-
Notifications
You must be signed in to change notification settings - Fork 14.9k
[DirectX] Add emulation for fp16 types of llvm.is.fpclass #157505
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
Conversation
fixes llvm#157504 This changes adds the emulation we need for IsNaN, IsNormal, & IsFinite This change only applies these emulations to the llvm.is.fpclass cases of fp16. Since there is no DX intrinsics yet for these cases, applying the emulation to the necessary intrinsics is left for future implementers of - llvm#99132 - llvm#156069 - llvm#99131
@llvm/pr-subscribers-backend-directx Author: Farzon Lotfi (farzonl) Changesfixes #157504 This changes adds the emulation we need for IsNaN, IsNormal, & IsFinite This change only applies these emulations to the llvm.is.fpclass cases of fp16. Since there is no DX intrinsics yet for these cases, applying the emulation to the necessary intrinsics is left for future implementers of
Full diff: https://github.com/llvm/llvm-project/pull/157505.diff 2 Files Affected:
diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index c613b351d85b6..b19f4d94a0bb6 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -86,6 +86,115 @@ static Value *expand16BitIsInf(CallInst *Orig) {
return B3;
}
+static Value *expand16BitIsNaN(CallInst *Orig) {
+ Module *M = Orig->getModule();
+ if (M->getTargetTriple().getDXILVersion() >= VersionTuple(1, 9))
+ return nullptr;
+
+ Value *Val = Orig->getOperand(0);
+ Type *ValTy = Val->getType();
+ if (!ValTy->getScalarType()->isHalfTy())
+ return nullptr;
+
+ IRBuilder<> Builder(Orig);
+ Type *IType = Type::getInt16Ty(M->getContext());
+
+ Constant *ExpBitMask =
+ ValTy->isVectorTy()
+ ? ConstantVector::getSplat(
+ ElementCount::getFixed(
+ cast<FixedVectorType>(ValTy)->getNumElements()),
+ ConstantInt::get(IType, 0x7c00))
+ : ConstantInt::get(IType, 0x7c00);
+ Constant *SigBitMask =
+ ValTy->isVectorTy()
+ ? ConstantVector::getSplat(
+ ElementCount::getFixed(
+ cast<FixedVectorType>(ValTy)->getNumElements()),
+ ConstantInt::get(IType, 0x3ff))
+ : ConstantInt::get(IType, 0x3ff);
+
+ Constant *Zero =
+ ValTy->isVectorTy()
+ ? ConstantVector::getSplat(
+ ElementCount::getFixed(
+ cast<FixedVectorType>(ValTy)->getNumElements()),
+ ConstantInt::get(IType, 0))
+ : ConstantInt::get(IType, 0);
+
+ Value *IVal = Builder.CreateBitCast(Val, ExpBitMask->getType());
+ Value *Exp = Builder.CreateAnd(IVal, ExpBitMask);
+ Value *B1 = Builder.CreateICmpEQ(Exp, ExpBitMask);
+
+ Value *Sig = Builder.CreateAnd(IVal, SigBitMask);
+ Value *B2 = Builder.CreateICmpNE(Sig, Zero);
+ Value *B3 = Builder.CreateAnd(B1, B2);
+ return B3;
+}
+
+static Value *expand16BitIsFinite(CallInst *Orig) {
+ Module *M = Orig->getModule();
+ if (M->getTargetTriple().getDXILVersion() >= VersionTuple(1, 9))
+ return nullptr;
+
+ Value *Val = Orig->getOperand(0);
+ Type *ValTy = Val->getType();
+ if (!ValTy->getScalarType()->isHalfTy())
+ return nullptr;
+
+ IRBuilder<> Builder(Orig);
+ Type *IType = Type::getInt16Ty(M->getContext());
+
+ Constant *ExpBitMask =
+ ValTy->isVectorTy()
+ ? ConstantVector::getSplat(
+ ElementCount::getFixed(
+ cast<FixedVectorType>(ValTy)->getNumElements()),
+ ConstantInt::get(IType, 0x7c00))
+ : ConstantInt::get(IType, 0x7c00);
+
+ Value *IVal = Builder.CreateBitCast(Val, ExpBitMask->getType());
+ Value *Exp = Builder.CreateAnd(IVal, ExpBitMask);
+ Value *B1 = Builder.CreateICmpNE(Exp, ExpBitMask);
+ return B1;
+}
+
+static Value *expand16BitIsNormal(CallInst *Orig) {
+ Module *M = Orig->getModule();
+ if (M->getTargetTriple().getDXILVersion() >= VersionTuple(1, 9))
+ return nullptr;
+
+ Value *Val = Orig->getOperand(0);
+ Type *ValTy = Val->getType();
+ if (!ValTy->getScalarType()->isHalfTy())
+ return nullptr;
+
+ IRBuilder<> Builder(Orig);
+ Type *IType = Type::getInt16Ty(M->getContext());
+
+ Constant *ExpBitMask =
+ ValTy->isVectorTy()
+ ? ConstantVector::getSplat(
+ ElementCount::getFixed(
+ cast<FixedVectorType>(ValTy)->getNumElements()),
+ ConstantInt::get(IType, 0x7c00))
+ : ConstantInt::get(IType, 0x7c00);
+ Constant *Zero =
+ ValTy->isVectorTy()
+ ? ConstantVector::getSplat(
+ ElementCount::getFixed(
+ cast<FixedVectorType>(ValTy)->getNumElements()),
+ ConstantInt::get(IType, 0))
+ : ConstantInt::get(IType, 0);
+
+ Value *IVal = Builder.CreateBitCast(Val, ExpBitMask->getType());
+ Value *Exp = Builder.CreateAnd(IVal, ExpBitMask);
+ Value *NotAllZeroes = Builder.CreateICmpNE(Exp, Zero);
+ Value *NotAllOnes = Builder.CreateICmpNE(Exp, ExpBitMask);
+ Value *B1 = Builder.CreateAnd(NotAllZeroes, NotAllOnes);
+ return B1;
+}
+
static bool isIntrinsicExpansion(Function &F) {
switch (F.getIntrinsicID()) {
case Intrinsic::abs:
@@ -342,8 +451,11 @@ static Value *expandIsFPClass(CallInst *Orig) {
case FPClassTest::fcInf:
return expand16BitIsInf(Orig);
case FPClassTest::fcNan:
+ return expand16BitIsNaN(Orig);
case FPClassTest::fcNormal:
+ return expand16BitIsNormal(Orig);
case FPClassTest::fcFinite:
+ return expand16BitIsFinite(Orig);
return nullptr;
}
diff --git a/llvm/test/CodeGen/DirectX/is_fpclass.ll b/llvm/test/CodeGen/DirectX/is_fpclass.ll
index 1796e8bd794d8..f5804f3418853 100644
--- a/llvm/test/CodeGen/DirectX/is_fpclass.ll
+++ b/llvm/test/CodeGen/DirectX/is_fpclass.ll
@@ -46,6 +46,25 @@ entry:
ret i1 %0
}
+define noundef i1 @isnanh(half noundef %a) {
+; CHECK-LABEL: define noundef i1 @isnanh(
+; CHECK-SAME: half noundef [[A:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; SM69CHECK-NEXT: [[TMP0:%.*]] = call i1 @dx.op.isSpecialFloat.f16(i32 8, half [[A]]) #[[ATTR0:[0-9]+]]
+; SMOLDCHECK-NEXT: [[BITCAST:%.*]] = bitcast half [[A]] to i16
+; SMOLDCHECK-NEXT: [[AND:%.*]] = and i16 [[BITCAST]], 31744
+; SMOLDCHECK-NEXT: [[CMPHIGH:%.*]] = icmp eq i16 [[AND]], 31744
+; SMOLDCHECK-NEXT: [[ANDLOW:%.*]] = and i16 [[BITCAST]], 1023
+; SMOLDCHECK-NEXT: [[CMPZERO:%.*]] = icmp ne i16 [[ANDLOW]], 0
+; SMOLDCHECK-NEXT: [[ANDLOW:%.*]] = and i1 [[CMPHIGH]], [[CMPZERO]]
+; SMOLDCHECK-NEXT: ret i1 [[ANDLOW]]
+; SM69CHECK-NEXT: ret i1 [[TMP0]]
+;
+entry:
+ %0 = call i1 @llvm.is.fpclass.f16(half %a, i32 3)
+ ret i1 %0
+}
+
define noundef <2 x i1> @isnanv2(<2 x float> noundef %a) {
; CHECK-LABEL: define noundef <2 x i1> @isnanv2(
; CHECK-SAME: <2 x float> noundef [[A:%.*]]) {
@@ -63,6 +82,40 @@ entry:
ret <2 x i1> %0
}
+define noundef <2 x i1> @isnanhv2(<2 x half> noundef %a) {
+; CHECK-LABEL: define noundef <2 x i1> @isnanhv2(
+; CHECK-SAME: <2 x half> noundef [[A:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[A_I0:%.*]] = extractelement <2 x half> [[A]], i64 0
+; SM69CHECK-NEXT: [[DOTI02:%.*]] = call i1 @dx.op.isSpecialFloat.f16(i32 8, half [[A_I0]]) #[[ATTR0:[0-9]+]]
+; SM69CHECK-NEXT: [[A_I1:%.*]] = extractelement <2 x half> [[A]], i64 1
+; SM69CHECK-NEXT: [[DOTI11:%.*]] = call i1 @dx.op.isSpecialFloat.f16(i32 8, half [[A_I1]]) #[[ATTR0]]
+; SM69CHECK-NEXT: [[DOTUPTO0:%.*]] = insertelement <2 x i1> poison, i1 [[DOTI02]], i64 0
+; SM69CHECK-NEXT: [[TMP0:%.*]] = insertelement <2 x i1> [[DOTUPTO0]], i1 [[DOTI11]], i64 1
+; SM69CHECK-NEXT: ret <2 x i1> [[TMP0]]
+;
+; SMOLDCHECK-NEXT: [[DOTI0:%.*]] = bitcast half [[A_I0]] to i16
+; SMOLDCHECK-NEXT: [[A_I1:%.*]] = extractelement <2 x half> [[A]], i64 1
+; SMOLDCHECK-NEXT: [[DOTI1:%.*]] = bitcast half [[A_I1]] to i16
+; SMOLDCHECK-NEXT: [[DOTI01:%.*]] = and i16 [[DOTI0]], 31744
+; SMOLDCHECK-NEXT: [[DOTI12:%.*]] = and i16 [[DOTI1]], 31744
+; SMOLDCHECK-NEXT: [[DOTI03:%.*]] = icmp eq i16 [[DOTI01]], 31744
+; SMOLDCHECK-NEXT: [[DOTI14:%.*]] = icmp eq i16 [[DOTI12]], 31744
+; SMOLDCHECK-NEXT: [[DOTI05:%.*]] = and i16 [[DOTI0]], 1023
+; SMOLDCHECK-NEXT: [[DOTI16:%.*]] = and i16 [[DOTI1]], 1023
+; SMOLDCHECK-NEXT: [[DOTI07:%.*]] = icmp ne i16 [[DOTI05]], 0
+; SMOLDCHECK-NEXT: [[DOTI18:%.*]] = icmp ne i16 [[DOTI16]], 0
+; SMOLDCHECK-NEXT: [[DOTI09:%.*]] = and i1 [[DOTI03]], [[DOTI07]]
+; SMOLDCHECK-NEXT: [[DOTI110:%.*]] = and i1 [[DOTI14]], [[DOTI18]]
+; SMOLDCHECK-NEXT: [[DOTUPTO015:%.*]] = insertelement <2 x i1> poison, i1 [[DOTI09]], i64 0
+; SMOLDCHECK-NEXT: [[TMP0:%.*]] = insertelement <2 x i1> [[DOTUPTO015]], i1 [[DOTI110]], i64 1
+; SMOLDCHECK-NEXT: ret <2 x i1> [[TMP0]]
+;
+entry:
+ %0 = call <2 x i1> @llvm.is.fpclass.v2f16(<2 x half> %a, i32 3)
+ ret <2 x i1> %0
+}
+
define noundef i1 @isinf(float noundef %a) {
; CHECK-LABEL: define noundef i1 @isinf(
; CHECK-SAME: float noundef [[A:%.*]]) {
@@ -80,7 +133,7 @@ define noundef i1 @isinfh(half noundef %a) {
; CHECK-SAME: half noundef [[A:%.*]]) {
; CHECK-NEXT: [[ENTRY:.*:]]
; SM69CHECK-NEXT: [[ISINF:%.*]] = call i1 @dx.op.isSpecialFloat.f16(i32 9, half [[A]]) #[[ATTR0]]
-; SMOLDCHECK-NEXT: [[BITCAST:%.*]] = bitcast half %a to i16
+; SMOLDCHECK-NEXT: [[BITCAST:%.*]] = bitcast half [[A]] to i16
; SMOLDCHECK-NEXT: [[CMPHIGH:%.*]] = icmp eq i16 [[BITCAST]], 31744
; SMOLDCHECK-NEXT: [[CMPLOW:%.*]] = icmp eq i16 [[BITCAST]], -1024
; SMOLDCHECK-NEXT: [[OR:%.*]] = or i1 [[CMPHIGH]], [[CMPLOW]]
@@ -121,6 +174,22 @@ entry:
ret i1 %0
}
+define noundef i1 @isfiniteh(half noundef %a) {
+; CHECK-LABEL: define noundef i1 @isfiniteh(
+; CHECK-SAME: half noundef [[A:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; SM69CHECK-NEXT: [[TMP0:%.*]] = call i1 @dx.op.isSpecialFloat.f16(i32 10, half [[A]]) #[[ATTR0]]
+; SMOLDCHECK-NEXT: [[BITCAST:%.*]] = bitcast half [[A]] to i16
+; SMOLDCHECK-NEXT: [[AND:%.*]] = and i16 [[BITCAST]], 31744
+; SMOLDCHECK-NEXT: [[CMPHIGH:%.*]] = icmp ne i16 [[AND]], 31744
+; SMOLDCHECK-NEXT: ret i1 [[CMPHIGH]]
+; SM69CHECK-NEXT: ret i1 [[TMP0]]
+;
+entry:
+ %0 = call i1 @llvm.is.fpclass.f16(half %a, i32 504)
+ ret i1 %0
+}
+
define noundef <2 x i1> @isfinitev2(<2 x float> noundef %a) {
; CHECK-LABEL: define noundef <2 x i1> @isfinitev2(
; CHECK-SAME: <2 x float> noundef [[A:%.*]]) {
@@ -138,6 +207,35 @@ entry:
ret <2 x i1> %0
}
+define noundef <2 x i1> @isfinitehv2(<2 x half> noundef %a) {
+; CHECK-LABEL: define noundef <2 x i1> @isfinitehv2(
+; CHECK-SAME: <2 x half> noundef [[A:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; SM69CHECK-NEXT: [[A_I0:%.*]] = extractelement <2 x half> [[A]], i64 0
+; SM69CHECK-NEXT: [[DOTI02:%.*]] = call i1 @dx.op.isSpecialFloat.f16(i32 10, half [[A_I0]]) #[[ATTR0:[0-9]+]]
+; SM69CHECK-NEXT: [[A_I1:%.*]] = extractelement <2 x half> [[A]], i64 1
+; SM69CHECK-NEXT: [[DOTI11:%.*]] = call i1 @dx.op.isSpecialFloat.f16(i32 10, half [[A_I1]]) #[[ATTR0]]
+; SM69CHECK-NEXT: [[DOTUPTO0:%.*]] = insertelement <2 x i1> poison, i1 [[DOTI02]], i64 0
+; SM69CHECK-NEXT: [[TMP0:%.*]] = insertelement <2 x i1> [[DOTUPTO0]], i1 [[DOTI11]], i64 1
+; SM69CHECK-NEXT: ret <2 x i1> [[TMP0]]
+;
+; SMOLDCHECK-NEXT: [[A_I0:%.*]] = extractelement <2 x half> [[A]], i64 0
+; SMOLDCHECK-NEXT: [[DOTI0:%.*]] = bitcast half [[A_I0]] to i16
+; SMOLDCHECK-NEXT: [[A_I1:%.*]] = extractelement <2 x half> [[A]], i64 1
+; SMOLDCHECK-NEXT: [[DOTI1:%.*]] = bitcast half [[A_I1]] to i16
+; SMOLDCHECK-NEXT: [[DOTI01:%.*]] = and i16 [[DOTI0]], 31744
+; SMOLDCHECK-NEXT: [[DOTI12:%.*]] = and i16 [[DOTI1]], 31744
+; SMOLDCHECK-NEXT: [[DOTI03:%.*]] = icmp ne i16 [[DOTI01]], 31744
+; SMOLDCHECK-NEXT: [[DOTI14:%.*]] = icmp ne i16 [[DOTI12]], 31744
+; SMOLDCHECK-NEXT: [[DOTUPTO06:%.*]] = insertelement <2 x i1> poison, i1 [[DOTI03]], i64 0
+; SMOLDCHECK-NEXT: [[TMP0:%.*]] = insertelement <2 x i1> [[DOTUPTO06]], i1 [[DOTI14]], i64 1
+; SMOLDCHECK-NEXT: ret <2 x i1> [[TMP0]]
+;
+entry:
+ %0 = call <2 x i1> @llvm.is.fpclass.v2f16(<2 x half> %a, i32 504)
+ ret <2 x i1> %0
+}
+
define noundef i1 @isnormal(float noundef %a) {
; CHECK-LABEL: define noundef i1 @isnormal(
; CHECK-SAME: float noundef [[A:%.*]]) {
@@ -150,6 +248,24 @@ entry:
ret i1 %0
}
+define noundef i1 @isnormalh(half noundef %a) {
+; CHECK-LABEL: define noundef i1 @isnormalh(
+; CHECK-SAME: half noundef [[A:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; SM69CHECK-NEXT: [[TMP0:%.*]] = call i1 @dx.op.isSpecialFloat.f16(i32 11, half [[A]]) #[[ATTR0]]
+; SMOLDCHECK-NEXT: [[BITCAST:%.*]] = bitcast half [[A]] to i16
+; SMOLDCHECK-NEXT: [[AND:%.*]] = and i16 [[BITCAST]], 31744
+; SMOLDCHECK-NEXT: [[CMPZERO:%.*]] = icmp ne i16 [[AND]], 0
+; SMOLDCHECK-NEXT: [[CMPHIGH:%.*]] = icmp ne i16 [[AND]], 31744
+; SMOLDCHECK-NEXT: [[ANDCMP:%.*]] = and i1 [[CMPZERO]], [[CMPHIGH]]
+; SMOLDCHECK-NEXT: ret i1 [[ANDCMP]]
+; SM69CHECK-NEXT: ret i1 [[TMP0]]
+;
+entry:
+ %0 = call i1 @llvm.is.fpclass.f16(half %a, i32 264)
+ ret i1 %0
+}
+
define noundef <2 x i1> @isnormalv2(<2 x float> noundef %a) {
; CHECK-LABEL: define noundef <2 x i1> @isnormalv2(
; CHECK-SAME: <2 x float> noundef [[A:%.*]]) {
@@ -167,5 +283,37 @@ entry:
ret <2 x i1> %0
}
+define noundef <2 x i1> @isnormalhv2(<2 x half> noundef %a) {
+; CHECK-LABEL: define noundef <2 x i1> @isnormalhv2(
+; CHECK-SAME: <2 x half> noundef [[A:%.*]]) {
+; CHECK-NEXT: [[ENTRY:.*:]]
+; CHECK-NEXT: [[A_I0:%.*]] = extractelement <2 x half> [[A]], i64 0
+; SM69CHECK-NEXT: [[DOTI02:%.*]] = call i1 @dx.op.isSpecialFloat.f16(i32 11, half [[A_I0]]) #[[ATTR0:[0-9]+]]
+; SM69CHECK-NEXT: [[A_I1:%.*]] = extractelement <2 x half> [[A]], i64 1
+; SM69CHECK-NEXT: [[DOTI11:%.*]] = call i1 @dx.op.isSpecialFloat.f16(i32 11, half [[A_I1]]) #[[ATTR0]]
+; SM69CHECK-NEXT: [[DOTUPTO0:%.*]] = insertelement <2 x i1> poison, i1 [[DOTI02]], i64 0
+; SM69CHECK-NEXT: [[TMP0:%.*]] = insertelement <2 x i1> [[DOTUPTO0]], i1 [[DOTI11]], i64 1
+; SM69CHECK-NEXT: ret <2 x i1> [[TMP0]]
+;
+; SMOLDCHECK-NEXT: [[DOTI0:%.*]] = bitcast half [[A_I0]] to i16
+; SMOLDCHECK-NEXT: [[A_I1:%.*]] = extractelement <2 x half> [[A]], i64 1
+; SMOLDCHECK-NEXT: [[DOTI1:%.*]] = bitcast half [[A_I1]] to i16
+; SMOLDCHECK-NEXT: [[DOTI01:%.*]] = and i16 [[DOTI0]], 31744
+; SMOLDCHECK-NEXT: [[DOTI12:%.*]] = and i16 [[DOTI1]], 31744
+; SMOLDCHECK-NEXT: [[DOTI03:%.*]] = icmp ne i16 [[DOTI01]], 0
+; SMOLDCHECK-NEXT: [[DOTI14:%.*]] = icmp ne i16 [[DOTI12]], 0
+; SMOLDCHECK-NEXT: [[DOTI05:%.*]] = icmp ne i16 [[DOTI01]], 31744
+; SMOLDCHECK-NEXT: [[DOTI16:%.*]] = icmp ne i16 [[DOTI12]], 31744
+; SMOLDCHECK-NEXT: [[DOTI07:%.*]] = and i1 [[DOTI03]], [[DOTI05]]
+; SMOLDCHECK-NEXT: [[DOTI18:%.*]] = and i1 [[DOTI14]], [[DOTI16]]
+; SMOLDCHECK-NEXT: [[DOTUPTO012:%.*]] = insertelement <2 x i1> poison, i1 [[DOTI07]], i64 0
+; SMOLDCHECK-NEXT: [[TMP0:%.*]] = insertelement <2 x i1> [[DOTUPTO012]], i1 [[DOTI18]], i64 1
+; SMOLDCHECK-NEXT: ret <2 x i1> [[TMP0]]
+;
+entry:
+ %0 = call <2 x i1> @llvm.is.fpclass.v2f16(<2 x half> %a, i32 264)
+ ret <2 x i1> %0
+}
+
declare i1 @llvm.is.fpclass.f32(float, i32 immarg)
declare <2 x i1> @llvm.is.fpclass.v2f32(<2 x float>, i32 immarg)
|
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
lgtm
failed tests are not related
|
LLVM Buildbot has detected a new failure on builder Full details are available at: https://lab.llvm.org/buildbot/#/builders/157/builds/39245 Here is the relevant piece of the build log for the reference
|
fixes #157504
This changes adds the emulation we need for IsNaN, IsNormal, & IsFinite This change only applies these emulations to the llvm.is.fpclass cases of fp16.
Since there is no DX intrinsics yet for these cases, applying the emulation to the necessary intrinsics is left for future implementers of
isnan
HLSL Function #99132isfinite
HLSL Function #99131