Reintroduce Native Low Precision module flag, and refactor module flags

Author: Icohedron

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, 23, NativeLowPrecision, "Use native low precision")
+SHADER_FEATURE_FLAG(18, -1, NativeLowPrecision, "Native low precision data types")
 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")
@@ -117,6 +117,7 @@ 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,  NativeLowPrecisionMode, "Use native low precision")
 DXIL_MODULE_FLAG(33,  ResMayNotAlias, "Any UAV may not alias any other UAV")
 
 #undef DXIL_MODULE_FLAG

llvm/lib/Target/DirectX/DXILShaderFlags.cpp

@@ -143,7 +143,7 @@ void ModuleShaderFlags::updateFunctionFlags(ComputedShaderFlags &CSF,
   }
 
   if (CSF.LowPrecisionPresent) {
-    if (NativeLowPrecisionAvailable)
+    if (NativeLowPrecisionMode)
       CSF.NativeLowPrecision = true;
     else
       CSF.MinimumPrecision = true;
@@ -220,15 +220,15 @@ void ModuleShaderFlags::initialize(Module &M, DXILResourceTypeMap &DRTM,
     if (RMA->getValue() != 0)
       CanSetResMayNotAlias = false;
 
-  // NativeLowPrecision can only be set when the command line option
+  // NativeLowPrecisionMode 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;
+  NativeLowPrecisionMode = false;
   if (auto *NativeLowPrec = mdconst::extract_or_null<ConstantInt>(
           M.getModuleFlag("dx.nativelowprec")))
     if (MMDI.ShaderModelVersion >= VersionTuple(6, 2) &&
         NativeLowPrec->getValue() != 0)
-      NativeLowPrecisionAvailable = true;
+      NativeLowPrecisionMode = true;
 
   CallGraph CG(M);
 
@@ -254,11 +254,6 @@ void ModuleShaderFlags::initialize(Module &M, DXILResourceTypeMap &DRTM,
         continue;
       }
 
-      // Set ResMayNotAlias to true if DXIL validator version < 1.8 and there
-      // are UAVs present globally.
-      if (CanSetResMayNotAlias && MMDI.ValidatorVersion < VersionTuple(1, 8))
-        SCCSF.ResMayNotAlias = !DRM.uavs().empty();
-
       ComputedShaderFlags CSF;
       for (const auto &BB : *F)
         for (const auto &I : BB)
@@ -295,6 +290,17 @@ void ModuleShaderFlags::initialize(Module &M, DXILResourceTypeMap &DRTM,
             *(EntryFunProps.Entry), "Inconsistent optnone attribute "));
   }
 
+  // Set ResMayNotAlias to true if DXIL validator version < 1.8 and there
+  // are UAVs present globally.
+  if (CanSetResMayNotAlias && MMDI.ValidatorVersion < VersionTuple(1, 8))
+    CombinedSFMask.ResMayNotAlias = !DRM.uavs().empty();
+
+  // Set the module flag that enables native low-precision execution mode. This
+  // is needed even if the module does not use 16-bit types because a
+  // corresponding debug module may include 16-bit types, and tools that use the
+  // debug module may expect it to have the same flags as the original
+  CombinedSFMask.NativeLowPrecisionMode = NativeLowPrecisionMode;
+
   // Set the Max64UAVs flag if the number of UAVs is > 8
   uint32_t NumUAVs = 0;
   for (auto &UAV : DRM.uavs())

llvm/lib/Target/DirectX/DXILShaderFlags.h

@@ -91,8 +91,10 @@ struct ModuleShaderFlags {
   const ComputedShaderFlags &getCombinedFlags() const { return CombinedSFMask; }
 
 private:
-  bool CanSetResMayNotAlias;
-  bool NativeLowPrecisionAvailable;
+  // Booleans set by module flags
+  bool CanSetResMayNotAlias;   // dx.resmayalias
+  bool NativeLowPrecisionMode; // dx.nativelowprec
+
   /// 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

llvm/test/CodeGen/DirectX/ShaderFlags/res-may-not-alias-sm6.7.ll

@@ -2,8 +2,8 @@
 
 ; This test checks to ensure the behavior of the DXIL shader flag analysis
 ; for the flag ResMayNotAlias is correct when the DXIL Version is >= 1.7 and the
-; DXIL Validator Version < 1.8. The ResMayNotAlias flag (0x20000000) should be
-; set on all functions if there are one or more UAVs present globally in the
+; DXIL Validator Version < 1.8. The ResMayNotAlias module flag (0x20000000)
+; should be set if there are one or more UAVs present globally in the
 ; module.
 
 target triple = "dxil-pc-shadermodel6.7-library"
@@ -19,7 +19,7 @@ target triple = "dxil-pc-shadermodel6.7-library"
 ; CHECK:       Any UAV may not alias any other UAV
 ;
 
-; CHECK: Function loadUAV : 0x200000000
+; CHECK: Function loadUAV : 0x00000000
 define float @loadUAV() #0 {
   %res = call target("dx.TypedBuffer", float, 1, 0, 0)
       @llvm.dx.resource.handlefrombinding(i32 0, i32 0, i32 1, i32 0, i1 false)
@@ -29,7 +29,7 @@ define float @loadUAV() #0 {
   ret float %val
 }
 
-; CHECK: Function loadSRV : 0x200000010
+; CHECK: Function loadSRV : 0x00000010
 define float @loadSRV() #0 {
   %res = tail call target("dx.RawBuffer", float, 0, 0)
       @llvm.dx.resource.handlefrombinding(i32 0, i32 0, i32 1, i32 0, i1 false)

llvm/test/CodeGen/DirectX/ShaderFlags/use-native-low-precision-1.ll

@@ -11,28 +11,26 @@ 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: ;       Native low precision data types
 ;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
 
-;CHECK-LABEL: ; Function add_i16 : 0x00800020
+;CHECK-LABEL: ; Function add_i16 : 0x00000020
 define i16 @add_i16(i16 %a, i16 %b) "hlsl.export" {
   %sum = add i16 %a, %b
   ret i16 %sum
 }
 
-; 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
 }
 
-;CHECK-LABEL: ; Function add_half : 0x00800020
+;CHECK-LABEL: ; Function add_half : 0x00000020
 define half @add_half(half %a, half %b) "hlsl.export" {
   %sum = fadd half %a, %b
   ret half %sum