MSL: Treat pointer to vector as scalar type when bitcasting

Otherwise, they will be misregarded as vectors and can be used in as_type, which does not support pointer type as destination.
KhronosGroup · Mar 4, 2025 · 764abd5 · 764abd5
1 parent 2c32b6b
commit 764abd5
Show file tree

Hide file tree

Showing 4 changed files with 259 additions and 1 deletion.
diff --git a/reference/opt/shaders-msl/asm/frag/bitcast-ptr-vec.argument-tier-1.msl23.asm.frag b/reference/opt/shaders-msl/asm/frag/bitcast-ptr-vec.argument-tier-1.msl23.asm.frag
@@ -0,0 +1,67 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+template<typename T>
+struct spvDescriptor
+{
+    T value;
+};
+
+template<typename T>
+struct spvDescriptorArray
+{
+    spvDescriptorArray(const device spvDescriptor<T>* ptr) : ptr(&ptr->value)
+    {
+    }
+    const device T& operator [] (size_t i) const
+    {
+        return ptr[i];
+    }
+    const device T* ptr;
+};
+
+struct type_ConstantBuffer_PushConstants
+{
+    ulong VertexShaderConstants;
+    ulong PixelShaderConstants;
+    ulong SharedConstants;
+};
+
+struct main0_out
+{
+    float4 out_var_SV_Target [[color(0)]];
+};
+
+fragment main0_out main0(constant type_ConstantBuffer_PushConstants& g_PushConstants [[buffer(0)]], const device spvDescriptor<texture2d<float>>* g_Texture2DDescriptorHeap_ [[buffer(1)]], float4 gl_FragCoord [[position]])
+{
+    spvDescriptorArray<texture2d<float>> g_Texture2DDescriptorHeap {g_Texture2DDescriptorHeap_};
+
+    main0_out out = {};
+    int2 _55 = int2(gl_FragCoord.xy) - (*(reinterpret_cast<device int2*>(g_PushConstants.SharedConstants + 16ul)));
+    bool _66;
+    if (!any(_55 < int2(0)))
+    {
+        _66 = any(_55 >= (*(reinterpret_cast<device int2*>(g_PushConstants.SharedConstants + 24ul))));
+    }
+    else
+    {
+        _66 = true;
+    }
+    float4 _77;
+    if (_66)
+    {
+        _77 = float4(0.0);
+    }
+    else
+    {
+        _77 = g_Texture2DDescriptorHeap[*(reinterpret_cast<device uint*>(g_PushConstants.SharedConstants + 12ul))].read(uint2(int3(select(_55, int2(0), bool2(_66)), 0).xy), 0);
+    }
+    float3 _81 = powr(_77.xyz, *(reinterpret_cast<device float3*>(g_PushConstants.SharedConstants)));
+    out.out_var_SV_Target = float4(_81.x, _81.y, _81.z, _77.w);
+    return out;
+}
+
diff --git a/reference/shaders-msl/asm/frag/bitcast-ptr-vec.argument-tier-1.msl23.asm.frag b/reference/shaders-msl/asm/frag/bitcast-ptr-vec.argument-tier-1.msl23.asm.frag
@@ -0,0 +1,67 @@
+#pragma clang diagnostic ignored "-Wmissing-prototypes"
+
+#include <metal_stdlib>
+#include <simd/simd.h>
+
+using namespace metal;
+
+template<typename T>
+struct spvDescriptor
+{
+    T value;
+};
+
+template<typename T>
+struct spvDescriptorArray
+{
+    spvDescriptorArray(const device spvDescriptor<T>* ptr) : ptr(&ptr->value)
+    {
+    }
+    const device T& operator [] (size_t i) const
+    {
+        return ptr[i];
+    }
+    const device T* ptr;
+};
+
+struct type_ConstantBuffer_PushConstants
+{
+    ulong VertexShaderConstants;
+    ulong PixelShaderConstants;
+    ulong SharedConstants;
+};
+
+struct main0_out
+{
+    float4 out_var_SV_Target [[color(0)]];
+};
+
+fragment main0_out main0(constant type_ConstantBuffer_PushConstants& g_PushConstants [[buffer(0)]], const device spvDescriptor<texture2d<float>>* g_Texture2DDescriptorHeap_ [[buffer(1)]], float4 gl_FragCoord [[position]])
+{
+    spvDescriptorArray<texture2d<float>> g_Texture2DDescriptorHeap {g_Texture2DDescriptorHeap_};
+
+    main0_out out = {};
+    int2 _55 = int2(gl_FragCoord.xy) - (*(reinterpret_cast<device int2*>(g_PushConstants.SharedConstants + 16ul)));
+    bool _66;
+    if (!any(_55 < int2(0)))
+    {
+        _66 = any(_55 >= (*(reinterpret_cast<device int2*>(g_PushConstants.SharedConstants + 24ul))));
+    }
+    else
+    {
+        _66 = true;
+    }
+    float4 _77;
+    if (_66)
+    {
+        _77 = float4(0.0);
+    }
+    else
+    {
+        _77 = g_Texture2DDescriptorHeap[*(reinterpret_cast<device uint*>(g_PushConstants.SharedConstants + 12ul))].read(uint2(int3(select(_55, int2(0), bool2(_66)), 0).xy), 0);
+    }
+    float3 _81 = powr(_77.xyz, *(reinterpret_cast<device float3*>(g_PushConstants.SharedConstants)));
+    out.out_var_SV_Target = float4(_81.x, _81.y, _81.z, _77.w);
+    return out;
+}
+
diff --git a/shaders-msl/asm/frag/bitcast-ptr-vec.argument-tier-1.msl23.asm.frag b/shaders-msl/asm/frag/bitcast-ptr-vec.argument-tier-1.msl23.asm.frag
@@ -0,0 +1,124 @@
+; SPIR-V
+; Version: 1.0
+; Generator: Google spiregg; 0
+; Bound: 83
+; Schema: 0
+               OpCapability Shader
+               OpCapability Int64
+               OpCapability RuntimeDescriptorArray
+               OpCapability PhysicalStorageBufferAddresses
+               OpExtension "SPV_EXT_descriptor_indexing"
+               OpExtension "SPV_KHR_physical_storage_buffer"
+          %1 = OpExtInstImport "GLSL.std.450"
+               OpMemoryModel PhysicalStorageBuffer64 GLSL450
+               OpEntryPoint Fragment %2 "main" %3 %4
+               OpExecutionMode %2 OriginUpperLeft
+               OpSource HLSL 600
+               OpName %5 "type.ConstantBuffer.PushConstants"
+               OpMemberName %5 0 "VertexShaderConstants"
+               OpMemberName %5 1 "PixelShaderConstants"
+               OpMemberName %5 2 "SharedConstants"
+               OpName %6 "g_PushConstants"
+               OpName %7 "type.2d.image"
+               OpName %8 "g_Texture2DDescriptorHeap"
+               OpName %4 "out.var.SV_Target"
+               OpName %2 "main"
+               OpDecorate %3 BuiltIn FragCoord
+               OpDecorate %4 Location 0
+               OpDecorate %8 DescriptorSet 0
+               OpDecorate %8 Binding 0
+               OpMemberDecorate %5 0 Offset 0
+               OpMemberDecorate %5 1 Offset 8
+               OpMemberDecorate %5 2 Offset 16
+               OpDecorate %5 Block
+          %9 = OpTypeInt 32 0
+         %10 = OpTypeInt 32 1
+         %11 = OpConstant %10 2
+         %12 = OpTypeInt 64 0
+         %13 = OpConstant %12 12
+         %14 = OpConstant %12 16
+         %15 = OpConstant %10 0
+         %16 = OpTypeVector %10 2
+         %17 = OpConstantComposite %16 %15 %15
+         %18 = OpTypeBool
+         %19 = OpConstantTrue %18
+         %20 = OpConstant %12 24
+         %21 = OpTypeFloat 32
+         %22 = OpConstant %21 0
+         %23 = OpTypeVector %21 4
+         %24 = OpConstantComposite %23 %22 %22 %22 %22
+          %5 = OpTypeStruct %12 %12 %12
+         %25 = OpTypePointer PushConstant %5
+          %7 = OpTypeImage %21 2D 2 0 0 1 Unknown
+         %26 = OpTypeRuntimeArray %7
+         %27 = OpTypePointer UniformConstant %26
+         %28 = OpTypePointer Input %23
+         %29 = OpTypePointer Output %23
+         %30 = OpTypeVoid
+         %31 = OpTypeFunction %30
+         %32 = OpTypePointer PushConstant %12
+         %33 = OpTypePointer PhysicalStorageBuffer %9
+         %34 = OpTypePointer UniformConstant %7
+         %35 = OpTypeVector %21 2
+         %36 = OpTypePointer PhysicalStorageBuffer %16
+         %37 = OpTypeVector %18 2
+         %38 = OpTypeVector %10 3
+         %39 = OpTypeVector %21 3
+         %40 = OpTypePointer PhysicalStorageBuffer %39
+          %6 = OpVariable %25 PushConstant
+          %8 = OpVariable %27 UniformConstant
+          %3 = OpVariable %28 Input
+          %4 = OpVariable %29 Output
+          %2 = OpFunction %30 None %31
+         %41 = OpLabel
+         %42 = OpLoad %23 %3
+         %43 = OpAccessChain %32 %6 %11
+         %44 = OpLoad %12 %43
+         %45 = OpIAdd %12 %44 %13
+         %46 = OpBitcast %33 %45
+         %47 = OpLoad %9 %46 Aligned 4
+         %48 = OpAccessChain %34 %8 %47
+         %49 = OpLoad %7 %48
+         %50 = OpVectorShuffle %35 %42 %42 0 1
+         %51 = OpConvertFToS %16 %50
+         %52 = OpIAdd %12 %44 %14
+         %53 = OpBitcast %36 %52
+         %54 = OpLoad %16 %53 Aligned 4
+         %55 = OpISub %16 %51 %54
+         %56 = OpSLessThan %37 %55 %17
+         %57 = OpAny %18 %56
+         %58 = OpLogicalNot %18 %57
+               OpSelectionMerge %59 None
+               OpBranchConditional %58 %60 %59
+         %60 = OpLabel
+         %61 = OpIAdd %12 %44 %20
+         %62 = OpBitcast %36 %61
+         %63 = OpLoad %16 %62 Aligned 4
+         %64 = OpSGreaterThanEqual %37 %55 %63
+         %65 = OpAny %18 %64
+               OpBranch %59
+         %59 = OpLabel
+         %66 = OpPhi %18 %19 %41 %65 %60
+         %67 = OpCompositeConstruct %37 %66 %66
+         %68 = OpSelect %16 %67 %17 %55
+               OpSelectionMerge %69 None
+               OpBranchConditional %66 %70 %71
+         %70 = OpLabel
+               OpBranch %69
+         %71 = OpLabel
+         %72 = OpCompositeExtract %10 %68 0
+         %73 = OpCompositeExtract %10 %68 1
+         %74 = OpCompositeConstruct %38 %72 %73 %15
+         %75 = OpVectorShuffle %16 %74 %74 0 1
+         %76 = OpImageFetch %23 %49 %75 Lod %15
+               OpBranch %69
+         %69 = OpLabel
+         %77 = OpPhi %23 %24 %70 %76 %71
+         %78 = OpVectorShuffle %39 %77 %77 0 1 2
+         %79 = OpBitcast %40 %44
+         %80 = OpLoad %39 %79 Aligned 4
+         %81 = OpExtInst %39 %1 Pow %78 %80
+         %82 = OpVectorShuffle %23 %77 %81 4 5 6 3
+               OpStore %4 %82
+               OpReturn
+               OpFunctionEnd
diff --git a/spirv_msl.cpp b/spirv_msl.cpp
@@ -10008,7 +10008,7 @@ void CompilerMSL::emit_instruction(const Instruction &instruction)
 		{
 			string op;
 
-			if (type.vecsize == 1 && input_type.vecsize == 1)
+			if ((type.vecsize == 1 || type.pointer) && (input_type.vecsize == 1 || input_type.pointer))
 				op = join("reinterpret_cast<", type_to_glsl(type), ">(", to_unpacked_expression(ops[2]), ")");
 			else if (input_type.vecsize == 2)
 				op = join("reinterpret_cast<", type_to_glsl(type), ">(as_type<ulong>(", to_unpacked_expression(ops[2]), "))");