Skip to content

[DirectX] Add isinf f16 emulation for SM6.8 and lower #156932

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.

Sign up for GitHub

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

Jump to bottom
Merged
merged 4 commits into from
Sep 5, 2025
Merged

[DirectX] Add isinf f16 emulation for SM6.8 and lower #156932

merged 4 commits into from
Sep 5, 2025
+126 −14

Conversation

farzonl
Copy link
Member

@farzonl farzonl commented Sep 4, 2025 •
edited
Loading

fixes #156068

  • We needed to add a new sub arch to the target tripple so we can test that emulation does not happen when targeting SM6.9
  • The HLSL toolchain needed to be updated to handle the conversion of strings to enums for the new sub arch.
  • The emulation is done in DXILIntrinsicExpansion.cpp and needs to be able to convert both llvm.is.fpclass and lvm.dx.isinf to the proper emulation
  • test updates in TargetParser/TripleTest.cpp, isinf.ll, is_fpclass.ll, and DXCModeTest.cpp

@llvmbot llvmbot added clang Clang issues not falling into any other category clang:driver 'clang' and 'clang++' user-facing binaries. Not 'clang-cl' backend:DirectX HLSL HLSL Language Support labels Sep 4, 2025
@llvmbot
Copy link
Member

llvmbot commented Sep 4, 2025 •
edited
Loading

@llvm/pr-subscribers-clang
@llvm/pr-subscribers-backend-directx

@llvm/pr-subscribers-clang-driver

Author: Farzon Lotfi (farzonl)

Changes

fixes #156068

  • We needed to add a new sub arch to the target tripple so we can test that emulation does not happen when targeting SM6.9
  • The HLSL toolchain needed to be updated to handle the conversion of strings to enums for the new sub arch.
  • The emulation is done in DXILIntrinsicExpansion.cpp and needs to be able to convert both llvm.is.fpclass and lvm.dx.isinf to the proper emulation
  • test updates in TargetParser/TripleTest.cpp, isinf.ll, and is_fpclass.ll

Full diff: https://github.com/llvm/llvm-project/pull/156932.diff

7 Files Affected:

  • (modified) clang/lib/Driver/ToolChains/HLSL.cpp (+3)
  • (modified) llvm/include/llvm/TargetParser/Triple.h (+2-1)
  • (modified) llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp (+42-1)
  • (modified) llvm/lib/TargetParser/Triple.cpp (+8-1)
  • (modified) llvm/test/CodeGen/DirectX/is_fpclass.ll (+19-2)
  • (modified) llvm/test/CodeGen/DirectX/isinf.ll (+37-6)
  • (modified) llvm/unittests/TargetParser/TripleTest.cpp (+13)
diff --git a/clang/lib/Driver/ToolChains/HLSL.cpp b/clang/lib/Driver/ToolChains/HLSL.cpp
index 660661945d62a..559af32dc3808 100644
--- a/clang/lib/Driver/ToolChains/HLSL.cpp
+++ b/clang/lib/Driver/ToolChains/HLSL.cpp
@@ -132,6 +132,9 @@ std::optional<std::string> tryParseProfile(StringRef Profile) {
   case 8:
     SubArch = llvm::Triple::DXILSubArch_v1_8;
     break;
+  case 9:
+    SubArch = llvm::Triple::DXILSubArch_v1_9;
+    break;
   case OfflineLibMinor:
     // Always consider minor version x as the latest supported DXIL version
     SubArch = llvm::Triple::LatestDXILSubArch;
diff --git a/llvm/include/llvm/TargetParser/Triple.h b/llvm/include/llvm/TargetParser/Triple.h
index f85984ed4f328..8e12c6852075d 100644
--- a/llvm/include/llvm/TargetParser/Triple.h
+++ b/llvm/include/llvm/TargetParser/Triple.h
@@ -180,7 +180,8 @@ class Triple {
     DXILSubArch_v1_6,
     DXILSubArch_v1_7,
     DXILSubArch_v1_8,
-    LatestDXILSubArch = DXILSubArch_v1_8,
+    DXILSubArch_v1_9,
+    LatestDXILSubArch = DXILSubArch_v1_9,
   };
   enum VendorType {
     UnknownVendor,
diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index ee1db54446cb8..7966e3f1d6f3d 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -51,6 +51,42 @@ static bool resourceAccessNeeds64BitExpansion(Module *M, Type *OverloadTy,
   return ScalarTy->isDoubleTy() || ScalarTy->isIntegerTy(64);
 }
 
+  static Value* expand16BitIsInf(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->isHalfTy() ||
+      (ValTy->isVectorTy() &&
+       cast<FixedVectorType>(ValTy)->getElementType()->isHalfTy())))
+        return nullptr;
+
+    IRBuilder<> Builder(Orig);
+    Type *IType = Type::getInt16Ty(M->getContext());
+    Constant *PosInf = ValTy->isVectorTy()
+                       ? ConstantVector::getSplat(
+                             ElementCount::getFixed(
+                                 cast<FixedVectorType>(ValTy)->getNumElements()),
+                             ConstantInt::get(IType, 0x7c00))
+                       : ConstantInt::get(IType, 0x7c00);
+
+    Constant *NegInf = ValTy->isVectorTy()
+                       ? ConstantVector::getSplat(
+                             ElementCount::getFixed(
+                                 cast<FixedVectorType>(ValTy)->getNumElements()),
+                             ConstantInt::get(IType, 0xfc00))
+                       : ConstantInt::get(IType, 0xfc00);
+
+
+    Value *IVal = Builder.CreateBitCast(Val, PosInf->getType());
+    Value *B1 = Builder.CreateICmpEQ(IVal, PosInf);
+    Value *B2 = Builder.CreateICmpEQ(IVal, NegInf);
+    Value *B3 = Builder.CreateOr(B1, B2);
+    return B3;
+  }
+
 static bool isIntrinsicExpansion(Function &F) {
   switch (F.getIntrinsicID()) {
   case Intrinsic::abs:
@@ -68,6 +104,7 @@ static bool isIntrinsicExpansion(Function &F) {
   case Intrinsic::dx_sclamp:
   case Intrinsic::dx_nclamp:
   case Intrinsic::dx_degrees:
+  case Intrinsic::dx_isinf:
   case Intrinsic::dx_lerp:
   case Intrinsic::dx_normalize:
   case Intrinsic::dx_fdot:
@@ -301,9 +338,10 @@ static Value *expandIsFPClass(CallInst *Orig) {
   auto *TCI = dyn_cast<ConstantInt>(T);
 
   // These FPClassTest cases have DXIL opcodes, so they will be handled in
-  // DXIL Op Lowering instead.
+  // DXIL Op Lowering instead for all non f16 cases.
   switch (TCI->getZExtValue()) {
   case FPClassTest::fcInf:
+    return expand16BitIsInf(Orig);
   case FPClassTest::fcNan:
   case FPClassTest::fcNormal:
   case FPClassTest::fcFinite:
@@ -873,6 +911,9 @@ static bool expandIntrinsic(Function &F, CallInst *Orig) {
   case Intrinsic::dx_degrees:
     Result = expandDegreesIntrinsic(Orig);
     break;
+  case Intrinsic::dx_isinf:
+    Result = expand16BitIsInf(Orig);
+    break;
   case Intrinsic::dx_lerp:
     Result = expandLerpIntrinsic(Orig);
     break;
diff --git a/llvm/lib/TargetParser/Triple.cpp b/llvm/lib/TargetParser/Triple.cpp
index ec15f235b8624..2ee27ba4a7de4 100644
--- a/llvm/lib/TargetParser/Triple.cpp
+++ b/llvm/lib/TargetParser/Triple.cpp
@@ -158,6 +158,8 @@ StringRef Triple::getArchName(ArchType Kind, SubArchType SubArch) {
       return "dxilv1.7";
     case Triple::DXILSubArch_v1_8:
       return "dxilv1.8";
+    case Triple::DXILSubArch_v1_9:
+      return "dxilv1.9";
     default:
       break;
     }
@@ -650,6 +652,8 @@ static Triple::ArchType parseArch(StringRef ArchName) {
                 .Cases("dxil", "dxilv1.0", "dxilv1.1", "dxilv1.2", "dxilv1.3",
                        "dxilv1.4", "dxilv1.5", "dxilv1.6", "dxilv1.7",
                        "dxilv1.8", Triple::dxil)
+                //Note: Cases has max limit of 10.
+                .Case("dxilv1.9", Triple::dxil)
                 .Case("xtensa", Triple::xtensa)
                 .Default(Triple::UnknownArch);
 
@@ -842,6 +846,7 @@ static Triple::SubArchType parseSubArch(StringRef SubArchName) {
         .EndsWith("v1.6", Triple::DXILSubArch_v1_6)
         .EndsWith("v1.7", Triple::DXILSubArch_v1_7)
         .EndsWith("v1.8", Triple::DXILSubArch_v1_8)
+        .EndsWith("v1.9", Triple::DXILSubArch_v1_9)
         .Default(Triple::NoSubArch);
 
   StringRef ARMSubArch = ARM::getCanonicalArchName(SubArchName);
@@ -1111,7 +1116,7 @@ static StringRef getDXILArchNameFromShaderModel(StringRef ShaderModelStr) {
   VersionTuple Ver =
       parseVersionFromName(ShaderModelStr.drop_front(strlen("shadermodel")));
   // Default DXIL minor version when Shader Model version is anything other
-  // than 6.[0...8] or 6.x (which translates to latest current SM version)
+  // than 6.[0...9] or 6.x (which translates to latest current SM version)
   const unsigned SMMajor = 6;
   if (!Ver.empty()) {
     if (Ver.getMajor() == SMMajor) {
@@ -1135,6 +1140,8 @@ static StringRef getDXILArchNameFromShaderModel(StringRef ShaderModelStr) {
           return Triple::getArchName(Triple::dxil, Triple::DXILSubArch_v1_7);
         case 8:
           return Triple::getArchName(Triple::dxil, Triple::DXILSubArch_v1_8);
+        case 9:
+          return Triple::getArchName(Triple::dxil, Triple::DXILSubArch_v1_9);
         default:
           report_fatal_error("Unsupported Shader Model version", false);
         }
diff --git a/llvm/test/CodeGen/DirectX/is_fpclass.ll b/llvm/test/CodeGen/DirectX/is_fpclass.ll
index a628096aacd7d..1796e8bd794d8 100644
--- a/llvm/test/CodeGen/DirectX/is_fpclass.ll
+++ b/llvm/test/CodeGen/DirectX/is_fpclass.ll
@@ -1,5 +1,5 @@
-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
-; RUN: opt -S -dxil-intrinsic-expansion -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
+; RUN: opt -S -dxil-intrinsic-expansion -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.9-library %s | FileCheck %s --check-prefixes=CHECK,SM69CHECK
+; RUN: opt -S -dxil-intrinsic-expansion -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.8-library %s | FileCheck %s --check-prefixes=CHECK,SMOLDCHECK
 
 
 define noundef i1 @isnegzero(float noundef %a) {
@@ -75,6 +75,23 @@ entry:
   ret i1 %0
 }
 
+define noundef i1 @isinfh(half noundef %a) {
+; CHECK-LABEL: define noundef i1 @isinfh(
+; 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: [[CMPHIGH:%.*]] = icmp eq i16 [[BITCAST]], 31744
+; SMOLDCHECK-NEXT: [[CMPLOW:%.*]] = icmp eq i16 [[BITCAST]], -1024
+; SMOLDCHECK-NEXT: [[OR:%.*]] = or i1 [[CMPHIGH]], [[CMPLOW]]
+; SMOLDCHECK-NEXT:   ret i1 [[OR]]
+; SM69CHECK-NEXT:    ret i1 [[ISINF]]
+;
+entry:
+  %0 = call i1 @llvm.is.fpclass.f16(half %a, i32 516)
+  ret i1 %0
+}
+
 define noundef <2 x i1> @isinfv2(<2 x float> noundef %a) {
 ; CHECK-LABEL: define noundef <2 x i1> @isinfv2(
 ; CHECK-SAME: <2 x float> noundef [[A:%.*]]) {
diff --git a/llvm/test/CodeGen/DirectX/isinf.ll b/llvm/test/CodeGen/DirectX/isinf.ll
index 461553b533ae1..bf31363ee114c 100644
--- a/llvm/test/CodeGen/DirectX/isinf.ll
+++ b/llvm/test/CodeGen/DirectX/isinf.ll
@@ -1,4 +1,5 @@
-; RUN: opt -S -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
+; RUN: opt -S -dxil-intrinsic-expansion -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.9-library %s | FileCheck %s --check-prefixes=CHECK,SM69CHECK
+; RUN: opt -S -dxil-intrinsic-expansion -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.8-library %s | FileCheck %s --check-prefixes=CHECK,SMOLDCHECK
 
 ; Make sure dxil operation function calls for isinf are generated for float and half.
 
@@ -11,17 +12,47 @@ entry:
 
 define noundef i1 @isinf_half(half noundef %a) {
 entry:
-  ; CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half %{{.*}}) #[[#ATTR]]
+  ; SM69CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half %{{.*}}) #[[#ATTR]]
+  ; SMOLDCHECK: [[BITCAST:%.*]] = bitcast half %a to i16
+  ; SMOLDCHECK: [[CMPHIGH:%.*]] = icmp eq i16 [[BITCAST]], 31744
+  ; SMOLDCHECK: [[CMPLOW:%.*]] = icmp eq i16 [[BITCAST]], -1024
+  ; SMOLDCHECK: [[OR:%.*]] = or i1 [[CMPHIGH]], [[CMPLOW]]
   %dx.isinf = call i1 @llvm.dx.isinf.f16(half %a)
   ret i1 %dx.isinf
 }
 
 define noundef <4 x i1> @isinf_half4(<4 x half> noundef %p0) {
 entry:
-  ; CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
-  ; CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
-  ; CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
-  ; CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
+  ; SM69CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
+  ; SM69CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
+  ; SM69CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
+  ; SM69CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
+
+  ; SMOLDCHECK: [[ee0:%.*]] = extractelement <4 x half> %p0, i64 0
+  ; SMOLDCHECK: [[BITCAST0:%.*]] = bitcast half [[ee0]] to i16
+  ; SMOLDCHECK: [[ee1:%.*]] = extractelement <4 x half> %p0, i64 1
+  ; SMOLDCHECK: [[BITCAST1:%.*]] = bitcast half [[ee1]] to i16
+  ; SMOLDCHECK:[[ee2:%.*]] = extractelement <4 x half> %p0, i64 2
+  ; SMOLDCHECK: [[BITCAST2:%.*]] = bitcast half [[ee2]] to i16
+  ; SMOLDCHECK: [[ee3:%.*]] = extractelement <4 x half> %p0, i64 3
+  ; SMOLDCHECK: [[BITCAST3:%.*]] = bitcast half [[ee3]] to i16
+  ; SMOLDCHECK: [[ICMPHIGH0:%.*]] = icmp eq i16 [[BITCAST0]], 31744
+  ; SMOLDCHECK: [[ICMPHIGH1:%.*]] = icmp eq i16 [[BITCAST1]], 31744
+  ; SMOLDCHECK: [[ICMPHIGH2:%.*]] = icmp eq i16 [[BITCAST2]], 31744
+  ; SMOLDCHECK: [[ICMPHIGH3:%.*]] = icmp eq i16 [[BITCAST3]], 31744
+  ; SMOLDCHECK: [[ICMPLOW0:%.*]] = icmp eq i16 [[BITCAST0]], -1024
+  ; SMOLDCHECK: [[ICMPLOW1:%.*]] = icmp eq i16 [[BITCAST1]], -1024
+  ; SMOLDCHECK: [[ICMPLOW2:%.*]] = icmp eq i16 [[BITCAST2]], -1024
+  ; SMOLDCHECK: [[ICMPLOW3:%.*]] = icmp eq i16 [[BITCAST3]], -1024
+  ; SMOLDCHECK: [[OR0:%.*]] = or i1 [[ICMPHIGH0]], [[ICMPLOW0]]
+  ; SMOLDCHECK: [[OR1:%.*]] = or i1 [[ICMPHIGH1]], [[ICMPLOW1]]
+  ; SMOLDCHECK: [[OR2:%.*]] = or i1 [[ICMPHIGH2]], [[ICMPLOW2]]
+  ; SMOLDCHECK: [[OR3:%.*]] = or i1 [[ICMPHIGH3]], [[ICMPLOW3]]
+  ; SMOLDCHECK: %.upto019 = insertelement <4 x i1> poison, i1 [[OR0]], i64 0
+  ; SMOLDCHECK: %.upto120 = insertelement <4 x i1> %.upto019, i1 [[OR1]], i64 1
+  ; SMOLDCHECK: %.upto221 = insertelement <4 x i1> %.upto120, i1 [[OR2]], i64 2
+  ; SMOLDCHECK: %0 = insertelement <4 x i1> %.upto221, i1 [[OR3]], i64 3
+
   %hlsl.isinf = call <4 x i1> @llvm.dx.isinf.v4f16(<4 x half> %p0)
   ret <4 x i1> %hlsl.isinf
 }
diff --git a/llvm/unittests/TargetParser/TripleTest.cpp b/llvm/unittests/TargetParser/TripleTest.cpp
index 7d07615d273d7..942912bcfe0e2 100644
--- a/llvm/unittests/TargetParser/TripleTest.cpp
+++ b/llvm/unittests/TargetParser/TripleTest.cpp
@@ -553,6 +553,13 @@ TEST(TripleTest, ParsedIDs) {
   EXPECT_EQ(Triple::ShaderModel, T.getOS());
   EXPECT_EQ(VersionTuple(1, 8), T.getDXILVersion());
 
+  T = Triple("dxilv1.9-unknown-shadermodel6.15-library");
+  EXPECT_EQ(Triple::dxil, T.getArch());
+  EXPECT_EQ(Triple::DXILSubArch_v1_9, T.getSubArch());
+  EXPECT_EQ(Triple::UnknownVendor, T.getVendor());
+  EXPECT_EQ(Triple::ShaderModel, T.getOS());
+  EXPECT_EQ(VersionTuple(1, 9), T.getDXILVersion());
+
   T = Triple("x86_64-unknown-fuchsia");
   EXPECT_EQ(Triple::x86_64, T.getArch());
   EXPECT_EQ(Triple::UnknownVendor, T.getVendor());
@@ -1270,6 +1277,12 @@ TEST(TripleTest, ParsedIDs) {
   EXPECT_EQ(Triple::UnknownVendor, T.getVendor());
   EXPECT_EQ(Triple::UnknownOS, T.getOS());
 
+  T = Triple("dxilv1.9-unknown-unknown");
+  EXPECT_EQ(Triple::dxil, T.getArch());
+  EXPECT_EQ(Triple::DXILSubArch_v1_9, T.getSubArch());
+  EXPECT_EQ(Triple::UnknownVendor, T.getVendor());
+  EXPECT_EQ(Triple::UnknownOS, T.getOS());
+
   // Check specification of unknown SubArch results in
   // unknown architecture.
   T = Triple("dxilv1.999-unknown-unknown");

@llvmbot
Copy link
Member

llvmbot commented Sep 4, 2025

@llvm/pr-subscribers-hlsl

Author: Farzon Lotfi (farzonl)

Changes

fixes #156068

  • We needed to add a new sub arch to the target tripple so we can test that emulation does not happen when targeting SM6.9
  • The HLSL toolchain needed to be updated to handle the conversion of strings to enums for the new sub arch.
  • The emulation is done in DXILIntrinsicExpansion.cpp and needs to be able to convert both llvm.is.fpclass and lvm.dx.isinf to the proper emulation
  • test updates in TargetParser/TripleTest.cpp, isinf.ll, and is_fpclass.ll

Full diff: https://github.com/llvm/llvm-project/pull/156932.diff

7 Files Affected:

  • (modified) clang/lib/Driver/ToolChains/HLSL.cpp (+3)
  • (modified) llvm/include/llvm/TargetParser/Triple.h (+2-1)
  • (modified) llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp (+42-1)
  • (modified) llvm/lib/TargetParser/Triple.cpp (+8-1)
  • (modified) llvm/test/CodeGen/DirectX/is_fpclass.ll (+19-2)
  • (modified) llvm/test/CodeGen/DirectX/isinf.ll (+37-6)
  • (modified) llvm/unittests/TargetParser/TripleTest.cpp (+13)
diff --git a/clang/lib/Driver/ToolChains/HLSL.cpp b/clang/lib/Driver/ToolChains/HLSL.cpp
index 660661945d62a..559af32dc3808 100644
--- a/clang/lib/Driver/ToolChains/HLSL.cpp
+++ b/clang/lib/Driver/ToolChains/HLSL.cpp
@@ -132,6 +132,9 @@ std::optional<std::string> tryParseProfile(StringRef Profile) {
   case 8:
     SubArch = llvm::Triple::DXILSubArch_v1_8;
     break;
+  case 9:
+    SubArch = llvm::Triple::DXILSubArch_v1_9;
+    break;
   case OfflineLibMinor:
     // Always consider minor version x as the latest supported DXIL version
     SubArch = llvm::Triple::LatestDXILSubArch;
diff --git a/llvm/include/llvm/TargetParser/Triple.h b/llvm/include/llvm/TargetParser/Triple.h
index f85984ed4f328..8e12c6852075d 100644
--- a/llvm/include/llvm/TargetParser/Triple.h
+++ b/llvm/include/llvm/TargetParser/Triple.h
@@ -180,7 +180,8 @@ class Triple {
     DXILSubArch_v1_6,
     DXILSubArch_v1_7,
     DXILSubArch_v1_8,
-    LatestDXILSubArch = DXILSubArch_v1_8,
+    DXILSubArch_v1_9,
+    LatestDXILSubArch = DXILSubArch_v1_9,
   };
   enum VendorType {
     UnknownVendor,
diff --git a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
index ee1db54446cb8..7966e3f1d6f3d 100644
--- a/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
+++ b/llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp
@@ -51,6 +51,42 @@ static bool resourceAccessNeeds64BitExpansion(Module *M, Type *OverloadTy,
   return ScalarTy->isDoubleTy() || ScalarTy->isIntegerTy(64);
 }
 
+  static Value* expand16BitIsInf(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->isHalfTy() ||
+      (ValTy->isVectorTy() &&
+       cast<FixedVectorType>(ValTy)->getElementType()->isHalfTy())))
+        return nullptr;
+
+    IRBuilder<> Builder(Orig);
+    Type *IType = Type::getInt16Ty(M->getContext());
+    Constant *PosInf = ValTy->isVectorTy()
+                       ? ConstantVector::getSplat(
+                             ElementCount::getFixed(
+                                 cast<FixedVectorType>(ValTy)->getNumElements()),
+                             ConstantInt::get(IType, 0x7c00))
+                       : ConstantInt::get(IType, 0x7c00);
+
+    Constant *NegInf = ValTy->isVectorTy()
+                       ? ConstantVector::getSplat(
+                             ElementCount::getFixed(
+                                 cast<FixedVectorType>(ValTy)->getNumElements()),
+                             ConstantInt::get(IType, 0xfc00))
+                       : ConstantInt::get(IType, 0xfc00);
+
+
+    Value *IVal = Builder.CreateBitCast(Val, PosInf->getType());
+    Value *B1 = Builder.CreateICmpEQ(IVal, PosInf);
+    Value *B2 = Builder.CreateICmpEQ(IVal, NegInf);
+    Value *B3 = Builder.CreateOr(B1, B2);
+    return B3;
+  }
+
 static bool isIntrinsicExpansion(Function &F) {
   switch (F.getIntrinsicID()) {
   case Intrinsic::abs:
@@ -68,6 +104,7 @@ static bool isIntrinsicExpansion(Function &F) {
   case Intrinsic::dx_sclamp:
   case Intrinsic::dx_nclamp:
   case Intrinsic::dx_degrees:
+  case Intrinsic::dx_isinf:
   case Intrinsic::dx_lerp:
   case Intrinsic::dx_normalize:
   case Intrinsic::dx_fdot:
@@ -301,9 +338,10 @@ static Value *expandIsFPClass(CallInst *Orig) {
   auto *TCI = dyn_cast<ConstantInt>(T);
 
   // These FPClassTest cases have DXIL opcodes, so they will be handled in
-  // DXIL Op Lowering instead.
+  // DXIL Op Lowering instead for all non f16 cases.
   switch (TCI->getZExtValue()) {
   case FPClassTest::fcInf:
+    return expand16BitIsInf(Orig);
   case FPClassTest::fcNan:
   case FPClassTest::fcNormal:
   case FPClassTest::fcFinite:
@@ -873,6 +911,9 @@ static bool expandIntrinsic(Function &F, CallInst *Orig) {
   case Intrinsic::dx_degrees:
     Result = expandDegreesIntrinsic(Orig);
     break;
+  case Intrinsic::dx_isinf:
+    Result = expand16BitIsInf(Orig);
+    break;
   case Intrinsic::dx_lerp:
     Result = expandLerpIntrinsic(Orig);
     break;
diff --git a/llvm/lib/TargetParser/Triple.cpp b/llvm/lib/TargetParser/Triple.cpp
index ec15f235b8624..2ee27ba4a7de4 100644
--- a/llvm/lib/TargetParser/Triple.cpp
+++ b/llvm/lib/TargetParser/Triple.cpp
@@ -158,6 +158,8 @@ StringRef Triple::getArchName(ArchType Kind, SubArchType SubArch) {
       return "dxilv1.7";
     case Triple::DXILSubArch_v1_8:
       return "dxilv1.8";
+    case Triple::DXILSubArch_v1_9:
+      return "dxilv1.9";
     default:
       break;
     }
@@ -650,6 +652,8 @@ static Triple::ArchType parseArch(StringRef ArchName) {
                 .Cases("dxil", "dxilv1.0", "dxilv1.1", "dxilv1.2", "dxilv1.3",
                        "dxilv1.4", "dxilv1.5", "dxilv1.6", "dxilv1.7",
                        "dxilv1.8", Triple::dxil)
+                //Note: Cases has max limit of 10.
+                .Case("dxilv1.9", Triple::dxil)
                 .Case("xtensa", Triple::xtensa)
                 .Default(Triple::UnknownArch);
 
@@ -842,6 +846,7 @@ static Triple::SubArchType parseSubArch(StringRef SubArchName) {
         .EndsWith("v1.6", Triple::DXILSubArch_v1_6)
         .EndsWith("v1.7", Triple::DXILSubArch_v1_7)
         .EndsWith("v1.8", Triple::DXILSubArch_v1_8)
+        .EndsWith("v1.9", Triple::DXILSubArch_v1_9)
         .Default(Triple::NoSubArch);
 
   StringRef ARMSubArch = ARM::getCanonicalArchName(SubArchName);
@@ -1111,7 +1116,7 @@ static StringRef getDXILArchNameFromShaderModel(StringRef ShaderModelStr) {
   VersionTuple Ver =
       parseVersionFromName(ShaderModelStr.drop_front(strlen("shadermodel")));
   // Default DXIL minor version when Shader Model version is anything other
-  // than 6.[0...8] or 6.x (which translates to latest current SM version)
+  // than 6.[0...9] or 6.x (which translates to latest current SM version)
   const unsigned SMMajor = 6;
   if (!Ver.empty()) {
     if (Ver.getMajor() == SMMajor) {
@@ -1135,6 +1140,8 @@ static StringRef getDXILArchNameFromShaderModel(StringRef ShaderModelStr) {
           return Triple::getArchName(Triple::dxil, Triple::DXILSubArch_v1_7);
         case 8:
           return Triple::getArchName(Triple::dxil, Triple::DXILSubArch_v1_8);
+        case 9:
+          return Triple::getArchName(Triple::dxil, Triple::DXILSubArch_v1_9);
         default:
           report_fatal_error("Unsupported Shader Model version", false);
         }
diff --git a/llvm/test/CodeGen/DirectX/is_fpclass.ll b/llvm/test/CodeGen/DirectX/is_fpclass.ll
index a628096aacd7d..1796e8bd794d8 100644
--- a/llvm/test/CodeGen/DirectX/is_fpclass.ll
+++ b/llvm/test/CodeGen/DirectX/is_fpclass.ll
@@ -1,5 +1,5 @@
-; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
-; RUN: opt -S -dxil-intrinsic-expansion -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
+; RUN: opt -S -dxil-intrinsic-expansion -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.9-library %s | FileCheck %s --check-prefixes=CHECK,SM69CHECK
+; RUN: opt -S -dxil-intrinsic-expansion -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.8-library %s | FileCheck %s --check-prefixes=CHECK,SMOLDCHECK
 
 
 define noundef i1 @isnegzero(float noundef %a) {
@@ -75,6 +75,23 @@ entry:
   ret i1 %0
 }
 
+define noundef i1 @isinfh(half noundef %a) {
+; CHECK-LABEL: define noundef i1 @isinfh(
+; 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: [[CMPHIGH:%.*]] = icmp eq i16 [[BITCAST]], 31744
+; SMOLDCHECK-NEXT: [[CMPLOW:%.*]] = icmp eq i16 [[BITCAST]], -1024
+; SMOLDCHECK-NEXT: [[OR:%.*]] = or i1 [[CMPHIGH]], [[CMPLOW]]
+; SMOLDCHECK-NEXT:   ret i1 [[OR]]
+; SM69CHECK-NEXT:    ret i1 [[ISINF]]
+;
+entry:
+  %0 = call i1 @llvm.is.fpclass.f16(half %a, i32 516)
+  ret i1 %0
+}
+
 define noundef <2 x i1> @isinfv2(<2 x float> noundef %a) {
 ; CHECK-LABEL: define noundef <2 x i1> @isinfv2(
 ; CHECK-SAME: <2 x float> noundef [[A:%.*]]) {
diff --git a/llvm/test/CodeGen/DirectX/isinf.ll b/llvm/test/CodeGen/DirectX/isinf.ll
index 461553b533ae1..bf31363ee114c 100644
--- a/llvm/test/CodeGen/DirectX/isinf.ll
+++ b/llvm/test/CodeGen/DirectX/isinf.ll
@@ -1,4 +1,5 @@
-; RUN: opt -S -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.3-library %s | FileCheck %s
+; RUN: opt -S -dxil-intrinsic-expansion -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.9-library %s | FileCheck %s --check-prefixes=CHECK,SM69CHECK
+; RUN: opt -S -dxil-intrinsic-expansion -scalarizer -dxil-op-lower -mtriple=dxil-pc-shadermodel6.8-library %s | FileCheck %s --check-prefixes=CHECK,SMOLDCHECK
 
 ; Make sure dxil operation function calls for isinf are generated for float and half.
 
@@ -11,17 +12,47 @@ entry:
 
 define noundef i1 @isinf_half(half noundef %a) {
 entry:
-  ; CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half %{{.*}}) #[[#ATTR]]
+  ; SM69CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half %{{.*}}) #[[#ATTR]]
+  ; SMOLDCHECK: [[BITCAST:%.*]] = bitcast half %a to i16
+  ; SMOLDCHECK: [[CMPHIGH:%.*]] = icmp eq i16 [[BITCAST]], 31744
+  ; SMOLDCHECK: [[CMPLOW:%.*]] = icmp eq i16 [[BITCAST]], -1024
+  ; SMOLDCHECK: [[OR:%.*]] = or i1 [[CMPHIGH]], [[CMPLOW]]
   %dx.isinf = call i1 @llvm.dx.isinf.f16(half %a)
   ret i1 %dx.isinf
 }
 
 define noundef <4 x i1> @isinf_half4(<4 x half> noundef %p0) {
 entry:
-  ; CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
-  ; CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
-  ; CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
-  ; CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
+  ; SM69CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
+  ; SM69CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
+  ; SM69CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
+  ; SM69CHECK: call i1 @dx.op.isSpecialFloat.f16(i32 9, half
+
+  ; SMOLDCHECK: [[ee0:%.*]] = extractelement <4 x half> %p0, i64 0
+  ; SMOLDCHECK: [[BITCAST0:%.*]] = bitcast half [[ee0]] to i16
+  ; SMOLDCHECK: [[ee1:%.*]] = extractelement <4 x half> %p0, i64 1
+  ; SMOLDCHECK: [[BITCAST1:%.*]] = bitcast half [[ee1]] to i16
+  ; SMOLDCHECK:[[ee2:%.*]] = extractelement <4 x half> %p0, i64 2
+  ; SMOLDCHECK: [[BITCAST2:%.*]] = bitcast half [[ee2]] to i16
+  ; SMOLDCHECK: [[ee3:%.*]] = extractelement <4 x half> %p0, i64 3
+  ; SMOLDCHECK: [[BITCAST3:%.*]] = bitcast half [[ee3]] to i16
+  ; SMOLDCHECK: [[ICMPHIGH0:%.*]] = icmp eq i16 [[BITCAST0]], 31744
+  ; SMOLDCHECK: [[ICMPHIGH1:%.*]] = icmp eq i16 [[BITCAST1]], 31744
+  ; SMOLDCHECK: [[ICMPHIGH2:%.*]] = icmp eq i16 [[BITCAST2]], 31744
+  ; SMOLDCHECK: [[ICMPHIGH3:%.*]] = icmp eq i16 [[BITCAST3]], 31744
+  ; SMOLDCHECK: [[ICMPLOW0:%.*]] = icmp eq i16 [[BITCAST0]], -1024
+  ; SMOLDCHECK: [[ICMPLOW1:%.*]] = icmp eq i16 [[BITCAST1]], -1024
+  ; SMOLDCHECK: [[ICMPLOW2:%.*]] = icmp eq i16 [[BITCAST2]], -1024
+  ; SMOLDCHECK: [[ICMPLOW3:%.*]] = icmp eq i16 [[BITCAST3]], -1024
+  ; SMOLDCHECK: [[OR0:%.*]] = or i1 [[ICMPHIGH0]], [[ICMPLOW0]]
+  ; SMOLDCHECK: [[OR1:%.*]] = or i1 [[ICMPHIGH1]], [[ICMPLOW1]]
+  ; SMOLDCHECK: [[OR2:%.*]] = or i1 [[ICMPHIGH2]], [[ICMPLOW2]]
+  ; SMOLDCHECK: [[OR3:%.*]] = or i1 [[ICMPHIGH3]], [[ICMPLOW3]]
+  ; SMOLDCHECK: %.upto019 = insertelement <4 x i1> poison, i1 [[OR0]], i64 0
+  ; SMOLDCHECK: %.upto120 = insertelement <4 x i1> %.upto019, i1 [[OR1]], i64 1
+  ; SMOLDCHECK: %.upto221 = insertelement <4 x i1> %.upto120, i1 [[OR2]], i64 2
+  ; SMOLDCHECK: %0 = insertelement <4 x i1> %.upto221, i1 [[OR3]], i64 3
+
   %hlsl.isinf = call <4 x i1> @llvm.dx.isinf.v4f16(<4 x half> %p0)
   ret <4 x i1> %hlsl.isinf
 }
diff --git a/llvm/unittests/TargetParser/TripleTest.cpp b/llvm/unittests/TargetParser/TripleTest.cpp
index 7d07615d273d7..942912bcfe0e2 100644
--- a/llvm/unittests/TargetParser/TripleTest.cpp
+++ b/llvm/unittests/TargetParser/TripleTest.cpp
@@ -553,6 +553,13 @@ TEST(TripleTest, ParsedIDs) {
   EXPECT_EQ(Triple::ShaderModel, T.getOS());
   EXPECT_EQ(VersionTuple(1, 8), T.getDXILVersion());
 
+  T = Triple("dxilv1.9-unknown-shadermodel6.15-library");
+  EXPECT_EQ(Triple::dxil, T.getArch());
+  EXPECT_EQ(Triple::DXILSubArch_v1_9, T.getSubArch());
+  EXPECT_EQ(Triple::UnknownVendor, T.getVendor());
+  EXPECT_EQ(Triple::ShaderModel, T.getOS());
+  EXPECT_EQ(VersionTuple(1, 9), T.getDXILVersion());
+
   T = Triple("x86_64-unknown-fuchsia");
   EXPECT_EQ(Triple::x86_64, T.getArch());
   EXPECT_EQ(Triple::UnknownVendor, T.getVendor());
@@ -1270,6 +1277,12 @@ TEST(TripleTest, ParsedIDs) {
   EXPECT_EQ(Triple::UnknownVendor, T.getVendor());
   EXPECT_EQ(Triple::UnknownOS, T.getOS());
 
+  T = Triple("dxilv1.9-unknown-unknown");
+  EXPECT_EQ(Triple::dxil, T.getArch());
+  EXPECT_EQ(Triple::DXILSubArch_v1_9, T.getSubArch());
+  EXPECT_EQ(Triple::UnknownVendor, T.getVendor());
+  EXPECT_EQ(Triple::UnknownOS, T.getOS());
+
   // Check specification of unknown SubArch results in
   // unknown architecture.
   T = Triple("dxilv1.999-unknown-unknown");

@github-actions GitHub Actions
Copy link

github-actions bot commented Sep 4, 2025 •
edited
Loading

✅ With the latest revision this PR passed the C/C++ code formatter.

@farzonl
[DirectX] Add isinf f16 emulation for SM6.8 and lower
8c31865 
fixes llvm#156068

- We needed to add a new sub arch to the target tripple so we can test that
emulation does not happen when targeting SM6.9
- The HLSL toolchain needed to be updated to handle the conversion of
  strings to enums for the new sub arch.
- The emulation is done in DXILIntrinsicExpansion.cpp and needs to be able to convert both llvm.is.fpclass and
  lvm.dx.isinf to the proper emulation
- test updates in TargetParser/TripleTest.cpp, isinf.ll, and is_fpclass.ll
@farzonl farzonl force-pushed the feature/isinf_f16_emulation_issue-156068 branch from dcc6ae9 to 8c31865 Compare September 4, 2025 17:24
spall
spall reviewed Sep 5, 2025
View reviewed changes
llvm/include/llvm/TargetParser/Triple.h
@@ -180,7 +180,8 @@ class Triple {
DXILSubArch_v1_6,
DXILSubArch_v1_7,
DXILSubArch_v1_8,
LatestDXILSubArch = DXILSubArch_v1_8,
DXILSubArch_v1_9,
LatestDXILSubArch = DXILSubArch_v1_9,
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

i'm not familiar with this code. What is the effect of the latestdxilsubarch? is it ok to set it to 1.9?

Copy link
Member Author

@farzonl farzonl Sep 5, 2025 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LatestDXILSubArch must always match the latest dxil version. Its uses for the 6.x target triple format.

llvm/lib/Target/DirectX/DXILIntrinsicExpansion.cpp Outdated

Value *Val = Orig->getOperand(0);
Type *ValTy = Val->getType();
if (!(ValTy->isHalfTy() ||
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Could you use getScalarType() here to prevent casting the vector?

!ValTy->getScalarType()->isHalfTy()

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

random though, are we going to have to update this code when matrix supported is added?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

yeah maybe that would be better getScalarType I think also works on matricies.

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

And yeah this whole file will need to be revisited for matricies.

Icohedron
Icohedron approved these changes Sep 5, 2025
View reviewed changes
Copy link
Contributor

@Icohedron Icohedron left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LGTM

@farzonl farzonl merged commit 16661b5 into llvm:main Sep 5, 2025
10 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Reviewers

@spall spall spall left review comments

@Icohedron Icohedron Icohedron approved these changes

Assignees
No one assigned
Labels
backend:DirectX clang:driver 'clang' and 'clang++' user-facing binaries. Not 'clang-cl' clang Clang issues not falling into any other category HLSL HLSL Language Support
Projects
None yet
Milestone
No milestone
Development

Successfully merging this pull request may close these issues.

[HLSL][DXIL] 16 bit overload of isinf should be emulated in sm6.8 and earlier.
4 participants
@farzonl @llvmbot @spall @Icohedron