https://github.com/danbrown-amd updated https://github.com/llvm/llvm-project/pull/233535
>From 2a2d5fd762edc659b1ed663d72da1608572f299e Mon Sep 17 00:00:00 2001 From: Dan Brown <[email protected]> Date: Mon, 19 Oct 2026 10:32:18 -0600 Subject: [PATCH] [HLSL] Use icmp ne instead of trunc on boolean load (#221105) Fixes #140824 Assisted-by: Claude Sonnet 4 --- .../clang/Basic/DiagnosticFrontendKinds.td | 3 ++ clang/lib/Driver/ToolChains/Clang.cpp | 1 + clang/lib/Frontend/CompilerInvocation.cpp | 10 +++++ clang/test/CodeGenHLSL/bool-vector-load.hlsl | 19 ++++++++ .../CodeGenHLSL/builtins/BoolSwizzles.hlsl | 45 +++++++------------ clang/test/CodeGenHLSL/strict-bool.hlsl | 15 +++---- .../test/Driver/HLSL/load-bool-from-mem.hlsl | 3 ++ 7 files changed, 58 insertions(+), 38 deletions(-) create mode 100644 clang/test/CodeGenHLSL/bool-vector-load.hlsl create mode 100644 clang/test/Driver/HLSL/load-bool-from-mem.hlsl diff --git a/clang/include/clang/Basic/DiagnosticFrontendKinds.td b/clang/include/clang/Basic/DiagnosticFrontendKinds.td index 3df87215a0476..b51aa3ff0a0cb 100644 --- a/clang/include/clang/Basic/DiagnosticFrontendKinds.td +++ b/clang/include/clang/Basic/DiagnosticFrontendKinds.td @@ -512,6 +512,9 @@ def warn_hlsl_signature_packing_ignored : Warning< def warn_hlsl_failed_to_embed_source : Warning<"failed to embed source for '%0' into dx.source.contents">, InGroup<HLSLDXCCompat>; +def warn_hlsl_bool_from_mem + : Warning<"-load-bool-from-mem=%0 ignored for HLSL (always nonzero)">, + InGroup<OptionIgnored>, DefaultWarn; def err_data_layout_mismatch : Error<"backend data layout '%0' does not match " "expected target description '%1'">; diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp index 608d7bb68a4e6..bd546dbb1fed2 100644 --- a/clang/lib/Driver/ToolChains/Clang.cpp +++ b/clang/lib/Driver/ToolChains/Clang.cpp @@ -4052,6 +4052,7 @@ static void RenderHLSLOptions(const Driver &D, const ArgList &Args, CmdArgs.push_back("-mllvm"); CmdArgs.push_back("--dx-pdb-in-private"); } + CmdArgs.push_back("-load-bool-from-mem=nonzero"); } static void RenderOpenACCOptions(const Driver &D, const ArgList &Args, diff --git a/clang/lib/Frontend/CompilerInvocation.cpp b/clang/lib/Frontend/CompilerInvocation.cpp index 225962359129b..c11ebf5a1a189 100644 --- a/clang/lib/Frontend/CompilerInvocation.cpp +++ b/clang/lib/Frontend/CompilerInvocation.cpp @@ -657,6 +657,16 @@ static bool FixupInvocation(CompilerInvocation &Invocation, Warnings.insert(Warnings.begin(), "vector-conversion"); if (!llvm::is_contained(Warnings, "matrix-conversion")) Warnings.insert(Warnings.begin(), "matrix-conversion"); + + // HLSL booleans in memory are i32; any non-zero value is true. + // Use icmp-ne rather than truncation to avoid InstCombine rewriting + // trunc i32 to i1 into an illegal wide <N x i1> bitcast (issue #140824). + if (Args.hasArg(OPT_load_bool_from_mem) && + CodeGenOpts.getLoadBoolFromMem() != + CodeGenOptions::BoolFromMem::NonZero) + Diags.Report(diag::warn_hlsl_bool_from_mem) + << Args.getLastArg(OPT_load_bool_from_mem)->getValue(); + CodeGenOpts.setLoadBoolFromMem(CodeGenOptions::BoolFromMem::NonZero); } // When these options are used, the compiler is allowed to apply diff --git a/clang/test/CodeGenHLSL/bool-vector-load.hlsl b/clang/test/CodeGenHLSL/bool-vector-load.hlsl new file mode 100644 index 0000000000000..c3e6c868bba3e --- /dev/null +++ b/clang/test/CodeGenHLSL/bool-vector-load.hlsl @@ -0,0 +1,19 @@ +// RUN: %clang_dxc -T cs_6_3 %s | FileCheck %s +// RUN: %clang_dxc -T cs_6_3 -Xclang -load-bool-from-mem=truncate %s 2>&1 | FileCheck %s --check-prefix=CHECK,TRUNCATE + +// TRUNCATE: -load-bool-from-mem=truncate ignored for HLSL (always nonzero) + +StructuredBuffer<bool4> In : register(t0); +RWStructuredBuffer<bool4> Out : register(u1); + +// CHECK-LABEL: define {{.*}}main +// CHECK: [[Load:%.*]] = call {{[^ ]*}} @dx.op.rawBufferLoad.i32(i32 139, %dx.types.Handle %{{[^,]*}}, i32 0, i32 12, i8 1, i32 4) +// CHECK: [[Elem:%.*]] = extractvalue {{.*}} [[Load]], 0 +// CHECK-NEXT: [[Bool:%.*]] = icmp ne i32 [[Elem]], 0 +// CHECK: [[Wide:%.*]] = zext i1 [[Bool]] to i32 +// CHECK: call void @dx.op.rawBufferStore.i32(i32 140, %dx.types.Handle %{{[^,]*}},{{[^,]*}},{{[^,]*}}, i32 [[Wide]] +[numthreads(1,1,1)] +void main(void) { + Out[1].w = In[0].w; +} + diff --git a/clang/test/CodeGenHLSL/builtins/BoolSwizzles.hlsl b/clang/test/CodeGenHLSL/builtins/BoolSwizzles.hlsl index 88058255e94c3..191a6ad1d604c 100644 --- a/clang/test/CodeGenHLSL/builtins/BoolSwizzles.hlsl +++ b/clang/test/CodeGenHLSL/builtins/BoolSwizzles.hlsl @@ -1,12 +1,10 @@ -// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -x hlsl -load-bool-from-mem=truncate -triple dxil-pc-shadermodel6.3-library %s -emit-llvm -disable-llvm-passes -o - | FileCheck -check-prefixes=CHECK,CHECK-TR %s -// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -x hlsl -load-bool-from-mem=nonzero -triple dxil-pc-shadermodel6.3-library %s -emit-llvm -disable-llvm-passes -o - | FileCheck -check-prefixes=CHECK,CHECK-NZ %s +// RUN: %clang_cc1 -std=hlsl2021 -finclude-default-header -x hlsl -load-bool-from-mem=nonzero -triple dxil-pc-shadermodel6.3-library %s -emit-llvm -disable-llvm-passes -o - | FileCheck %s // CHECK-LABEL: ToFourBools // CHECK: {{%.*}} = zext i1 {{.*}} to i32 // CHECK: [[splat:%.*]] = insertelement <1 x i32> poison, i32 {{.*}}, i64 0 // CHECK-NEXT: [[vec4:%.*]] = shufflevector <1 x i32> [[splat]], <1 x i32> poison, <4 x i32> zeroinitializer -// CHECK-NZ-NEXT: [[vec2Ret:%.*]] = icmp ne <4 x i32> [[vec4]], zeroinitializer -// CHECK-TR-NEXT: [[vec2Ret:%.*]] = trunc <4 x i32> [[vec4]] to <4 x i1> +// CHECK-NEXT: [[vec2Ret:%.*]] = icmp ne <4 x i32> [[vec4]], zeroinitializer // CHECK-NEXT: ret <4 x i1> [[vec2Ret]] bool4 ToFourBools(bool V) { return V.rrrr; @@ -17,8 +15,7 @@ bool4 ToFourBools(bool V) { // CHECK-NEXT: store <1 x i32> splat (i32 1), ptr [[Tmp]], align 4 // CHECK-NEXT: [[Vec1:%.*]] = load <1 x i32>, ptr [[Tmp]], align 4 // CHECK-NEXT: [[Vec2:%.*]] = shufflevector <1 x i32> [[Vec1]], <1 x i32> poison, <2 x i32> zeroinitializer -// CHECK-NZ-NEXT: [[Vec2Ret:%.*]] = icmp ne <2 x i32> [[Vec2]], zeroinitializer -// CHECK-TR-NEXT: [[Vec2Ret:%.*]] = trunc <2 x i32> [[Vec2]] to <2 x i1> +// CHECK-NEXT: [[Vec2Ret:%.*]] = icmp ne <2 x i32> [[Vec2]], zeroinitializer // CHECK-NEXT: ret <2 x i1> [[Vec2Ret]] bool2 FillTrue() { return true.xx; @@ -32,14 +29,12 @@ bool2 FillTrue() { // CHECK-NEXT: [[VVal:%.*]] = load i32, ptr [[VAddr]], align 4 // CHECK-NEXT: [[Splat:%.*]] = insertelement <1 x i32> poison, i32 [[VVal]], i64 0 // CHECK-NEXT: [[Vec2:%.*]] = shufflevector <1 x i32> [[Splat]], <1 x i32> poison, <2 x i32> zeroinitializer -// CHECK-NZ-NEXT: [[Trunc:%.*]] = icmp ne <2 x i32> [[Vec2]], zeroinitializer -// CHECK-TR-NEXT: [[Trunc:%.*]] = trunc <2 x i32> [[Vec2]] to <2 x i1> +// CHECK-NEXT: [[Trunc:%.*]] = icmp ne <2 x i32> [[Vec2]], zeroinitializer // CHECK-NEXT: [[Ext:%.*]] = zext <2 x i1> [[Trunc]] to <2 x i32> // CHECK-NEXT: store <2 x i32> [[Ext]], ptr [[Vec2Ptr]], align 4 // CHECK-NEXT: [[V2:%.*]] = load <2 x i32>, ptr [[Vec2Ptr]], align 4 // CHECK-NEXT: [[V3:%.*]] = shufflevector <2 x i32> [[V2]], <2 x i32> poison, <2 x i32> zeroinitializer -// CHECK-NZ-NEXT: [[LV1:%.*]] = icmp ne <2 x i32> [[V3]], zeroinitializer -// CHECK-TR-NEXT: [[LV1:%.*]] = trunc <2 x i32> [[V3]] to <2 x i1> +// CHECK-NEXT: [[LV1:%.*]] = icmp ne <2 x i32> [[V3]], zeroinitializer // CHECK-NEXT: ret <2 x i1> [[LV1]] bool2 HowManyBools(bool V) { return V.rr.rr; @@ -53,23 +48,20 @@ bool2 HowManyBools(bool V) { // CHECK-NEXT: [[X:%.*]] = load i32, ptr [[VAddr]], align 4 // CHECK-NEXT: [[Splat:%.*]] = insertelement <1 x i32> poison, i32 [[X]], i64 0 // CHECK-NEXT: [[Y:%.*]] = extractelement <1 x i32> [[Splat]], i32 0 -// CHECK-NZ-NEXT: [[Z:%.*]] = icmp ne i32 [[Y]], 0 -// CHECK-TR-NEXT: [[Z:%.*]] = trunc i32 [[Y]] to i1 +// CHECK-NEXT: [[Z:%.*]] = icmp ne i32 [[Y]], 0 // CHECK-NEXT: [[A:%.*]] = zext i1 [[Z]] to i32 // CHECK-NEXT: store i32 [[A]], ptr [[XAddr]], align 4 // CHECK-NEXT: [[B:%.*]] = load i32, ptr [[VAddr]], align 4 // CHECK-NEXT: [[Splat2:%.*]] = insertelement <1 x i32> poison, i32 [[B]], i64 0 // CHECK-NEXT: [[C:%.*]] = extractelement <1 x i32> [[Splat2]], i32 0 -// CHECK-NZ-NEXT: [[D:%.*]] = icmp ne i32 [[C]], 0 -// CHECK-TR-NEXT: [[D:%.*]] = trunc i32 [[C]] to i1 +// CHECK-NEXT: [[D:%.*]] = icmp ne i32 [[C]], 0 // CHECK-NEXT: br i1 [[D]], label %lor.end, label %lor.rhs // CHECK: lor.rhs: // CHECK-NEXT: [[E:%.*]] = load i32, ptr [[VAddr]], align 4 // CHECK-NEXT: [[Splat3:%.*]] = insertelement <1 x i32> poison, i32 [[E]], i64 0 // CHECK-NEXT: [[F:%.*]] = extractelement <1 x i32> [[Splat3]], i32 0 -// CHECK-NZ-NEXT: [[G:%.*]] = icmp ne i32 [[F]], 0 -// CHECK-TR-NEXT: [[G:%.*]] = trunc i32 [[F]] to i1 +// CHECK-NEXT: [[G:%.*]] = icmp ne i32 [[F]], 0 // CHECK-NEXT: br label %lor.end // CHECK: lor.end: @@ -77,8 +69,7 @@ bool2 HowManyBools(bool V) { // CHECK-NEXT: [[J:%.*]] = zext i1 [[H]] to i32 // CHECK-NEXT: store i32 [[J]], ptr [[XAddr]], align 4 // CHECK-NEXT: [[I:%.*]] = load i32, ptr [[XAddr]], align 4 -// CHECK-NZ-NEXT: [[LoadV:%.*]] = icmp ne i32 [[I]], 0 -// CHECK-TR-NEXT: [[LoadV:%.*]] = trunc i32 [[I]] to i1 +// CHECK-NEXT: [[LoadV:%.*]] = icmp ne i32 [[I]], 0 // CHECK-NEXT: ret i1 [[LoadV]] bool AssignBool(bool V) { bool X = V.x; @@ -95,13 +86,11 @@ bool AssignBool(bool V) { // CHECK-NEXT: store <1 x i32> splat (i32 1), ptr [[Tmp]], align 4 // CHECK-NEXT: [[Y:%.*]] = load <1 x i32>, ptr [[Tmp]], align 4 // CHECK-NEXT: [[Z:%.*]] = shufflevector <1 x i32> [[Y]], <1 x i32> poison, <2 x i32> zeroinitializer -// CHECK-NZ-NEXT: [[LV:%.*]] = icmp ne <2 x i32> [[Z]], zeroinitializer -// CHECK-TR-NEXT: [[LV:%.*]] = trunc <2 x i32> [[Z]] to <2 x i1> +// CHECK-NEXT: [[LV:%.*]] = icmp ne <2 x i32> [[Z]], zeroinitializer // CHECK-NEXT: [[A:%.*]] = zext <2 x i1> [[LV]] to <2 x i32> // CHECK-NEXT: store <2 x i32> [[A]], ptr [[X]], align 4 // CHECK-NEXT: [[B:%.*]] = load i32, ptr [[VAddr]], align 4 -// CHECK-NZ-NEXT: [[LV1:%.*]] = icmp ne i32 [[B]], 0 -// CHECK-TR-NEXT: [[LV1:%.*]] = trunc i32 [[B]] to i1 +// CHECK-NEXT: [[LV1:%.*]] = icmp ne i32 [[B]], 0 // CHECK-NEXT: [[D:%.*]] = zext i1 [[LV1]] to i32 // CHECK-NEXT: [[C:%.*]] = getelementptr <2 x i32>, ptr [[X]], i32 0, i32 1 // CHECK-NEXT: store i32 [[D]], ptr [[C]], align 4 @@ -118,8 +107,7 @@ void AssignBool2(bool V) { // CHECK-NEXT: store <2 x i32> [[Y]], ptr [[VAddr]], align 4 // CHECK-NEXT: store <2 x i32> splat (i32 1), ptr [[X]], align 4 // CHECK-NEXT: [[Z:%.*]] = load <2 x i32>, ptr [[VAddr]], align 4 -// CHECK-NZ-NEXT: [[LV:%.*]] = icmp ne <2 x i32> [[Z]], zeroinitializer -// CHECK-TR-NEXT: [[LV:%.*]] = trunc <2 x i32> [[Z]] to <2 x i1> +// CHECK-NEXT: [[LV:%.*]] = icmp ne <2 x i32> [[Z]], zeroinitializer // CHECK-NEXT: [[B:%.*]] = zext <2 x i1> [[LV]] to <2 x i32> // CHECK-NEXT: [[V1:%.*]] = extractelement <2 x i32> [[B]], i32 0 // CHECK-NEXT: store i32 [[V1]], ptr [[X]], align 4 @@ -139,14 +127,12 @@ void AssignBool3(bool2 V) { // CHECK-NEXT: store <1 x i32> splat (i32 1), ptr [[Tmp]], align 4 // CHECK-NEXT: [[Y:%.*]] = load <1 x i32>, ptr [[Tmp]], align 4 // CHECK-NEXT: [[Z:%.*]] = shufflevector <1 x i32> [[Y]], <1 x i32> poison, <4 x i32> zeroinitializer -// CHECK-NZ-NEXT: [[LV:%.*]] = icmp ne <4 x i32> [[Z]], zeroinitializer -// CHECK-TR-NEXT: [[LV:%.*]] = trunc <4 x i32> [[Vec2]] to <4 x i1> +// CHECK-NEXT: [[LV:%.*]] = icmp ne <4 x i32> [[Z]], zeroinitializer // CHECK-NEXT: [[A:%.*]] = zext <4 x i1> [[LV]] to <4 x i32> // CHECK-NEXT: store <4 x i32> [[A]], ptr [[X]], align 4 // CHECK-NEXT: [[B:%.*]] = load <4 x i32>, ptr [[X]], align 4 // CHECK-NEXT: [[C:%.*]] = shufflevector <4 x i32> [[B]], <4 x i32> poison, <2 x i32> <i32 2, i32 3> -// CHECK-NZ-NEXT: [[LV1:%.*]] = icmp ne <2 x i32> [[C]], zeroinitializer -// CHECK-TR-NEXT: [[LV1:%.*]] = trunc <2 x i32> [[C]] to <2 x i1> +// CHECK-NEXT: [[LV1:%.*]] = icmp ne <2 x i32> [[C]], zeroinitializer // CHECK-NEXT: ret <2 x i1> [[LV1]] bool2 AccessBools() { bool4 X = true.xxxx; @@ -160,8 +146,7 @@ bool2 AccessBools() { // CHECK-NEXT: store <1 x i32> zeroinitializer, ptr [[Tmp]], align 4 // CHECK-NEXT: [[L0:%.*]] = load <1 x i32>, ptr [[Tmp]], align 4 // CHECK-NEXT: [[L1:%.*]] = shufflevector <1 x i32> [[L0]], <1 x i32> poison, <3 x i32> zeroinitializer -// CHECK-NZ-NEXT: [[TruncV:%.*]] = icmp ne <3 x i32> [[L1]], zeroinitializer -// CHECK-TR-NEXT: [[TruncV:%.*]] = trunc <3 x i32> [[Vec2]] to <3 x i1> +// CHECK-NEXT: [[TruncV:%.*]] = icmp ne <3 x i32> [[L1]], zeroinitializer // CHECK-NEXT: [[L2:%.*]] = zext <3 x i1> [[TruncV]] to <3 x i32> // CHECK-NEXT: [[V1:%.*]] = extractelement <3 x i32> [[L2]], i32 0 // CHECK-NEXT: store i32 [[V1]], ptr %B, align 4 diff --git a/clang/test/CodeGenHLSL/strict-bool.hlsl b/clang/test/CodeGenHLSL/strict-bool.hlsl index 31867065e77a3..c9bacffc04469 100644 --- a/clang/test/CodeGenHLSL/strict-bool.hlsl +++ b/clang/test/CodeGenHLSL/strict-bool.hlsl @@ -1,15 +1,14 @@ -// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.3-library -O1 -load-bool-from-mem=truncate -emit-llvm -o - %s | FileCheck %s -check-prefixes=CHECK-TRUNCATE -// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.3-library -O1 -load-bool-from-mem=nonzero -emit-llvm -o - %s | FileCheck %s -check-prefixes=CHECK-NONZERO -// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.3-library -O0 -emit-llvm -o - %s | FileCheck %s -check-prefixes=CHECK-NONZERO +// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.3-library -O1 -load-bool-from-mem=truncate -emit-llvm -o - %s 2>&1 | FileCheck %s --check-prefix=CHECK,TRUNCATE +// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.3-library -O1 -load-bool-from-mem=nonzero -emit-llvm -o - %s | FileCheck %s +// RUN: %clang_cc1 -triple dxil-pc-shadermodel6.3-library -O0 -emit-llvm -o - %s | FileCheck %s + +// TRUNCATE: -load-bool-from-mem=truncate ignored for HLSL (always nonzero) typedef bool bool8_t __attribute__((ext_vector_type(8))); extern bool8_t vec; bool8_t getvec(void) { - // CHECK-TRUNCATE: [[BOOL:%.+]] = load <8 x i32> - // CHECK-TRUNCATE: trunc <8 x i32> [[BOOL]] to <8 x i1> - - // CHECK-NONZERO: [[BOOL:%.+]] = load <8 x i32> - // CHECK-NONZERO: icmp ne <8 x i32> [[BOOL]], zeroinitializer + // CHECK: [[BOOL:%.+]] = load <8 x i32> + // CHECK: icmp ne <8 x i32> [[BOOL]], zeroinitializer return vec; } diff --git a/clang/test/Driver/HLSL/load-bool-from-mem.hlsl b/clang/test/Driver/HLSL/load-bool-from-mem.hlsl new file mode 100644 index 0000000000000..23ee8cca8b6de --- /dev/null +++ b/clang/test/Driver/HLSL/load-bool-from-mem.hlsl @@ -0,0 +1,3 @@ +// RUN: %clang_dxc -T cs_6_3 -### %s 2>&1 | FileCheck %s + +// CHECK: "-load-bool-from-mem=nonzero" _______________________________________________ cfe-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
