[llvm] [DirectX] Set shader feature flags MinimumPrecision and NativeLowPrecision (PR #139623)
Deric C. via llvm-commits
llvm-commits at lists.llvm.org
Tue May 13 11:45:38 PDT 2025
https://github.com/Icohedron updated https://github.com/llvm/llvm-project/pull/139623
>From 2039fa7a08f488f0029e31d0402447da0b142634 Mon Sep 17 00:00:00 2001
From: Icohedron <cheung.deric at gmail.com>
Date: Mon, 12 May 2025 20:58:36 +0000
Subject: [PATCH 1/4] Set shader feature flags MinimumPrecision and
NativeLowPrecision
---
llvm/lib/Target/DirectX/DXILShaderFlags.cpp | 8 ++++++++
.../DirectX/ShaderFlags/low-precision.ll | 20 ++++++++++++++++---
.../ShaderFlags/use-native-low-precision-0.ll | 18 +++++++++++++----
.../ShaderFlags/use-native-low-precision-1.ll | 19 +++++++++++++++---
4 files changed, 55 insertions(+), 10 deletions(-)
diff --git a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
index b50a9b5d6051c..36f909ffa6ef5 100644
--- a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
+++ b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
@@ -296,6 +296,14 @@ void ModuleShaderFlags::initialize(Module &M, DXILResourceTypeMap &DRTM,
if (NumUAVs > 8)
CombinedSFMask.Max64UAVs = true;
+ // Set the Shader Feature Info flags related to low-precision datatypes
+ if (CombinedSFMask.LowPrecisionPresent) {
+ if (CombinedSFMask.UseNativeLowPrecision)
+ CombinedSFMask.NativeLowPrecision = true;
+ else
+ CombinedSFMask.MinimumPrecision = true;
+ }
+
CombinedSFMask.UAVsAtEveryStage = hasUAVsAtEveryStage(DRM, MMDI);
}
diff --git a/llvm/test/CodeGen/DirectX/ShaderFlags/low-precision.ll b/llvm/test/CodeGen/DirectX/ShaderFlags/low-precision.ll
index 561e09bb1e9dc..bccc14d2b38cb 100644
--- a/llvm/test/CodeGen/DirectX/ShaderFlags/low-precision.ll
+++ b/llvm/test/CodeGen/DirectX/ShaderFlags/low-precision.ll
@@ -1,4 +1,10 @@
; RUN: opt -S --passes="print-dx-shader-flags" 2>&1 %s | FileCheck %s
+; RUN: llc %s --filetype=obj -o - | obj2yaml | FileCheck %s --check-prefix=DXC
+
+; Check that when the dx.nativelowprec module flag is not specified, the
+; module-level shader flag UseNativeLowPrecision is not set, and the
+; MinimumPrecision feature flag is set due to the presence of half and i16
+; without native low precision.
target triple = "dxil-pc-shadermodel6.7-library"
@@ -6,25 +12,33 @@ target triple = "dxil-pc-shadermodel6.7-library"
;CHECK-NEXT: ; Shader Flags Value: 0x00000020
;CHECK-NEXT: ;
;CHECK-NEXT: ; Note: shader requires additional functionality:
+;CHECK-NEXT: ; Minimum-precision data types
;CHECK-NEXT: ; Note: extra DXIL module flags:
;CHECK-NEXT: ; Low-precision data types
;CHECK-NEXT: ;
;CHECK-NEXT: ; Shader Flags for Module Functions
;CHECK-LABEL: ; Function add_i16 : 0x00000020
-define i16 @add_i16(i16 %a, i16 %b) {
+define i16 @add_i16(i16 %a, i16 %b) "hlsl.export" {
%sum = add i16 %a, %b
ret i16 %sum
}
;CHECK-LABEL: ; Function add_i32 : 0x00000000
-define i32 @add_i32(i32 %a, i32 %b) {
+define i32 @add_i32(i32 %a, i32 %b) "hlsl.export" {
%sum = add i32 %a, %b
ret i32 %sum
}
;CHECK-LABEL: ; Function add_half : 0x00000020
-define half @add_half(half %a, half %b) {
+define half @add_half(half %a, half %b) "hlsl.export" {
%sum = fadd half %a, %b
ret half %sum
}
+
+; DXC: - Name: SFI0
+; DXC-NEXT: Size: 8
+; DXC-NEXT: Flags:
+; DXC: MinimumPrecision: true
+; DXC: NativeLowPrecision: false
+; DXC: ...
diff --git a/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-0.ll b/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-0.ll
index c537a01482f39..e4d55d3d89013 100644
--- a/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-0.ll
+++ b/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-0.ll
@@ -1,7 +1,9 @@
; RUN: opt -S --passes="print-dx-shader-flags" 2>&1 %s | FileCheck %s
+; RUN: llc %s --filetype=obj -o - | obj2yaml | FileCheck %s --check-prefix=DXC
; Check that when the dx.nativelowprec module flag is set to 0, the module-level
-; shader flag UseNativeLowPrecision is not set
+; shader flag UseNativeLowPrecision is not set, and the MinimumPrecision feature
+; flag is set due to the presence of half and i16 without native low precision.
target triple = "dxil-pc-shadermodel6.7-library"
@@ -9,6 +11,7 @@ target triple = "dxil-pc-shadermodel6.7-library"
;CHECK-NEXT: ; Shader Flags Value: 0x00000020
;CHECK-NEXT: ;
;CHECK-NEXT: ; Note: shader requires additional functionality:
+;CHECK-NEXT: ; Minimum-precision data types
;CHECK-NEXT: ; Note: extra DXIL module flags:
;CHECK-NEXT: ; Low-precision data types
;CHECK-NOT: ; Native 16bit types enabled
@@ -16,22 +19,29 @@ target triple = "dxil-pc-shadermodel6.7-library"
;CHECK-NEXT: ; Shader Flags for Module Functions
;CHECK-LABEL: ; Function add_i16 : 0x00000020
-define i16 @add_i16(i16 %a, i16 %b) {
+define i16 @add_i16(i16 %a, i16 %b) "hlsl.export" {
%sum = add i16 %a, %b
ret i16 %sum
}
;CHECK-LABEL: ; Function add_i32 : 0x00000000
-define i32 @add_i32(i32 %a, i32 %b) {
+define i32 @add_i32(i32 %a, i32 %b) "hlsl.export" {
%sum = add i32 %a, %b
ret i32 %sum
}
;CHECK-LABEL: ; Function add_half : 0x00000020
-define half @add_half(half %a, half %b) {
+define half @add_half(half %a, half %b) "hlsl.export" {
%sum = fadd half %a, %b
ret half %sum
}
!llvm.module.flags = !{!0}
!0 = !{i32 1, !"dx.nativelowprec", i32 0}
+
+; DXC: - Name: SFI0
+; DXC-NEXT: Size: 8
+; DXC-NEXT: Flags:
+; DXC: MinimumPrecision: true
+; DXC: NativeLowPrecision: false
+; DXC: ...
diff --git a/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-1.ll b/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-1.ll
index 07c4b9064d666..6ee592de087a6 100644
--- a/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-1.ll
+++ b/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-1.ll
@@ -1,4 +1,9 @@
; RUN: opt -S --passes="print-dx-shader-flags" 2>&1 %s | FileCheck %s
+; RUN: llc %s --filetype=obj -o - | obj2yaml | FileCheck %s --check-prefix=DXC
+
+; Check that when the dx.nativelowprec module flag is set to 1, the module-level
+; shader flag UseNativeLowPrecision is set, and the NativeLowPrecision feature
+; flag is set
target triple = "dxil-pc-shadermodel6.7-library"
@@ -6,6 +11,7 @@ target triple = "dxil-pc-shadermodel6.7-library"
;CHECK-NEXT: ; Shader Flags Value: 0x00800020
;CHECK-NEXT: ;
;CHECK-NEXT: ; Note: shader requires additional functionality:
+;CHECK-NEXT: ; Use native low precision
;CHECK-NEXT: ; Note: extra DXIL module flags:
;CHECK-NEXT: ; Low-precision data types
;CHECK-NEXT: ; Use native low precision
@@ -13,7 +19,7 @@ target triple = "dxil-pc-shadermodel6.7-library"
;CHECK-NEXT: ; Shader Flags for Module Functions
;CHECK-LABEL: ; Function add_i16 : 0x00800020
-define i16 @add_i16(i16 %a, i16 %b) {
+define i16 @add_i16(i16 %a, i16 %b) "hlsl.export" {
%sum = add i16 %a, %b
ret i16 %sum
}
@@ -22,16 +28,23 @@ define i16 @add_i16(i16 %a, i16 %b) {
; in the module regardless of whether or not the function uses low precision
; data types (flag 0x20). This matches the behavior in DXC
;CHECK-LABEL: ; Function add_i32 : 0x00800000
-define i32 @add_i32(i32 %a, i32 %b) {
+define i32 @add_i32(i32 %a, i32 %b) "hlsl.export" {
%sum = add i32 %a, %b
ret i32 %sum
}
;CHECK-LABEL: ; Function add_half : 0x00800020
-define half @add_half(half %a, half %b) {
+define half @add_half(half %a, half %b) "hlsl.export" {
%sum = fadd half %a, %b
ret half %sum
}
!llvm.module.flags = !{!0}
!0 = !{i32 1, !"dx.nativelowprec", i32 1}
+
+; DXC: - Name: SFI0
+; DXC-NEXT: Size: 8
+; DXC-NEXT: Flags:
+; DXC: MinimumPrecision: false
+; DXC: NativeLowPrecision: true
+; DXC: ...
>From aa824aca6fe81c2a3528edd41a8db932087aa667 Mon Sep 17 00:00:00 2001
From: Icohedron <cheung.deric at gmail.com>
Date: Tue, 13 May 2025 17:40:29 +0000
Subject: [PATCH 2/4] Refactor flags related to low precision data types and
combine the NativeLowPrecision module and feature flags
---
.../BinaryFormat/DXContainerConstants.def | 3 +-
llvm/lib/Target/DirectX/DXILShaderFlags.cpp | 35 +++++++++----------
llvm/lib/Target/DirectX/DXILShaderFlags.h | 1 +
.../ShaderFlags/use-native-low-precision-1.ll | 9 +++--
4 files changed, 23 insertions(+), 25 deletions(-)
diff --git a/llvm/include/llvm/BinaryFormat/DXContainerConstants.def b/llvm/include/llvm/BinaryFormat/DXContainerConstants.def
index 1645018aebedb..ed259524570a1 100644
--- a/llvm/include/llvm/BinaryFormat/DXContainerConstants.def
+++ b/llvm/include/llvm/BinaryFormat/DXContainerConstants.def
@@ -34,7 +34,7 @@ SHADER_FEATURE_FLAG(14, 19, WaveOps, "Wave level operations")
SHADER_FEATURE_FLAG(15, 20, Int64Ops, "64-Bit integer")
SHADER_FEATURE_FLAG(16, 21, ViewID, "View Instancing")
SHADER_FEATURE_FLAG(17, 22, Barycentrics, "Barycentrics")
-SHADER_FEATURE_FLAG(18, -1, NativeLowPrecision, "Use native low precision")
+SHADER_FEATURE_FLAG(18, 23, NativeLowPrecision, "Use native low precision")
SHADER_FEATURE_FLAG(19, 24, ShadingRate, "Shading Rate")
SHADER_FEATURE_FLAG(20, 25, Raytracing_Tier_1_1, "Raytracing tier 1.1 features")
SHADER_FEATURE_FLAG(21, 26, SamplerFeedback, "Sampler feedback")
@@ -117,7 +117,6 @@ DXIL_MODULE_FLAG( 3, ForceEarlyDepthStencil, "Force early depth-stencil test")
DXIL_MODULE_FLAG( 4, EnableRawAndStructuredBuffers, "Raw and structured buffers")
DXIL_MODULE_FLAG( 5, LowPrecisionPresent, "Low-precision data types")
DXIL_MODULE_FLAG( 8, AllResourcesBound, "All resources bound for the duration of shader execution")
-DXIL_MODULE_FLAG(23, UseNativeLowPrecision, "Use native low precision")
DXIL_MODULE_FLAG(33, ResMayNotAlias, "Any UAV may not alias any other UAV")
#undef DXIL_MODULE_FLAG
diff --git a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
index 36f909ffa6ef5..dd9cd8849525f 100644
--- a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
+++ b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
@@ -142,6 +142,13 @@ void ModuleShaderFlags::updateFunctionFlags(ComputedShaderFlags &CSF,
}
}
+ if (CSF.LowPrecisionPresent) {
+ if (NativeLowPrecisionAvailable)
+ CSF.NativeLowPrecision = true;
+ else
+ CSF.MinimumPrecision = true;
+ }
+
if (!CSF.Int64Ops)
CSF.Int64Ops = I.getType()->isIntegerTy(64);
@@ -206,14 +213,21 @@ void ModuleShaderFlags::initialize(Module &M, DXILResourceTypeMap &DRTM,
const ModuleMetadataInfo &MMDI) {
CanSetResMayNotAlias = MMDI.DXILVersion >= VersionTuple(1, 7);
-
- // Check if -res-may-alias was provided on the command line.
- // The command line option will set the dx.resmayalias module flag to 1.
+ // The command line option -res-may-alias will set the dx.resmayalias module
+ // flag to 1 and disable the ability to set the ResMayNotAlias flag
if (auto *RMA = mdconst::extract_or_null<ConstantInt>(
M.getModuleFlag("dx.resmayalias")))
if (RMA->getValue() != 0)
CanSetResMayNotAlias = false;
+ // Set UseNativeLowPrecision using the dx.nativelowprec module flag set by
+ // the command line option -enable-16bit-types
+ if (auto *NativeLowPrec = mdconst::extract_or_null<ConstantInt>(
+ M.getModuleFlag("dx.nativelowprec")))
+ if (MMDI.ShaderModelVersion >= VersionTuple(6, 2) &&
+ NativeLowPrec->getValue() != 0)
+ NativeLowPrecisionAvailable = true;
+
CallGraph CG(M);
// Compute Shader Flags Mask for all functions using post-order visit of SCC
@@ -243,13 +257,6 @@ void ModuleShaderFlags::initialize(Module &M, DXILResourceTypeMap &DRTM,
if (CanSetResMayNotAlias && MMDI.ValidatorVersion < VersionTuple(1, 8))
SCCSF.ResMayNotAlias = !DRM.uavs().empty();
- // Set UseNativeLowPrecision using dx.nativelowprec module metadata
- if (auto *NativeLowPrec = mdconst::extract_or_null<ConstantInt>(
- M.getModuleFlag("dx.nativelowprec")))
- if (MMDI.ShaderModelVersion >= VersionTuple(6, 2) &&
- NativeLowPrec->getValue() != 0)
- SCCSF.UseNativeLowPrecision = true;
-
ComputedShaderFlags CSF;
for (const auto &BB : *F)
for (const auto &I : BB)
@@ -296,14 +303,6 @@ void ModuleShaderFlags::initialize(Module &M, DXILResourceTypeMap &DRTM,
if (NumUAVs > 8)
CombinedSFMask.Max64UAVs = true;
- // Set the Shader Feature Info flags related to low-precision datatypes
- if (CombinedSFMask.LowPrecisionPresent) {
- if (CombinedSFMask.UseNativeLowPrecision)
- CombinedSFMask.NativeLowPrecision = true;
- else
- CombinedSFMask.MinimumPrecision = true;
- }
-
CombinedSFMask.UAVsAtEveryStage = hasUAVsAtEveryStage(DRM, MMDI);
}
diff --git a/llvm/lib/Target/DirectX/DXILShaderFlags.h b/llvm/lib/Target/DirectX/DXILShaderFlags.h
index 0e0bd0036349e..5fa5cf48c16f8 100644
--- a/llvm/lib/Target/DirectX/DXILShaderFlags.h
+++ b/llvm/lib/Target/DirectX/DXILShaderFlags.h
@@ -92,6 +92,7 @@ struct ModuleShaderFlags {
private:
bool CanSetResMayNotAlias;
+ bool NativeLowPrecisionAvailable;
/// Map of Function-Shader Flag Mask pairs representing properties of each of
/// the functions in the module. Shader Flags of each function represent both
/// module-level and function-level flags
diff --git a/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-1.ll b/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-1.ll
index 6ee592de087a6..5cc15be48f4a9 100644
--- a/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-1.ll
+++ b/llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-1.ll
@@ -14,7 +14,6 @@ target triple = "dxil-pc-shadermodel6.7-library"
;CHECK-NEXT: ; Use native low precision
;CHECK-NEXT: ; Note: extra DXIL module flags:
;CHECK-NEXT: ; Low-precision data types
-;CHECK-NEXT: ; Use native low precision
;CHECK-NEXT: ;
;CHECK-NEXT: ; Shader Flags for Module Functions
@@ -24,10 +23,10 @@ define i16 @add_i16(i16 %a, i16 %b) "hlsl.export" {
ret i16 %sum
}
-; NOTE: The flag for native low precision (0x80000) is set for every function
-; in the module regardless of whether or not the function uses low precision
-; data types (flag 0x20). This matches the behavior in DXC
-;CHECK-LABEL: ; Function add_i32 : 0x00800000
+; NOTE: In DXC, the flag for native low precision (0x80000) is set for every
+; function in the module regardless of whether or not the function uses low
+; precision data types (flag 0x20). However, this will not be the case for Clang
+;CHECK-LABEL: ; Function add_i32 : 0x00000000
define i32 @add_i32(i32 %a, i32 %b) "hlsl.export" {
%sum = add i32 %a, %b
ret i32 %sum
>From 4cb533b474a7e6f7dc5fd207561a6e2cfb3b4b64 Mon Sep 17 00:00:00 2001
From: Icohedron <cheung.deric at gmail.com>
Date: Tue, 13 May 2025 17:51:22 +0000
Subject: [PATCH 3/4] Fix comment about the -enable-16bit-types flag
---
llvm/lib/Target/DirectX/DXILShaderFlags.cpp | 5 +++--
1 file changed, 3 insertions(+), 2 deletions(-)
diff --git a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
index dd9cd8849525f..11f3a1e4a8c18 100644
--- a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
+++ b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
@@ -220,8 +220,9 @@ void ModuleShaderFlags::initialize(Module &M, DXILResourceTypeMap &DRTM,
if (RMA->getValue() != 0)
CanSetResMayNotAlias = false;
- // Set UseNativeLowPrecision using the dx.nativelowprec module flag set by
- // the command line option -enable-16bit-types
+ // NativeLowPrecision can only be set when the command line option
+ // -enable-16bit-types is provided. This is indicated by the dx.nativelowprec
+ // module flag being set
if (auto *NativeLowPrec = mdconst::extract_or_null<ConstantInt>(
M.getModuleFlag("dx.nativelowprec")))
if (MMDI.ShaderModelVersion >= VersionTuple(6, 2) &&
>From 610f4357502b9edb29ea69ba37e80039374b3f30 Mon Sep 17 00:00:00 2001
From: Icohedron <cheung.deric at gmail.com>
Date: Tue, 13 May 2025 18:45:26 +0000
Subject: [PATCH 4/4] Set default value of NativeLowPrecisionAvailable to false
---
llvm/lib/Target/DirectX/DXILShaderFlags.cpp | 1 +
1 file changed, 1 insertion(+)
diff --git a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
index 11f3a1e4a8c18..c359e3036b6cd 100644
--- a/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
+++ b/llvm/lib/Target/DirectX/DXILShaderFlags.cpp
@@ -223,6 +223,7 @@ void ModuleShaderFlags::initialize(Module &M, DXILResourceTypeMap &DRTM,
// NativeLowPrecision can only be set when the command line option
// -enable-16bit-types is provided. This is indicated by the dx.nativelowprec
// module flag being set
+ NativeLowPrecisionAvailable = false;
if (auto *NativeLowPrec = mdconst::extract_or_null<ConstantInt>(
M.getModuleFlag("dx.nativelowprec")))
if (MMDI.ShaderModelVersion >= VersionTuple(6, 2) &&
More information about the llvm-commits
mailing list