Name == "avx.dp.ps.256") // Added in 3.6
534-
return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx_dp_ps_256,
535-
NewFn);
536-
if (Name == "avx2.mpsadbw") // Added in 3.6
537-
return UpgradeX86IntrinsicsWith8BitMask(F, Intrinsic::x86_avx2_mpsadbw,
538-
NewFn);
539-
if (Name == "avx512.mask.cmp.pd.128") // Added in 7.0
540-
return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_pd_128,
541-
NewFn);
542-
if (Name == "avx512.mask.cmp.pd.256") // Added in 7.0
543-
return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_pd_256,
544-
NewFn);
545-
if (Name == "avx512.mask.cmp.pd.512") // Added in 7.0
546-
return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_pd_512,
547-
NewFn);
548-
if (Name == "avx512.mask.cmp.ps.128") // Added in 7.0
549-
return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_ps_128,
550-
NewFn);
551-
if (Name == "avx512.mask.cmp.ps.256") // Added in 7.0
552-
return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_ps_256,
553-
NewFn);
554-
if (Name == "avx512.mask.cmp.ps.512") // Added in 7.0
555-
return UpgradeX86MaskedFPCompare(F, Intrinsic::x86_avx512_mask_cmp_ps_512,
556-
NewFn);
557-
if (Name == "avx512bf16.cvtne2ps2bf16.128") // Added in 9.0
558-
return UpgradeX86BF16Intrinsic(
559-
F, Intrinsic::x86_avx512bf16_cvtne2ps2bf16_128, NewFn);
560-
if (Name == "avx512bf16.cvtne2ps2bf16.256") // Added in 9.0
561-
return UpgradeX86BF16Intrinsic(
562-
F, Intrinsic::x86_avx512bf16_cvtne2ps2bf16_256, NewFn);
563-
if (Name == "avx512bf16.cvtne2ps2bf16.512") // Added in 9.0
564-
return UpgradeX86BF16Intrinsic(
565-
F, Intrinsic::x86_avx512bf16_cvtne2ps2bf16_512, NewFn);
566-
if (Name == "avx512bf16.mask.cvtneps2bf16.128") // Added in 9.0
567-
return UpgradeX86BF16Intrinsic(
568-
F, Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128, NewFn);
569-
if (Name == "avx512bf16.cvtneps2bf16.256") // Added in 9.0
570-
return UpgradeX86BF16Intrinsic(
571-
F, Intrinsic::x86_avx512bf16_cvtneps2bf16_256, NewFn);
572-
if (Name == "avx512bf16.cvtneps2bf16.512") // Added in 9.0
573-
return UpgradeX86BF16Intrinsic(
574-
F, Intrinsic::x86_avx512bf16_cvtneps2bf16_512, NewFn);
575-
if (Name == "avx512bf16.dpbf16ps.128") // Added in 9.0
576-
return UpgradeX86BF16DPIntrinsic(
577-
F, Intrinsic::x86_avx512bf16_dpbf16ps_128, NewFn);
578-
if (Name == "avx512bf16.dpbf16ps.256") // Added in 9.0
579-
return UpgradeX86BF16DPIntrinsic(
580-
F, Intrinsic::x86_avx512bf16_dpbf16ps_256, NewFn);
581-
if (Name == "avx512bf16.dpbf16ps.512") // Added in 9.0
582-
return UpgradeX86BF16DPIntrinsic(
583-
F, Intrinsic::x86_avx512bf16_dpbf16ps_512, NewFn);
584-
585-
// frcz.ss/sd may need to have an argument dropped. Added in 3.2
586-
if (Name.starts_with("xop.vfrcz.ss") && F->arg_size() == 2) {
587-
rename(F);
588-
NewFn = Intrinsic::getDeclaration(F->getParent(),
589-
Intrinsic::x86_xop_vfrcz_ss);
590-
return true;
525+
526+
// Added in 3.6
527+
ID = StringSwitch<Intrinsic::ID>(Name)
528+
.Case("sse41.insertps", Intrinsic::x86_sse41_insertps)
529+
.Case("sse41.dppd", Intrinsic::x86_sse41_dppd)
530+
.Case("sse41.dpps", Intrinsic::x86_sse41_dpps)
531+
.Case("sse41.mpsadbw", Intrinsic::x86_sse41_mpsadbw)
532+
.Case("avx.dp.ps.256", Intrinsic::x86_avx_dp_ps_256)
533+
.Case("avx2.mpsadbw", Intrinsic::x86_avx2_mpsadbw)
534+
.Default(Intrinsic::not_intrinsic);
535+
if (ID != Intrinsic::not_intrinsic)
536+
return UpgradeX86IntrinsicsWith8BitMask(F, ID, NewFn);
537+
538+
if (Name.consume_front("avx512.mask.cmp.")) {
539+
// Added in 7.0
540+
ID = StringSwitch<Intrinsic::ID>(Name)
541+
.Case("pd.128", Intrinsic::x86_avx512_mask_cmp_pd_128)
542+
.Case("pd.256", Intrinsic::x86_avx512_mask_cmp_pd_256)
543+
.Case("pd.512", Intrinsic::x86_avx512_mask_cmp_pd_512)
544+
.Case("ps.128", Intrinsic::x86_avx512_mask_cmp_ps_128)
545+
.Case("ps.256", Intrinsic::x86_avx512_mask_cmp_ps_256)
546+
.Case("ps.512", Intrinsic::x86_avx512_mask_cmp_ps_512)
547+
.Default(Intrinsic::not_intrinsic);
548+
if (ID != Intrinsic::not_intrinsic)
549+
return UpgradeX86MaskedFPCompare(F, ID, NewFn);
550+
return false; // No other 'x86.avx523.mask.cmp.*'.
591551
}
592-
if (Name.starts_with("xop.vfrcz.sd") && F->arg_size() == 2) {
593-
rename(F);
594-
NewFn = Intrinsic::getDeclaration(F->getParent(),
595-
Intrinsic::x86_xop_vfrcz_sd);
596-
return true;
552+
553+
if (Name.consume_front("avx512bf16.")) {
554+
// Added in 9.0
555+
ID = StringSwitch<Intrinsic::ID>(Name)
556+
.Case("cvtne2ps2bf16.128",
557+
Intrinsic::x86_avx512bf16_cvtne2ps2bf16_128)
558+
.Case("cvtne2ps2bf16.256",
559+
Intrinsic::x86_avx512bf16_cvtne2ps2bf16_256)
560+
.Case("cvtne2ps2bf16.512",
561+
Intrinsic::x86_avx512bf16_cvtne2ps2bf16_512)
562+
.Case("mask.cvtneps2bf16.128",
563+
Intrinsic::x86_avx512bf16_mask_cvtneps2bf16_128)
564+
.Case("cvtneps2bf16.256",
565+
Intrinsic::x86_avx512bf16_cvtneps2bf16_256)
566+
.Case("cvtneps2bf16.512",
567+
Intrinsic::x86_avx512bf16_cvtneps2bf16_512)
568+
.Default(Intrinsic::not_intrinsic);
569+
if (ID != Intrinsic::not_intrinsic)
570+
return UpgradeX86BF16Intrinsic(F, ID, NewFn);
571+
572+
// Added in 9.0
573+
ID = StringSwitch<Intrinsic::ID>(Name)
574+
.Case("dpbf16ps.128", Intrinsic::x86_avx512bf16_dpbf16ps_128)
575+
.Case("dpbf16ps.256", Intrinsic::x86_avx512bf16_dpbf16ps_256)
576+
.Case("dpbf16ps.512", Intrinsic::x86_avx512bf16_dpbf16ps_512)
577+
.Default(Intrinsic::not_intrinsic);
578+
if (ID != Intrinsic::not_intrinsic)
579+
return UpgradeX86BF16DPIntrinsic(F, ID, NewFn);
580+
return false; // No other 'x86.avx512bf16.*'.
597581
}
598-
// Upgrade any XOP PERMIL2 index operand still using a float/double vector.
599-
if (Name.starts_with("xop.vpermil2")) { // Added in 3.9
600-
auto Idx = F->getFunctionType()->getParamType(2);
601-
if (Idx->isFPOrFPVectorTy()) {
582+
583+
if (Name.consume_front("xop.")) {
584+
Intrinsic::ID ID = Intrinsic::not_intrinsic;
585+
if (Name.startswith("vpermil2")) { // Added in 3.9
586+
// Upgrade any XOP PERMIL2 index operand still using a float/double
587+
// vector.
588+
auto Idx = F->getFunctionType()->getParamType(2);
589+
if (Idx->isFPOrFPVectorTy()) {
590+
unsigned IdxSize = Idx->getPrimitiveSizeInBits();
591+
unsigned EltSize = Idx->getScalarSizeInBits();
592+
if (EltSize == 64 && IdxSize == 128)
593+
ID = Intrinsic::x86_xop_vpermil2pd;
594+
else if (EltSize == 32 && IdxSize == 128)
595+
ID = Intrinsic::x86_xop_vpermil2ps;
596+
else if (EltSize == 64 && IdxSize == 256)
597+
ID = Intrinsic::x86_xop_vpermil2pd_256;
598+
else
599+
ID = Intrinsic::x86_xop_vpermil2ps_256;
600+
}
601+
} else if (F->arg_size() == 2)
602+
// frcz.ss/sd may need to have an argument dropped. Added in 3.2
603+
ID = StringSwitch<Intrinsic::ID>(Name)
604+
.Case("vfrcz.ss", Intrinsic::x86_xop_vfrcz_ss)
605+
.Case("vfrcz.sd", Intrinsic::x86_xop_vfrcz_sd)
606+
.Default(Intrinsic::not_intrinsic);
607+
608+
if (ID != Intrinsic::not_intrinsic) {
602609
rename(F);
603-
unsigned IdxSize = Idx->getPrimitiveSizeInBits();
604-
unsigned EltSize = Idx->getScalarSizeInBits();
605-
Intrinsic::ID Permil2ID;
606-
if (EltSize == 64 && IdxSize == 128)
607-
Permil2ID = Intrinsic::x86_xop_vpermil2pd;
608-
else if (EltSize == 32 && IdxSize == 128)
609-
Permil2ID = Intrinsic::x86_xop_vpermil2ps;
610-
else if (EltSize == 64 && IdxSize == 256)
611-
Permil2ID = Intrinsic::x86_xop_vpermil2pd_256;
612-
else
613-
Permil2ID = Intrinsic::x86_xop_vpermil2ps_256;
614-
NewFn = Intrinsic::getDeclaration(F->getParent(), Permil2ID);
610+
NewFn = Intrinsic::getDeclaration(F->getParent(), ID);
615611
return true;
616612
}
613+
return false; // No other 'x86.xop.*'
617614
}
618615

619616
if (Name == "seh.recoverfp") {

0 commit comments

Comments
 (0)