[clang] [llvm] [HLSL][Matrix] Support row-major `transpose` and `mul` by inserting matrix memory layout transformations (PR #186898)

[Finn Plummer via cfe-commits]

================
@@ -90,22 +93,31 @@ export float2x3 test_mat_scalar_mul(float2x3 a, float b) { return mul(a, b); }
 // -- Case 8: matrix * vector -> vector --
 
 // CHECK-LABEL: test_mat_vec_mul
-// CHECK: %hlsl.mul = {{.*}} call {{.*}} <2 x float> @llvm.matrix.multiply.v2f32.v6f32.v3f32(<6 x float> %m, <3 x float> %v, i32 2, i32 3, i32 1)
+// ROWMAJOR: {{.*}} = {{.*}} call {{.*}} <6 x float> @llvm.matrix.transpose.v6f32(<6 x float> %{{.*}}, i32 3, i32 2)
+// CHECK: %hlsl.mul = {{.*}} call {{.*}} <2 x float> @llvm.matrix.multiply.v2f32.v6f32.v3f32(<6 x float> %{{.*}}, <3 x float> %v, i32 2, i32 3, i32 1)
 // CHECK: ret <2 x float> %hlsl.mul
 export float2 test_mat_vec_mul(float2x3 m, float3 v) { return mul(m, v); }
 
 // -- Case 9: matrix * matrix -> matrix --
 
 // CHECK-LABEL: test_mat_mat_mul
-// CHECK: %hlsl.mul = {{.*}} call {{.*}} <8 x float> @llvm.matrix.multiply.v8f32.v6f32.v12f32(<6 x float> %a, <12 x float> %b, i32 2, i32 3, i32 4)
-// CHECK: ret <8 x float> %hlsl.mul
+// ROWMAJOR: {{.*}} = {{.*}} call {{.*}} <6 x float> @llvm.matrix.transpose.v6f32(<6 x float> %{{.*}}, i32 3, i32 2)
+// ROWMAJOR: {{.*}} = {{.*}} call {{.*}} <12 x float> @llvm.matrix.transpose.v12f32(<12 x float> %{{.*}}, i32 4, i32 3)
+// CHECK: %hlsl.mul = {{.*}} call {{.*}} <8 x float> @llvm.matrix.multiply.v8f32.v6f32.v12f32(<6 x float> %{{.*}}, <12 x float> %{{.*}}, i32 2, i32 3, i32 4)
+// COLMAJOR: ret <8 x float> %hlsl.mul
+// ROWMAJOR: %[[TRANSPOSE_RES:.*]] = {{.*}} call {{.*}} <8 x float> @llvm.matrix.transpose.v8f32(<8 x float> %hlsl.mul, i32 2, i32 4)
+// ROWMAJOR: ret <8 x float> %[[TRANSPOSE_RES]]
 export float2x4 test_mat_mat_mul(float2x3 a, float3x4 b) { return mul(a, b); }
 
 // -- Integer matrix multiply --
 
 // CHECK-LABEL: test_mat_mat_muli
-// CHECK: %hlsl.mul = {{.*}} call <8 x i32> @llvm.matrix.multiply.v8i32.v6i32.v12i32(<6 x i32> %a, <12 x i32> %b, i32 2, i32 3, i32 4)
-// CHECK: ret <8 x i32> %hlsl.mul
+// ROWMAJOR: {{.*}} = {{.*}} call <6 x i32> @llvm.matrix.transpose.v6i32(<6 x i32> %{{.*}}, i32 3, i32 2)
+// ROWMAJOR: {{.*}} = {{.*}} call <12 x i32> @llvm.matrix.transpose.v12i32(<12 x i32> %{{.*}}, i32 4, i32 3)
+// CHECK: %hlsl.mul = {{.*}} call <8 x i32> @llvm.matrix.multiply.v8i32.v6i32.v12i32(<6 x i32> %{{.*}}, <12 x i32> %{{.*}}, i32 2, i32 3, i32 4)
+// COLMAJOR: ret <8 x i32> %hlsl.mul
+// ROWMAJOR: %[[TRANSPOSE_RES:.*]] = {{.*}} call <8 x i32> @llvm.matrix.transpose.v8i32(<8 x i32> %hlsl.mul, i32 2, i32 4)
+// ROWMAJOR: ret <8 x i32> %[[TRANSPOSE_RES]]
----------------
inbelic wrote:

nit:

```suggestion
// CHECK: %[[MUL:.*]] = {{.*}} call <8 x i32> @llvm.matrix.multiply.v8i32.v6i32.v12i32(<6 x i32> %{{.*}}, <12 x i32> %{{.*}}, i32 2, i32 3, i32 4)
// ROWMAJOR: %[[TRANSPOSE_RES:.*]] = {{.*}} call <8 x i32> @llvm.matrix.transpose.v8i32(<8 x i32> hlsl.mul, i32 2, i32 4)
// ROWMAJOR: ret <8 x i32> %[[TRANSPOSE_RES]]
// COLMAJOR: ret <8 x i32> %[[MUL_RES]]
```
(and eleswhere, or at least capture `%hlsl.mul`?)

https://github.com/llvm/llvm-project/pull/186898