[SPIRV] Reimplement flat conversion.

tools/clang/lib/SPIRV/SpirvEmitter.h

@@ -759,10 +759,8 @@ class SpirvEmitter : public ASTConsumer {
 
 private:
   /// \brief Performs a FlatConversion implicit cast. Fills an instance of the
-  /// given type with initializer <result-id>. The initializer is of type
-  /// initType.
+  /// given type with initializer <result-id>.
   SpirvInstruction *processFlatConversion(const QualType type,
-                                          const QualType initType,
                                           SpirvInstruction *initId,
                                           SourceLocation,
                                           SourceRange range = {});
@@ -1248,6 +1246,26 @@ class SpirvEmitter : public ASTConsumer {
   /// Returns a function scope parameter with the same type as |param|.
   SpirvVariable *createFunctionScopeTempFromParameter(const ParmVarDecl *param);
 
+  /// Returns a vector of SpirvInstruction that is the decompostion of `inst`
+  /// into scalars. This is recursive. For example, a struct of a 4 element
+  /// vector will return 4 scalars.
+  std::vector<SpirvInstruction *> decomposeToScalars(SpirvInstruction *inst);
+
+  /// Returns a spirv instruction with the value of the given type and layout
+  /// rule that is obtained by assigning each scalar in `type` to corresponding
+  /// value in `scalars`. This is the inverse of `decomposeToScalars`.
+  SpirvInstruction *
+  generateFromScalars(QualType type, std::vector<SpirvInstruction *> &scalars,
+                      SpirvLayoutRule layoutRule);
+
+  /// Returns a spirv instruction with the value of the given type and layout
+  /// rule that is obtained by assigning `scalar` each scalar in `type`. This is
+  /// the same as calling `generateFromScalars` with a sufficiently large vector
+  /// where every element is `scalar`.
+  SpirvInstruction *splatScalarToGenerate(QualType type,
+                                          SpirvInstruction *scalar,
+                                          SpirvLayoutRule rule);
+
 public:
   /// \brief Wrapper method to create a fatal error message and report it
   /// in the diagnostic engine associated with this consumer.

tools/clang/test/CodeGenSPIRV/cast.flat-conversion.literal-initializer.hlsl

@@ -9,6 +9,16 @@ struct S {
   uint64_t f;
 };
 
+struct T {
+  int32_t i;
+  int64_t j;
+};
+
+struct UT {
+  uint32_t i;
+  uint64_t j;
+};
+
 void main() {
 
 // CHECK:              [[inf:%[0-9]+]] = OpFDiv %float %float_1 %float_0
@@ -47,4 +57,27 @@ void main() {
 // CHECK-NEXT:           {{%[0-9]+}} = OpCompositeConstruct %S [[a2_float]] [[a_float_0]] [[a2_double]] [[a]] [[a_int64]] [[a_uint64]]
   double a;
   S s1 = (S)(a);
+
+// TODO(6188): This is wrong because we lose most significant bits in the literal.
+// CHECK: [[lit:%[0-9]+]] = OpIAdd %int %int_0 %int_1
+// CHECK: [[longLit:%[0-9]+]] = OpSConvert %long [[lit]]
+// CHECK: [[t:%[0-9]+]] = OpCompositeConstruct %T [[lit]] [[longLit]]
+// CHECK: OpStore %t [[t]]
+  T t = (T)(0x100000000+1);
+
+// TODO(6188): This is wrong because we lose most significant bits in the literal.
+// CHECK: [[lit:%[0-9]+]] = OpIAdd %uint %uint_0 %uint_1
+// CHECK: [[longLit:%[0-9]+]] = OpUConvert %ulong [[lit]]
+// CHECK: [[t:%[0-9]+]] = OpCompositeConstruct %UT [[lit]] [[longLit]]
+// CHECK: OpStore %ut [[t]]
+  UT ut = (UT)(0x100000000ul+1);
+
+// TODO(6188): This is wrong because we lose most significant bits in the literal.
+// CHECK: [[longLit:%[0-9]+]] = OpIAdd %ulong %ulong_4294967296 %ulong_1
+// CHECK: [[lit:%[0-9]+]] = OpUConvert %uint [[longLit]]
+// CHECK: [[lit2:%[0-9]+]] = OpBitcast %int [[lit]]
+// CHECK: [[longLit2:%[0-9]+]] = OpBitcast %long [[longLit]]
+// CHECK: [[t:%[0-9]+]] = OpCompositeConstruct %T [[lit2]] [[longLit2]]
+// CHECK: OpStore %t2 [[t]]
+  T t2 = (T)(0x100000000ull+1);
 }

tools/clang/test/CodeGenSPIRV/cast.flat-conversion.matrix.hlsl

@@ -0,0 +1,52 @@
+// RUN: %dxc -T ps_6_0 -E main %s -spirv -Zpc | FileCheck %s -check-prefix=COL -check-prefix=CHECK
+// RUN: %dxc -T ps_6_0 -E main %s -spirv -Zpr | FileCheck %s -check-prefix=ROW -check-prefix=CHECK
+
+struct S {
+  float2x3 a;
+};
+
+struct T {
+  float a[6];
+};
+
+RWStructuredBuffer<S> s_output;
+RWStructuredBuffer<T> t_output;
+
+// See https://github.com/microsoft/DirectXShaderCompiler/blob/438781364eea22d188b337be1dfa4174ed7d928c/docs/SPIR-V.rst#L723.
+// COL: OpMemberDecorate %S 0 RowMajor
+// ROW: OpMemberDecorate %S 0 ColMajor
+
+// The DXIL generated for the two cases seem to produce the same results,
+// and this matches that behaviour.
+// CHECK: [[array_const:%[0-9]+]] = OpConstantComposite %_arr_float_uint_6 %float_0 %float_1 %float_2 %float_3 %float_4 %float_5
+// CHECK: [[t:%[0-9]+]] = OpConstantComposite %T [[array_const]]
+
+// The DXIL that is generates different order for the values depending on
+// whether the matrix is column or row major. However, for SPIR-V, the value
+// stored in both cases is the same because the decoration, which is checked
+// above, is what determines the layout in memory for the value.
+
+// CHECK: [[row0:%[0-9]+]] = OpConstantComposite %v3float %float_0 %float_1 %float_2
+// CHECK: [[row1:%[0-9]+]] = OpConstantComposite %v3float %float_3 %float_4 %float_5
+// CHECK: [[mat:%[0-9]+]] = OpConstantComposite %mat2v3float %33 %34
+// CHECK: [[s:%[0-9]+]] = OpConstantComposite %S %35
+
+void main() {
+  S s;
+  [unroll]
+  for( int i = 0; i < 2; ++i) {
+    [unroll]
+    for( int j = 0; j < 3; ++j) {
+      s.a[i][j] = i*3+j;
+    }
+  }
+// CHECK: [[ac:%[0-9]+]] = OpAccessChain %_ptr_Uniform_T %t_output %int_0 %uint_0
+// CHECK: OpStore [[ac]] [[t]]
+  T t = (T)(s);
+  t_output[0] = t;
+
+// CHECK: [[ac:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %s_output %int_0 %uint_0
+// CHECK: OpStore [[ac]] [[s]]
+  s = (S)t;
+  s_output[0] = s;
+}

tools/clang/test/CodeGenSPIRV/cast.flat-conversion.vector.hlsl

@@ -1,10 +1,24 @@
 // RUN: %dxc -T ps_6_0 -E main -fcgl  %s -spirv | FileCheck %s
 
-struct S {
+struct S1 {
     float2 data[2];
 };
 
-StructuredBuffer<S> MySB;
+StructuredBuffer<S1> MySB;
+
+struct S2 {
+  float b0;
+  float3 b1;
+};
+
+struct S3 {
+  float3 vec;
+};
+
+StructuredBuffer<float4> input2;
+
+
+
 
 float4 main() : SV_TARGET
 {
@@ -19,5 +33,28 @@ float4 main() : SV_TARGET
 // CHECK-NEXT:  [[val:%[0-9]+]] = OpCompositeConstruct %_arr_float_uint_4 [[v1]] [[v2]] [[v3]] [[v4]]
 // CHECK-NEXT:                 OpStore %data [[val]]
     float data[4] = (float[4])MySB[0].data;
-    return data[1];
+
+
+// CHECK: [[ac:%[0-9]+]] = OpAccessChain %_ptr_Uniform_v4float %input2 %int_0 %uint_0
+// CHECK-NEXT: [[ld:%[0-9]+]] = OpLoad %v4float [[ac]]
+// CHECK-NEXT: [[elem0:%[0-9]+]] = OpCompositeExtract %float [[ld]] 0
+// CHECK-NEXT: [[elem1:%[0-9]+]] = OpCompositeExtract %float [[ld]] 1
+// CHECK-NEXT: [[elem2:%[0-9]+]] = OpCompositeExtract %float [[ld]] 2
+// CHECK-NEXT: [[elem3:%[0-9]+]] = OpCompositeExtract %float [[ld]] 3
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeConstruct %v3float [[elem1]] [[elem2]] [[elem3]]
+// CHECK-NEXT: OpCompositeConstruct %S2 [[elem0]] [[vec]]
+    S2 d2 = (S2)input2[0];
+
+
+// CHECK: [[ac:%[0-9]+]] = OpAccessChain %_ptr_Uniform_v4float %input2 %int_0 %uint_0
+// CHECK-NEXT: [[ld:%[0-9]+]] = OpLoad %v4float [[ac]]
+// CHECK-NEXT: [[elem0:%[0-9]+]] = OpCompositeExtract %float [[ld]] 0
+// CHECK-NEXT: [[elem1:%[0-9]+]] = OpCompositeExtract %float [[ld]] 1
+// CHECK-NEXT: [[elem2:%[0-9]+]] = OpCompositeExtract %float [[ld]] 2
+// CHECK-NEXT: [[elem3:%[0-9]+]] = OpCompositeExtract %float [[ld]] 3
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeConstruct %v3float [[elem0]] [[elem1]] [[elem2]]
+// CHECK-NEXT: OpCompositeConstruct %S3 [[vec]]
+    S3 d3 = (S3)input2[0];
+
+    return 0;
 }

tools/clang/test/CodeGenSPIRV/cast.struct-to-int.hlsl

@@ -30,6 +30,7 @@ struct Vectors {
 
 RWStructuredBuffer<uint> buf : r0;
 RWStructuredBuffer<uint64_t> lbuf : r1;
+RWStructuredBuffer<Vectors> vbuf : r2;
 
 // CHECK: OpName [[BUF:%[^ ]*]] "buf"
 // CHECK: OpName [[LBUF:%[^ ]*]] "lbuf"
@@ -83,22 +84,21 @@ void main()
 // CHECK: [[COLORS:%[^ ]*]] = OpLoad [[TWOCOLORS]]
 // CHECK: [[COLORS0:%[^ ]*]] = OpCompositeExtract [[COLORRGBA]] [[COLORS]] 0
 // CHECK: [[COLORS00:%[^ ]*]] = OpCompositeExtract [[UINT]] [[COLORS0]] 0
-// CHECK: [[COLORS000:%[^ ]*]] = OpBitFieldUExtract [[UINT]] [[COLORS00]] [[U0]] [[U8]]
 // CHECK: [[BUF00:%[^ ]*]] = OpAccessChain %{{[^ ]*}} [[BUF]] [[I0]] [[U0]]
-// CHECK: OpStore [[BUF00]] [[COLORS000]]
+// CHECK: OpStore [[BUF00]] [[COLORS00]]
 
     buf[0] -= (uint) rgb;
 // CHECK: [[RGB:%[^ ]*]] = OpLoad [[COLORRGB]]
 // CHECK: [[RGB0:%[^ ]*]] = OpCompositeExtract [[UINT]] [[RGB]] 0
-// CHECK: [[RGB00:%[^ ]*]] = OpBitFieldUExtract [[UINT]] [[RGB0]] [[U0]] [[U8]]
 // CHECK: [[BUF00_0:%[^ ]*]] = OpAccessChain %{{[^ ]*}} [[BUF]] [[I0]] [[U0]]
 // CHECK: [[V1:%[^ ]*]] = OpLoad [[UINT]] [[BUF00_0]]
-// CHECK: [[V2:%[^ ]*]] = OpISub [[UINT]] [[V1]] [[RGB00]]
+// CHECK: [[V2:%[^ ]*]] = OpISub [[UINT]] [[V1]] [[RGB0]]
 // CHECK: OpStore [[BUF00_0]] [[V2]]
 
     lbuf[0] = (uint64_t) v;
 // CHECK: [[VECS:%[^ ]*]] = OpLoad [[VECTORS]]
-// CHECK: [[VECS00:%[^ ]*]] = OpCompositeExtract [[UINT]] [[VECS]] 0 0
+// CHECK: [[VECS0:%[^ ]*]] = OpCompositeExtract {{%v2uint}} [[VECS]] 0
+// CHECK: [[VECS00:%[^ ]*]] = OpCompositeExtract [[UINT]] [[VECS0]] 0
 // CHECK: [[V1_0:%[^ ]*]] = OpUConvert [[ULONG]] [[VECS00]]
 // CHECK: [[LBUF00:%[^ ]*]] = OpAccessChain %{{[^ ]*}} [[LBUF]] [[I0]] [[U0]]
 // CHECK: OpStore [[LBUF00]] [[V1_0]]
@@ -112,5 +112,22 @@ void main()
 // CHECK: [[V3:%[^ ]*]] = OpLoad [[ULONG]] [[LBUF00_0]]
 // CHECK: [[V4:%[^ ]*]] = OpIAdd [[ULONG]] [[V3]] [[V2_0]]
 // CHECK: OpStore [[LBUF00_0]] [[V4]]
+
+    vbuf[0] = (Vectors) colors;
+// CHECK: [[c0:%[^ ]*]] = OpLoad {{%[^ ]*}} %colors
+// CHECK: [[c0_0:%[^ ]+]] = OpCompositeExtract %ColorRGBA [[c0]] 0
+// The entire bit container extracted for each bitfield.
+// CHECK: [[c0_0_0:%[^ ]*]] = OpCompositeExtract %uint [[c0_0]] 0
+// CHECK: [[c0_0_1:%[^ ]*]] = OpCompositeExtract %uint [[c0_0]] 0
+// CHECK: [[c0_0_2:%[^ ]*]] = OpCompositeExtract %uint [[c0_0]] 0
+// CHECK: [[c0_0_3:%[^ ]*]] = OpCompositeExtract %uint [[c0_0]] 0
+// CHECK: [[v0:%[^ ]*]] = OpCompositeConstruct %v2uint [[c0_0_0]] [[c0_0_1]]
+// CHECK: [[v1:%[^ ]*]] = OpCompositeConstruct %v2uint [[c0_0_2]] [[c0_0_3]]
+// CHECK: [[v:%[^ ]*]] = OpCompositeConstruct %Vectors_0 [[v0]] [[v1]]
+// CHECK: [[vbuf:%[^ ]*]] = OpAccessChain %{{[^ ]*}} %vbuf [[I0]] [[U0]]
+// CHECK: [[v0:%[^ ]*]] = OpCompositeExtract %v2uint [[v]] 0
+// CHECK: [[v1:%[^ ]*]] = OpCompositeExtract %v2uint [[v]] 1
+// CHECK: [[v:%[^ ]*]] = OpCompositeConstruct %Vectors [[v0]] [[v1]]
+// CHECK: OpStore [[vbuf]] [[v]]
 }
 

tools/clang/test/CodeGenSPIRV/spirv.legal.cbuffer.hlsl

@@ -30,14 +30,24 @@ float4 main(in float4 pos : SV_Position) : SV_Target
 // Initializing a T with a ConstantBuffer<T> is a copy
 // CHECK:      [[val:%[0-9]+]] = OpLoad %type_ConstantBuffer_S %myCBuffer
 // CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeExtract %v4float [[val]] 0
+// CHECK-NEXT: [[e0:%[0-9]+]] = OpCompositeExtract %float [[vec]] 0
+// CHECK-NEXT: [[e1:%[0-9]+]] = OpCompositeExtract %float [[vec]] 1
+// CHECK-NEXT: [[e2:%[0-9]+]] = OpCompositeExtract %float [[vec]] 2
+// CHECK-NEXT: [[e3:%[0-9]+]] = OpCompositeExtract %float [[vec]] 3
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeConstruct %v4float  [[e0]] [[e1]] [[e2]] [[e3]]
 // CHECK-NEXT: [[tmp:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec]]
 // CHECK-NEXT:                OpStore %buffer1 [[tmp]]
     S buffer1 = myCBuffer;
 
 // Assigning a ConstantBuffer<T> to a T is a copy
 // CHECK:      [[val_0:%[0-9]+]] = OpLoad %type_ConstantBuffer_S %myCBuffer
-// CHECK-NEXT: [[vec_0:%[0-9]+]] = OpCompositeExtract %v4float [[val_0]] 0
-// CHECK-NEXT: [[tmp_0:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec_0]]
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeExtract %v4float [[val_0]] 0
+// CHECK-NEXT: [[e0:%[0-9]+]] = OpCompositeExtract %float [[vec]] 0
+// CHECK-NEXT: [[e1:%[0-9]+]] = OpCompositeExtract %float [[vec]] 1
+// CHECK-NEXT: [[e2:%[0-9]+]] = OpCompositeExtract %float [[vec]] 2
+// CHECK-NEXT: [[e3:%[0-9]+]] = OpCompositeExtract %float [[vec]] 3
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeConstruct %v4float  [[e0]] [[e1]] [[e2]] [[e3]]
+// CHECK-NEXT: [[tmp_0:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec]]
 // CHECK-NEXT:                OpStore %buffer2 [[tmp_0]]
     S buffer2;
     buffer2 = myCBuffer;
@@ -51,26 +61,41 @@ float4 main(in float4 pos : SV_Position) : SV_Target
 // Write out each component recursively
 // CHECK:      [[ptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %myASBuffer %uint_0 {{%[0-9]+}}
 // CHECK-NEXT: [[val_2:%[0-9]+]] = OpLoad %type_ConstantBuffer_S %myCBuffer
-// CHECK-NEXT: [[vec_1:%[0-9]+]] = OpCompositeExtract %v4float [[val_2]] 0
-// CHECK-NEXT: [[tmp_1:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec_1]]
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeExtract %v4float [[val_2]] 0
+// CHECK-NEXT: [[e0:%[0-9]+]] = OpCompositeExtract %float [[vec]] 0
+// CHECK-NEXT: [[e1:%[0-9]+]] = OpCompositeExtract %float [[vec]] 1
+// CHECK-NEXT: [[e2:%[0-9]+]] = OpCompositeExtract %float [[vec]] 2
+// CHECK-NEXT: [[e3:%[0-9]+]] = OpCompositeExtract %float [[vec]] 3
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeConstruct %v4float  [[e0]] [[e1]] [[e2]] [[e3]]
+// CHECK-NEXT: [[tmp_1:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec]]
 // CHECK-NEXT: [[vec_2:%[0-9]+]] = OpCompositeExtract %v4float [[tmp_1]] 0
 // CHECK-NEXT: [[tmp_2:%[0-9]+]] = OpCompositeConstruct %S [[vec_2]]
 // CHECK-NEXT:                OpStore [[ptr]] [[tmp_2]]
     myASBuffer.Append(myCBuffer);
 
 // Passing a ConstantBuffer<T> to a T parameter is a copy
 // CHECK:      [[val_3:%[0-9]+]] = OpLoad %type_ConstantBuffer_S %myCBuffer
-// CHECK-NEXT: [[vec_3:%[0-9]+]] = OpCompositeExtract %v4float [[val_3]] 0
-// CHECK-NEXT: [[tmp_3:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec_3]]
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeExtract %v4float [[val_3]] 0
+// CHECK-NEXT: [[e0:%[0-9]+]] = OpCompositeExtract %float [[vec]] 0
+// CHECK-NEXT: [[e1:%[0-9]+]] = OpCompositeExtract %float [[vec]] 1
+// CHECK-NEXT: [[e2:%[0-9]+]] = OpCompositeExtract %float [[vec]] 2
+// CHECK-NEXT: [[e3:%[0-9]+]] = OpCompositeExtract %float [[vec]] 3
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeConstruct %v4float  [[e0]] [[e1]] [[e2]] [[e3]]
+// CHECK-NEXT: [[tmp_3:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec]]
 // CHECK-NEXT:                OpStore %param_var_buffer [[tmp_3]]
     return doStuff(myCBuffer);
 }
 
 S retStuff() {
 // Returning a ConstantBuffer<T> as a T is a copy
 // CHECK:      [[val_4:%[0-9]+]] = OpLoad %type_ConstantBuffer_S %myCBuffer
-// CHECK-NEXT: [[vec_4:%[0-9]+]] = OpCompositeExtract %v4float [[val_4]] 0
-// CHECK-NEXT: [[ret:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec_4]]
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeExtract %v4float [[val_4]] 0
+// CHECK-NEXT: [[e0:%[0-9]+]] = OpCompositeExtract %float [[vec]] 0
+// CHECK-NEXT: [[e1:%[0-9]+]] = OpCompositeExtract %float [[vec]] 1
+// CHECK-NEXT: [[e2:%[0-9]+]] = OpCompositeExtract %float [[vec]] 2
+// CHECK-NEXT: [[e3:%[0-9]+]] = OpCompositeExtract %float [[vec]] 3
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeConstruct %v4float  [[e0]] [[e1]] [[e2]] [[e3]]
+// CHECK-NEXT: [[ret:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec]]
 // CHECK-NEXT:                OpReturnValue [[ret]]
     return myCBuffer;
 }

tools/clang/test/CodeGenSPIRV/spirv.legal.tbuffer.hlsl

@@ -34,14 +34,24 @@ float4 main(in float4 pos : SV_Position) : SV_Target
 // Initializing a T with a TextureBuffer<T> is a copy
 // CHECK:      [[val:%[0-9]+]] = OpLoad %type_TextureBuffer_S %myTBuffer
 // CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeExtract %v4float [[val]] 0
-// CHECK-NEXT: [[tmp:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec]]
+// CHECK-NEXT: [[e0:%[0-9]+]] = OpCompositeExtract %float [[vec]] 0
+// CHECK-NEXT: [[e1:%[0-9]+]] = OpCompositeExtract %float [[vec]] 1
+// CHECK-NEXT: [[e2:%[0-9]+]] = OpCompositeExtract %float [[vec]] 2
+// CHECK-NEXT: [[e3:%[0-9]+]] = OpCompositeExtract %float [[vec]] 3
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeConstruct %v4float  [[e0]] [[e1]] [[e2]] [[e3]]
+// CHECK: [[tmp:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec]]
 // CHECK-NEXT:                OpStore %buffer1 [[tmp]]
     S buffer1 = myTBuffer;
 
 // Assigning a TextureBuffer<T> to a T is a copy
 // CHECK:      [[val_0:%[0-9]+]] = OpLoad %type_TextureBuffer_S %myTBuffer
-// CHECK-NEXT: [[vec_0:%[0-9]+]] = OpCompositeExtract %v4float [[val_0]] 0
-// CHECK-NEXT: [[tmp_0:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec_0]]
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeExtract %v4float [[val_0]] 0
+// CHECK-NEXT: [[e0:%[0-9]+]] = OpCompositeExtract %float [[vec]] 0
+// CHECK-NEXT: [[e1:%[0-9]+]] = OpCompositeExtract %float [[vec]] 1
+// CHECK-NEXT: [[e2:%[0-9]+]] = OpCompositeExtract %float [[vec]] 2
+// CHECK-NEXT: [[e3:%[0-9]+]] = OpCompositeExtract %float [[vec]] 3
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeConstruct %v4float  [[e0]] [[e1]] [[e2]] [[e3]]
+// CHECK-NEXT: [[tmp_0:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec]]
 // CHECK-NEXT:                OpStore %buffer2 [[tmp_0]]
     S buffer2;
     buffer2 = myTBuffer;
@@ -57,6 +67,11 @@ float4 main(in float4 pos : SV_Position) : SV_Target
 // CHECK:      [[ptr:%[0-9]+]] = OpAccessChain %_ptr_Uniform_S %myASBuffer %uint_0 {{%[0-9]+}}
 // CHECK-NEXT:  [[tb:%[0-9]+]] = OpLoad %type_TextureBuffer_S %myTBuffer
 // CHECK-NEXT: [[vec_1:%[0-9]+]] = OpCompositeExtract %v4float [[tb]] 0
+// CHECK-NEXT: [[e0:%[0-9]+]] = OpCompositeExtract %float [[vec_1]] 0
+// CHECK-NEXT: [[e1:%[0-9]+]] = OpCompositeExtract %float [[vec_1]] 1
+// CHECK-NEXT: [[e2:%[0-9]+]] = OpCompositeExtract %float [[vec_1]] 2
+// CHECK-NEXT: [[e3:%[0-9]+]] = OpCompositeExtract %float [[vec_1]] 3
+// CHECK-NEXT: [[vec_1:%[0-9]+]] = OpCompositeConstruct %v4float  [[e0]] [[e1]] [[e2]] [[e3]]
 // CHECK-NEXT: [[loc:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec_1]]
 // CHECK-NEXT: [[vec_2:%[0-9]+]] = OpCompositeExtract %v4float [[loc]] 0
 // CHECK-NEXT: [[val_2:%[0-9]+]] = OpCompositeConstruct %S [[vec_2]]
@@ -65,17 +80,27 @@ float4 main(in float4 pos : SV_Position) : SV_Target
 
 // Passing a TextureBuffer<T> to a T parameter is a copy
 // CHECK:      [[val_3:%[0-9]+]] = OpLoad %type_TextureBuffer_S %myTBuffer
-// CHECK-NEXT: [[vec_3:%[0-9]+]] = OpCompositeExtract %v4float [[val_3]] 0
-// CHECK-NEXT: [[tmp_1:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec_3]]
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeExtract %v4float [[val_3]] 0
+// CHECK-NEXT: [[e0:%[0-9]+]] = OpCompositeExtract %float [[vec]] 0
+// CHECK-NEXT: [[e1:%[0-9]+]] = OpCompositeExtract %float [[vec]] 1
+// CHECK-NEXT: [[e2:%[0-9]+]] = OpCompositeExtract %float [[vec]] 2
+// CHECK-NEXT: [[e3:%[0-9]+]] = OpCompositeExtract %float [[vec]] 3
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeConstruct %v4float  [[e0]] [[e1]] [[e2]] [[e3]]
+// CHECK-NEXT: [[tmp_1:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec]]
 // CHECK-NEXT:                OpStore %param_var_buffer [[tmp_1]]
     return doStuff(myTBuffer);
 }
 
 S retStuff() {
 // Returning a TextureBuffer<T> as a T is a copy
 // CHECK:      [[val_4:%[0-9]+]] = OpLoad %type_TextureBuffer_S %myTBuffer
-// CHECK-NEXT: [[vec_4:%[0-9]+]] = OpCompositeExtract %v4float [[val_4]] 0
-// CHECK-NEXT: [[ret:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec_4]]
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeExtract %v4float [[val_4]] 0
+// CHECK-NEXT: [[e0:%[0-9]+]] = OpCompositeExtract %float [[vec]] 0
+// CHECK-NEXT: [[e1:%[0-9]+]] = OpCompositeExtract %float [[vec]] 1
+// CHECK-NEXT: [[e2:%[0-9]+]] = OpCompositeExtract %float [[vec]] 2
+// CHECK-NEXT: [[e3:%[0-9]+]] = OpCompositeExtract %float [[vec]] 3
+// CHECK-NEXT: [[vec:%[0-9]+]] = OpCompositeConstruct %v4float  [[e0]] [[e1]] [[e2]] [[e3]]
+// CHECK-NEXT: [[ret:%[0-9]+]] = OpCompositeConstruct %S_0 [[vec]]
 // CHECK-NEXT:                OpReturnValue [[ret]]
     return myTBuffer;
 }