tint: Preserve padding in matrices with three rows

The PreservePadding transform now decomposes writes to matrices with
three rows into separate column vector writes, to avoid modifying
padding between columns.

Bug: tint:1571
Change-Id: If575f79bb87f52810783fd3338e2f3ce3228ab2e
Reviewed-on: https://dawn-review.googlesource.com/c/dawn/+/121600
Auto-Submit: James Price <jrprice@google.com>
Kokoro: Kokoro <noreply+kokoro@google.com>
Reviewed-by: Ben Clayton <bclayton@google.com>
Commit-Queue: James Price <jrprice@google.com>

test/tint/buffer/uniform/std140/array/mat2x3_f32/to_storage.wgsl.expected.glsl

@@ -8,9 +8,24 @@ layout(binding = 1, std430) buffer u_block_ssbo {
   mat2x3 inner[4];
 } s;
 
+void assign_and_preserve_padding_1_s_X(uint dest[1], mat2x3 value) {
+  s.inner[dest[0]][0] = value[0u];
+  s.inner[dest[0]][1] = value[1u];
+}
+
+void assign_and_preserve_padding_s(mat2x3 value[4]) {
+  {
+    for(uint i = 0u; (i < 4u); i = (i + 1u)) {
+      uint tint_symbol[1] = uint[1](i);
+      assign_and_preserve_padding_1_s_X(tint_symbol, value[i]);
+    }
+  }
+}
+
 void f() {
-  s.inner = u.inner;
-  s.inner[1] = u.inner[2];
+  assign_and_preserve_padding_s(u.inner);
+  uint tint_symbol_1[1] = uint[1](1u);
+  assign_and_preserve_padding_1_s_X(tint_symbol_1, u.inner[2]);
   s.inner[1][0] = u.inner[0][1].zxy;
   s.inner[1][0].x = u.inner[0][1].x;
 }

test/tint/buffer/uniform/std140/array/mat2x3_f32/to_storage.wgsl.expected.msl

@@ -14,9 +14,20 @@ struct tint_array {
     T elements[N];
 };
 
+void assign_and_preserve_padding_1(device float2x3* const dest, float2x3 value) {
+  (*(dest))[0u] = value[0u];
+  (*(dest))[1u] = value[1u];
+}
+
+void assign_and_preserve_padding(device tint_array<float2x3, 4>* const dest, tint_array<float2x3, 4> value) {
+  for(uint i = 0u; (i < 4u); i = (i + 1u)) {
+    assign_and_preserve_padding_1(&((*(dest))[i]), value[i]);
+  }
+}
+
 kernel void f(device tint_array<float2x3, 4>* tint_symbol [[buffer(1)]], const constant tint_array<float2x3, 4>* tint_symbol_1 [[buffer(0)]]) {
-  *(tint_symbol) = *(tint_symbol_1);
-  (*(tint_symbol))[1] = (*(tint_symbol_1))[2];
+  assign_and_preserve_padding(tint_symbol, *(tint_symbol_1));
+  assign_and_preserve_padding_1(&((*(tint_symbol))[1]), (*(tint_symbol_1))[2]);
   (*(tint_symbol))[1][0] = float3((*(tint_symbol_1))[0][1]).zxy;
   (*(tint_symbol))[1][0][0] = (*(tint_symbol_1))[0][1][0];
   return;

test/tint/buffer/uniform/std140/array/mat2x3_f32/to_storage.wgsl.expected.spvasm

@@ -1,7 +1,7 @@
 ; SPIR-V
 ; Version: 1.3
 ; Generator: Google Tint Compiler; 0
-; Bound: 42
+; Bound: 83
 ; Schema: 0
                OpCapability Shader
                OpMemoryModel Logical GLSL450
@@ -11,6 +11,13 @@
                OpMemberName %u_block 0 "inner"
                OpName %u "u"
                OpName %s "s"
+               OpName %assign_and_preserve_padding_1_s_X "assign_and_preserve_padding_1_s_X"
+               OpName %dest "dest"
+               OpName %value "value"
+               OpName %assign_and_preserve_padding_s "assign_and_preserve_padding_s"
+               OpName %value_0 "value"
+               OpName %i "i"
+               OpName %var_for_index "var_for_index"
                OpName %f "f"
                OpDecorate %u_block Block
                OpMemberDecorate %u_block 0 Offset 0
@@ -22,6 +29,7 @@
                OpDecorate %u Binding 0
                OpDecorate %s DescriptorSet 0
                OpDecorate %s Binding 1
+               OpDecorate %_arr_uint_uint_1 ArrayStride 4
       %float = OpTypeFloat 32
     %v3float = OpTypeVector %float 3
 %mat2v3float = OpTypeMatrix %v3float 2
@@ -34,38 +42,94 @@
 %_ptr_StorageBuffer_u_block = OpTypePointer StorageBuffer %u_block
           %s = OpVariable %_ptr_StorageBuffer_u_block StorageBuffer
        %void = OpTypeVoid
-         %12 = OpTypeFunction %void
+     %uint_1 = OpConstant %uint 1
+%_arr_uint_uint_1 = OpTypeArray %uint %uint_1
+         %12 = OpTypeFunction %void %_arr_uint_uint_1 %mat2v3float
      %uint_0 = OpConstant %uint 0
-%_ptr_StorageBuffer__arr_mat2v3float_uint_4 = OpTypePointer StorageBuffer %_arr_mat2v3float_uint_4
-%_ptr_Uniform__arr_mat2v3float_uint_4 = OpTypePointer Uniform %_arr_mat2v3float_uint_4
         %int = OpTypeInt 32 1
+         %22 = OpConstantNull %int
+%_ptr_StorageBuffer_v3float = OpTypePointer StorageBuffer %v3float
+         %26 = OpConstantNull %uint
       %int_1 = OpConstant %int 1
-%_ptr_StorageBuffer_mat2v3float = OpTypePointer StorageBuffer %mat2v3float
+         %32 = OpTypeFunction %void %_arr_mat2v3float_uint_4
+%_ptr_Function_uint = OpTypePointer Function %uint
+       %bool = OpTypeBool
+%_ptr_Function__arr_mat2v3float_uint_4 = OpTypePointer Function %_arr_mat2v3float_uint_4
+         %50 = OpConstantNull %_arr_mat2v3float_uint_4
+%_ptr_Function_mat2v3float = OpTypePointer Function %mat2v3float
+         %60 = OpTypeFunction %void
+%_ptr_Uniform__arr_mat2v3float_uint_4 = OpTypePointer Uniform %_arr_mat2v3float_uint_4
+         %68 = OpConstantComposite %_arr_uint_uint_1 %uint_1
       %int_2 = OpConstant %int 2
 %_ptr_Uniform_mat2v3float = OpTypePointer Uniform %mat2v3float
-         %30 = OpConstantNull %int
-%_ptr_StorageBuffer_v3float = OpTypePointer StorageBuffer %v3float
 %_ptr_Uniform_v3float = OpTypePointer Uniform %v3float
 %_ptr_StorageBuffer_float = OpTypePointer StorageBuffer %float
 %_ptr_Uniform_float = OpTypePointer Uniform %float
-          %f = OpFunction %void None %12
-         %15 = OpLabel
-         %18 = OpAccessChain %_ptr_StorageBuffer__arr_mat2v3float_uint_4 %s %uint_0
-         %20 = OpAccessChain %_ptr_Uniform__arr_mat2v3float_uint_4 %u %uint_0
-         %21 = OpLoad %_arr_mat2v3float_uint_4 %20
-               OpStore %18 %21
-         %25 = OpAccessChain %_ptr_StorageBuffer_mat2v3float %s %uint_0 %int_1
-         %28 = OpAccessChain %_ptr_Uniform_mat2v3float %u %uint_0 %int_2
-         %29 = OpLoad %mat2v3float %28
-               OpStore %25 %29
-         %32 = OpAccessChain %_ptr_StorageBuffer_v3float %s %uint_0 %int_1 %30
-         %34 = OpAccessChain %_ptr_Uniform_v3float %u %uint_0 %30 %int_1
-         %35 = OpLoad %v3float %34
-         %36 = OpVectorShuffle %v3float %35 %35 2 0 1
-               OpStore %32 %36
-         %38 = OpAccessChain %_ptr_StorageBuffer_float %s %uint_0 %int_1 %30 %uint_0
-         %40 = OpAccessChain %_ptr_Uniform_float %u %uint_0 %30 %int_1 %uint_0
-         %41 = OpLoad %float %40
-               OpStore %38 %41
+%assign_and_preserve_padding_1_s_X = OpFunction %void None %12
+       %dest = OpFunctionParameter %_arr_uint_uint_1
+      %value = OpFunctionParameter %mat2v3float
+         %19 = OpLabel
+         %23 = OpCompositeExtract %uint %dest 0
+         %25 = OpAccessChain %_ptr_StorageBuffer_v3float %s %uint_0 %23 %22
+         %27 = OpCompositeExtract %v3float %value 0
+               OpStore %25 %27
+         %28 = OpCompositeExtract %uint %dest 0
+         %30 = OpAccessChain %_ptr_StorageBuffer_v3float %s %uint_0 %28 %int_1
+         %31 = OpCompositeExtract %v3float %value 1
+               OpStore %30 %31
+               OpReturn
+               OpFunctionEnd
+%assign_and_preserve_padding_s = OpFunction %void None %32
+    %value_0 = OpFunctionParameter %_arr_mat2v3float_uint_4
+         %35 = OpLabel
+          %i = OpVariable %_ptr_Function_uint Function %26
+%var_for_index = OpVariable %_ptr_Function__arr_mat2v3float_uint_4 Function %50
+               OpStore %i %26
+               OpBranch %38
+         %38 = OpLabel
+               OpLoopMerge %39 %40 None
+               OpBranch %41
+         %41 = OpLabel
+         %43 = OpLoad %uint %i
+         %44 = OpULessThan %bool %43 %uint_4
+         %42 = OpLogicalNot %bool %44
+               OpSelectionMerge %46 None
+               OpBranchConditional %42 %47 %46
+         %47 = OpLabel
+               OpBranch %39
+         %46 = OpLabel
+               OpStore %var_for_index %value_0
+         %52 = OpLoad %uint %i
+         %53 = OpCompositeConstruct %_arr_uint_uint_1 %52
+         %54 = OpLoad %uint %i
+         %56 = OpAccessChain %_ptr_Function_mat2v3float %var_for_index %54
+         %57 = OpLoad %mat2v3float %56
+         %51 = OpFunctionCall %void %assign_and_preserve_padding_1_s_X %53 %57
+               OpBranch %40
+         %40 = OpLabel
+         %58 = OpLoad %uint %i
+         %59 = OpIAdd %uint %58 %uint_1
+               OpStore %i %59
+               OpBranch %38
+         %39 = OpLabel
+               OpReturn
+               OpFunctionEnd
+          %f = OpFunction %void None %60
+         %62 = OpLabel
+         %65 = OpAccessChain %_ptr_Uniform__arr_mat2v3float_uint_4 %u %uint_0
+         %66 = OpLoad %_arr_mat2v3float_uint_4 %65
+         %63 = OpFunctionCall %void %assign_and_preserve_padding_s %66
+         %71 = OpAccessChain %_ptr_Uniform_mat2v3float %u %uint_0 %int_2
+         %72 = OpLoad %mat2v3float %71
+         %67 = OpFunctionCall %void %assign_and_preserve_padding_1_s_X %68 %72
+         %73 = OpAccessChain %_ptr_StorageBuffer_v3float %s %uint_0 %int_1 %22
+         %75 = OpAccessChain %_ptr_Uniform_v3float %u %uint_0 %22 %int_1
+         %76 = OpLoad %v3float %75
+         %77 = OpVectorShuffle %v3float %76 %76 2 0 1
+               OpStore %73 %77
+         %79 = OpAccessChain %_ptr_StorageBuffer_float %s %uint_0 %int_1 %22 %uint_0
+         %81 = OpAccessChain %_ptr_Uniform_float %u %uint_0 %22 %int_1 %uint_0
+         %82 = OpLoad %float %81
+               OpStore %79 %82
                OpReturn
                OpFunctionEnd