[SPIR-V] Fix cast elision with initializer lists (#6992)

When an a chain of casts were performed in an initializer list, only the source and destination types were considered. This means float -> uint was the same as float -> int -> uint. This however is not correct: intermediate casts can change the result. Removing the shortcut we took solves this issue. Fixes #6975 Signed-off-by: Nathan Gauër <[email protected]>
microsoft · Nov 1, 2024 · ac36a79 · ac36a79
1 parent 5704c47
commit ac36a79
Show file tree

Hide file tree

Showing 6 changed files with 96 additions and 43 deletions.
diff --git a/tools/clang/lib/SPIRV/InitListHandler.cpp b/tools/clang/lib/SPIRV/InitListHandler.cpp
@@ -77,11 +77,6 @@ void InitListHandler::flatten(const InitListExpr *expr) {
     const Expr *init = expr->getInit(i);
     if (const auto *subInitList = dyn_cast<InitListExpr>(init)) {
       flatten(subInitList);
-    } else if (const auto *subInitList = dyn_cast<InitListExpr>(
-                   // Ignore constructor casts which are no-ops
-                   // For cases like: <type>(<initializer-list>)
-                   init->IgnoreParenNoopCasts(theEmitter.getASTContext()))) {
-      flatten(subInitList);
     } else {
       auto *initializer = theEmitter.loadIfGLValue(init);
       if (!initializer) {

diff --git a/tools/clang/test/CodeGenSPIRV/init.list.cast.hlsl b/tools/clang/test/CodeGenSPIRV/init.list.cast.hlsl
@@ -0,0 +1,38 @@
+// RUN: %dxc -T cs_6_6 -E main -fcgl  %s -spirv | FileCheck %s
+
+RWStructuredBuffer<float> buffer;
+RWStructuredBuffer<float4> buffer4;
+
+[numthreads(1, 1, 1)]
+void main() {
+  {
+    // CHECK: [[f:%[0-9]+]] = OpLoad %float {{.*}}
+    // CHECK: [[s:%[0-9]+]] = OpConvertFToS %int [[f]]
+    // CHECK: [[u:%[0-9]+]] = OpBitcast %uint [[s]]
+    // CHECK:                 OpStore {{.*}} [[u]]
+    uint p = uint(int(buffer[0]));
+  }
+
+  {
+    // CHECK: [[f:%[0-9]+]] = OpLoad %v4float {{.*}}
+    // CHECK: [[s:%[0-9]+]] = OpConvertFToS %v4int [[f]]
+    // CHECK: [[u:%[0-9]+]] = OpBitcast %v4uint [[s]]
+    // CHECK:                 OpStore {{.*}} [[u]]
+    uint4 p4 = uint4(int4(buffer4[0]));
+  }
+
+  {
+    // CHECK: [[f:%[0-9]+]] = OpLoad %v4float {{.*}}
+    // CHECK: [[s:%[0-9]+]] = OpConvertFToS %v4int [[f]]
+    // CHECK: [[u:%[0-9]+]] = OpBitcast %v4uint [[s]]
+    // CHECK:                 OpStore {{.*}} [[u]]
+    uint4 p4 = int4(buffer4[0]);
+  }
+
+  {
+    // CHECK: [[f:%[0-9]+]] = OpLoad %v4float {{.*}}
+    // CHECK: [[u:%[0-9]+]] = OpConvertFToU %v4uint [[f]]
+    // CHECK:                 OpStore {{.*}} [[u]]
+    uint4 p4 = uint4(uint4(buffer4[0]));
+  }
+}
diff --git a/tools/clang/test/CodeGenSPIRV/shader.debug.line.composite.hlsl b/tools/clang/test/CodeGenSPIRV/shader.debug.line.composite.hlsl
@@ -45,71 +45,79 @@ void main() {
 
   int4 a = {
       float2(1, 0),
-// CHECK: DebugLine [[src]] %uint_50 %uint_50 %uint_7 %uint_19
-// CHECK: OpFunctionCall %int4_bool_float3_0 %test_struct
+// CHECK:      DebugLine [[src]] %uint_51 %uint_51 %uint_7 %uint_19
+// CHECK-NEXT: OpFunctionCall %int4_bool_float3_0 %test_struct
+// CHECK:      OpVectorShuffle %v2float
       test_struct().c.zx
-// CHECK:      OpCompositeExtract %float {{%[0-9]+}} 0
+// CHECK:      DebugLine [[src]] %uint_46 %uint_46 %uint_12 %uint_12
+// CHECK-NEXT: OpCompositeExtract %float {{%[0-9]+}} 0
 // CHECK-NEXT: OpCompositeExtract %float {{%[0-9]+}} 1
-// CHECK-NEXT: DebugLine [[src]] %uint_46 %uint_46 %uint_12 %uint_12
 // CHECK-NEXT: OpConvertFToS %int
 // CHECK-NEXT: OpConvertFToS %int
+// CHECK:      DebugLine [[src]] %uint_51 %uint_51 %uint_7 %uint_23
+// CHECK-NEXT: OpCompositeExtract %float {{%[0-9]+}} 0
+// CHECK-NEXT: OpCompositeExtract %float {{%[0-9]+}} 1
+// CHECK:      DebugLine [[src]] %uint_46 %uint_46 %uint_12 %uint_12
+// CHECK-NEXT: OpConvertFToS %int
+// CHECK-NEXT: OpConvertFToS %int
+// CHECK:      DebugLine [[src]] %uint_46 %uint_65 %uint_12 %uint_3
 // CHECK:      OpCompositeConstruct %v4int
   };
 
 // CHECK:                        OpFDiv %float {{%[0-9]+}} %float_2
-// CHECK-NEXT:                   DebugLine [[src]] %uint_64 %uint_64 %uint_16 %uint_57
+// CHECK-NEXT:                   DebugLine [[src]] %uint_72 %uint_72 %uint_16 %uint_57
 // CHECK-NEXT:  [[first:%[0-9]+]] = OpCompositeConstruct %v2float {{%[0-9]+}} {{%[0-9]+}}
 // CHECK-NEXT: [[second:%[0-9]+]] = OpCompositeConstruct %v2float {{%[0-9]+}} {{%[0-9]+}}
 // CHECK-NEXT:        {{%[0-9]+}} = OpCompositeConstruct %mat2v2float [[first]] [[second]]
   float2x2 b = float2x2(a.x, b._m00, 2 + a.y, b._m11 / 2);
 
-// CHECK:                   DebugLine [[src]] %uint_69 %uint_69 %uint_12 %uint_14
+// CHECK:                   DebugLine [[src]] %uint_77 %uint_77 %uint_12 %uint_14
 // CHECK-NEXT: [[y:%[0-9]+]] = OpAccessChain %_ptr_Uniform_int4_bool_float3 %CONSTANTS %int_0
 // CHECK-NEXT:   {{%[0-9]+}} = OpAccessChain %_ptr_Uniform_v4int [[y]] %int_0
   int4 c = y.a;
 
-// CHECK:                   DebugLine [[src]] %uint_76 %uint_76 %uint_3 %uint_3
+// CHECK:                   DebugLine [[src]] %uint_84 %uint_84 %uint_3 %uint_3
 // CHECK-NEXT: [[z:%[0-9]+]] = OpLoad %type_2d_image %z
 // CHECK-NEXT: [[z_0:%[0-9]+]] = OpImageRead %v4int [[z]] {{%[0-9]+}} None
 // CHECK-NEXT: [[z_1:%[0-9]+]] = OpVectorShuffle %v3int [[z_0]] [[z_0]] 0 1 2
 // CHECK:        {{%[0-9]+}} = OpCompositeInsert %v3int %int_16 [[z_1]] 0
   z[uint2(2, 3)].x = 16;
 
-// CHECK:      DebugLine [[src]] %uint_82 %uint_82 %uint_3 %uint_4
+// CHECK:      DebugLine [[src]] %uint_90 %uint_90 %uint_3 %uint_4
 // CHECK-NEXT: OpLoad %mat2v2float %b
-// CHECK:      DebugLine [[src]] %uint_82 %uint_82 %uint_3 %uint_4
+// CHECK:      DebugLine [[src]] %uint_90 %uint_90 %uint_3 %uint_4
 // CHECK-NEXT: OpFSub %v2float
   b--;
 
   int2x2 d;
-// CHECK:      DebugLine [[src]] %uint_91 %uint_91 %uint_8 %uint_8
+// CHECK:      DebugLine [[src]] %uint_99 %uint_99 %uint_8 %uint_8
 // CHECK-NEXT: OpLoad %mat2v2float %b
-// CHECK-NEXT: DebugLine [[src]] %uint_91 %uint_91 %uint_3 %uint_12
+// CHECK-NEXT: DebugLine [[src]] %uint_99 %uint_99 %uint_3 %uint_12
 // CHECK-NEXT: OpCompositeExtract %v2float
 // CHECK:      OpCompositeConstruct %_arr_v2int_uint_2
 // CHECK-NEXT: OpStore %d
   modf(b, d);
 
-// CHECK-TODO:      DebugLine [[src]] %uint_95 %uint_95 %uint_3 %uint_11
+// CHECK-TODO:      DebugLine [[src]] %uint_103 %uint_103 %uint_3 %uint_11
 // CHECK-NEXT-TODO: OpFunctionCall %void %S_inc %foo
 // TODO(greg-lunarg):  foo.inc();
 
-// CHECK-TODO:      DebugLine [[src]] %uint_99 %uint_99 %uint_3 %uint_14
+// CHECK-TODO:      DebugLine [[src]] %uint_107 %uint_107 %uint_3 %uint_14
 // CHECK-NEXT-TODO: OpFunctionCall %void %S_inc %temp_var_S
 // TODO(greg-lunarg):  getS().inc();
 
-// CHECK:      DebugLine [[src]] %uint_105 %uint_105 %uint_19 %uint_19
+// CHECK:      DebugLine [[src]] %uint_113 %uint_113 %uint_19 %uint_19
 // CHECK-NEXT: OpLoad %init %bar
-// CHECK:      DebugLine [[src]] %uint_105 %uint_105 %uint_12 %uint_12
+// CHECK:      DebugLine [[src]] %uint_113 %uint_113 %uint_12 %uint_12
 // CHECK-NEXT: OpConvertFToS %int
   int4 e = {1, 2, bar};
 
-// CHECK:      DebugLine [[src]] %uint_111 %uint_111 %uint_7 %uint_25
+// CHECK:      DebugLine [[src]] %uint_119 %uint_119 %uint_7 %uint_25
 // CHECK-NEXT: OpCompositeConstruct %v2float %float_1 %float_2
-// CHECK-NEXT: DebugLine [[src]] %uint_111 %uint_111 %uint_22 %uint_22
+// CHECK-NEXT: DebugLine [[src]] %uint_119 %uint_119 %uint_22 %uint_22
 // CHECK-NEXT: OpCompositeExtract %int
   b = float2x2(1, 2, bar);
-// CHECK:      DebugLine [[src]] %uint_111 %uint_111 %uint_3 %uint_25
+// CHECK:      DebugLine [[src]] %uint_119 %uint_119 %uint_3 %uint_25
 // CHECK-NEXT: OpStore %b
 
 // TODO(jaebaek): Update InitListHandler to properly emit debug info.

diff --git a/tools/clang/test/CodeGenSPIRV/spirv.debug.opline.composite.hlsl b/tools/clang/test/CodeGenSPIRV/spirv.debug.opline.composite.hlsl
@@ -53,63 +53,67 @@ void main() {
 // CHECK-NEXT: OpCompositeExtract %float {{%[0-9]+}} 1
 // CHECK-NEXT: OpConvertFToS %int
 // CHECK-NEXT: OpConvertFToS %int
+// CHECK-NEXT: OpCompositeExtract %float {{%[0-9]+}} 0
+// CHECK-NEXT: OpCompositeExtract %float {{%[0-9]+}} 1
+// CHECK-NEXT: OpConvertFToS %int
+// CHECK-NEXT: OpConvertFToS %int
 // CHECK-NEXT: OpCompositeConstruct %v4int
   };
 
 // CHECK:                        OpFDiv %float {{%[0-9]+}} %float_2
-// CHECK-NEXT:                   OpLine [[file]] 64 24
+// CHECK-NEXT:                   OpLine [[file]] 68 24
 // CHECK-NEXT:  [[first:%[0-9]+]] = OpCompositeConstruct %v2float {{%[0-9]+}} {{%[0-9]+}}
 // CHECK-NEXT: [[second:%[0-9]+]] = OpCompositeConstruct %v2float {{%[0-9]+}} {{%[0-9]+}}
 // CHECK-NEXT:        {{%[0-9]+}} = OpCompositeConstruct %mat2v2float [[first]] [[second]]
   float2x2 b = float2x2(a.x, b._m00, 2 + a.y, b._m11 / 2);
 
-// CHECK:                   OpLine [[file]] 69 12
+// CHECK:                   OpLine [[file]] 73 12
 // CHECK-NEXT: [[y:%[0-9]+]] = OpAccessChain %_ptr_Uniform_int4_bool_float3 %CONSTANTS %int_0
 // CHECK-NEXT:   {{%[0-9]+}} = OpAccessChain %_ptr_Uniform_v4int [[y]] %int_0
   int4 c = y.a;
 
-// CHECK:                   OpLine [[file]] 76 3
+// CHECK:                   OpLine [[file]] 80 3
 // CHECK-NEXT: [[z:%[0-9]+]] = OpLoad %type_2d_image %z
 // CHECK-NEXT: [[z_0:%[0-9]+]] = OpImageRead %v4int [[z]] {{%[0-9]+}} None
 // CHECK-NEXT: [[z_1:%[0-9]+]] = OpVectorShuffle %v3int [[z_0]] [[z_0]] 0 1 2
 // CHECK-NEXT:   {{%[0-9]+}} = OpCompositeInsert %v3int %int_16 [[z_1]] 0
   z[uint2(2, 3)].x = 16;
 
-// CHECK:      OpLine [[file]] 82 3
+// CHECK:      OpLine [[file]] 86 3
 // CHECK-NEXT: OpLoad %mat2v2float %b
-// CHECK:      OpLine [[file]] 82 4
+// CHECK:      OpLine [[file]] 86 4
 // CHECK-NEXT: OpFSub %v2float
   b--;
 
   int2x2 d;
-// CHECK:      OpLine [[file]] 91 8
+// CHECK:      OpLine [[file]] 95 8
 // CHECK-NEXT: OpLoad %mat2v2float %b
-// CHECK-NEXT: OpLine [[file]] 91 3
+// CHECK-NEXT: OpLine [[file]] 95 3
 // CHECK-NEXT: OpCompositeExtract %v2float
-// CHECK:      OpLine [[file]] 91 11
+// CHECK:      OpLine [[file]] 95 11
 // CHECK:      OpStore %d
   modf(b, d);
 
-// CHECK:      OpLine [[file]] 95 7
+// CHECK:      OpLine [[file]] 99 7
 // CHECK-NEXT: OpFunctionCall %void %S_inc %foo
   foo.inc();
 
-// CHECK:      OpLine [[file]] 99 10
+// CHECK:      OpLine [[file]] 103 10
 // CHECK-NEXT: OpFunctionCall %void %S_inc %temp_var_S
   getS().inc();
 
-// CHECK:      OpLine [[file]] 105 19
+// CHECK:      OpLine [[file]] 109 19
 // CHECK-NEXT: OpLoad %init %bar
-// CHECK:      OpLine [[file]] 105 12
+// CHECK:      OpLine [[file]] 109 12
 // CHECK-NEXT: OpConvertFToS %int
   int4 e = {1, 2, bar};
 
-// CHECK:      OpLine [[file]] 111 15
+// CHECK:      OpLine [[file]] 115 15
 // CHECK-NEXT: OpCompositeConstruct %v2float %float_1 %float_2
-// CHECK-NEXT: OpLine [[file]] 111 22
+// CHECK-NEXT: OpLine [[file]] 115 22
 // CHECK-NEXT: OpCompositeExtract %int
   b = float2x2(1, 2, bar);
-// CHECK:      OpLine [[file]] 111 3
+// CHECK:      OpLine [[file]] 115 3
 // CHECK-NEXT: OpStore %b
 
 // TODO(jaebaek): Update InitListHandler to properly emit debug info.

diff --git a/tools/clang/test/CodeGenSPIRV/type.template.function.empty-struct-argument.hlsl b/tools/clang/test/CodeGenSPIRV/type.template.function.empty-struct-argument.hlsl
@@ -1,5 +1,6 @@
 // RUN: %dxc -T ps_6_0 -E main -HV 2021 -fcgl %s -spirv | FileCheck %s
 
+// CHECK: [[const:%[0-9]+]] = OpConstantComposite %v4float %float_1 %float_2 %float_3 %float_4
 struct A {};
 
 template <typename T0, typename T1 = A>
@@ -9,7 +10,8 @@ struct B {
 };
 
 float4 main() : SV_Target {
-  // CHECK: {{%[0-9]+}} = OpCompositeConstruct %v4float %float_1 %float_2 %float_3 %float_4
+  // CHECK: [[A:%[0-9]+]] = OpCompositeConstruct %A
+  // CHECK:  {{%[0-9]+}} = OpCompositeConstruct %B [[const]] [[A]]
   B<float4> b = { float4(1, 2, 3, 4) };
   return b.m0;
 }
diff --git a/tools/clang/test/CodeGenSPIRV/var.init.matrix.mxn.hlsl b/tools/clang/test/CodeGenSPIRV/var.init.matrix.mxn.hlsl
@@ -3,8 +3,10 @@
 // TODO: optimize to generate constant composite for suitable initializers
 // TODO: decompose matrix in initializer
 
-// CHECK:      [[v3fc1:%[0-9]+]] = OpConstantComposite %v3float %float_1 %float_1 %float_1
-// CHECK-NEXT: [[v3fc0:%[0-9]+]] = OpConstantComposite %v3float %float_0 %float_0 %float_0
+// CHECK-DAG:  [[v3fc0:%[0-9]+]] = OpConstantComposite %v3float %float_0 %float_0 %float_0
+// CHECK-DAG: [[f2_1_2:%[0-9]+]] = OpConstantComposite %v2float %float_1 %float_2
+// CHECK-DAG: [[i2_1_2:%[0-9]+]] = OpConstantComposite %v2int %int_1 %int_2
+// CHECK-DAG:  [[v3fc1:%[0-9]+]] = OpConstantComposite %v3float %float_1 %float_1 %float_1
 
 void main() {
 // CHECK-LABEL: %bb_entry = OpLabel
@@ -58,7 +60,9 @@ void main() {
 // CHECK-NEXT: [[ce05:%[0-9]+]] = OpCompositeExtract %float [[vec2a]] 0
 // CHECK-NEXT: [[ce06:%[0-9]+]] = OpCompositeExtract %float [[vec2a]] 1
 // CHECK-NEXT: [[cc15:%[0-9]+]] = OpCompositeConstruct %v4float [[ce02]] [[ce03]] [[ce04]] [[ce05]]
-// CHECK-NEXT: [[cc16:%[0-9]+]] = OpCompositeConstruct %v4float [[ce06]] %float_1 %float_2 %float_3
+// CHECK-NEXT: [[f_1:%[0-9]+]] = OpCompositeExtract %float [[f2_1_2]] 0
+// CHECK-NEXT: [[f_2:%[0-9]+]] = OpCompositeExtract %float [[f2_1_2]] 1
+// CHECK-NEXT: [[cc16:%[0-9]+]] = OpCompositeConstruct %v4float [[ce06]] [[f_1]] [[f_2]] %float_3
 // CHECK-NEXT: [[cc17:%[0-9]+]] = OpCompositeConstruct %mat4v4float [[cc14]] [[cc15]] [[cc16]] [[vec4]]
 // CHECK-NEXT:  OpStore %mat5 [[cc17]]
     float4x4 mat5 = {scalar, vec1, vec2,  // [0]
@@ -193,7 +197,9 @@ void main() {
 // CHECK-NEXT: [[ce05_0:%[0-9]+]] = OpCompositeExtract %int [[vec2a_0]] 0
 // CHECK-NEXT: [[ce06_0:%[0-9]+]] = OpCompositeExtract %int [[vec2a_0]] 1
 // CHECK-NEXT: [[cc15_0:%[0-9]+]] = OpCompositeConstruct %v4int [[ce02_0]] [[ce03_0]] [[ce04_0]] [[ce05_0]]
-// CHECK-NEXT: [[cc16_0:%[0-9]+]] = OpCompositeConstruct %v4int [[ce06_0]] %int_1 %int_2 %int_3
+// CHECK-NEXT: [[i_1:%[0-9]+]] = OpCompositeExtract %int [[i2_1_2]] 0
+// CHECK-NEXT: [[i_2:%[0-9]+]] = OpCompositeExtract %int [[i2_1_2]] 1
+// CHECK-NEXT: [[cc16_0:%[0-9]+]] = OpCompositeConstruct %v4int [[ce06_0]] [[i_1]] [[i_2]] %int_3
 // CHECK-NEXT: [[cc17_0:%[0-9]+]] = OpCompositeConstruct %_arr_v4int_uint_4 [[cc14_0]] [[cc15_0]] [[cc16_0]] [[vec4_0]]
 // CHECK-NEXT:  OpStore %imat5 [[cc17_0]]
     int4x4 imat5 = {intScalar, intVec1, intVec2, // [0]