[Matrix] Adjust matrix pointer type for inline asm arguments.

[fhahn]

Summary

Matrix types in memory are represented as arrays, but accessed through
vector pointers, with the alignment specified on the access operation.

For inline assembly, update pointer arguments to use vector pointers.
Otherwise there will be a mis-match if the matrix is also an
input-argument which is represented as vector.

[fhahn]

There's still one remaining issue. The backend will still crash when the resulting vector does not fit into the target's vector register directly. But that's the same as with ext_vector_type I think. The snippet below should produce the same asm call as a float x 3 x 3 matrix.

using vec = float __attribute__((ext_vector_type(9)));
void f1 () {
    vec m;
    asm volatile("" : "+r,m"(m) : : "memory");
}

clang/lib/CodeGen/CGStmt.cpp

if (!Info.allowsMemory() && (hasScalarEvaluationKind(OutExpr->getType()) ||
       }
 
       // Update largest vector width for any vector types.
       if (auto *VT = dyn_cast<llvm::VectorType>(ResultRegTypes.back()))
         LargestVectorWidth =
             std::max((uint64_t)LargestVectorWidth,
                      VT->getPrimitiveSizeInBits().getKnownMinSize());
     } else {
-      ArgTypes.push_back(Dest.getAddress(*this).getType());
-      Args.push_back(Dest.getPointer(*this));
+      llvm::Type *DestAddrTy = Dest.getAddress(*this).getType();
+      llvm::Value *DestPtr = Dest.getPointer(*this);
+      // Matrix types in memory are represented by arrays, but accessed through
+      // vector pointers, with the alignment specified on the access operation.
+      // For inline assembly, update pointer arguments to use vector pointers.
+      // Otherwise there will be a mis-match if the matrix is also an
+      // input-argument which is represented as vector.
+      if (isa<MatrixType>(OutExpr->getType().getCanonicalType())) {
+        DestAddrTy = llvm::PointerType::get(
+            ConvertType(OutExpr->getType()),
+            cast<llvm::PointerType>(DestAddrTy)->getAddressSpace());
+        DestPtr = Builder.CreateBitCast(DestPtr, DestAddrTy);
+      }
+      ArgTypes.push_back(DestAddrTy);
+      Args.push_back(DestPtr);
       Constraints += "=*";
       Constraints += OutputConstraint;
       ReadOnly = ReadNone = false;
     }
 
     if (Info.isReadWrite()) {
       InOutConstraints += ',';
 

clang/test/CodeGen/matrix-type.c

void matrix_struct(Matrix *a, Matrix *b) {
   b->Data = a->Data;
 }
 
 typedef double dx4x4_t __attribute__((matrix_type(4, 4)));
 void matrix_inline_asm_memory_readwrite() {
   // CHECK-LABEL: define void @matrix_inline_asm_memory_readwrite()
   // CHECK-NEXT:  entry:
   // CHECK-NEXT:    [[ALLOCA:%.+]] = alloca [16 x double], align 8
-  // CHECK-NEXT:    [[PTR:%.+]] = bitcast [16 x double]* [[ALLOCA]] to <16 x double>*
-  // CHECK-NEXT:    [[VAL:%.+]] = load <16 x double>, <16 x double>* [[PTR]], align 8
-  // FIXME: Pointer element type does not match the vector type.
-  // CHECK-NEXT:    call void asm sideeffect "", "=*r|m,0,~{memory},~{dirflag},~{fpsr},~{flags}"([16 x double]* [[ALLOCA]], <16 x double> [[VAL]])
+  // CHECK-NEXT:    [[PTR1:%.+]] = bitcast [16 x double]* [[ALLOCA]] to <16 x double>*
+  // CHECK-NEXT:    [[PTR2:%.+]] = bitcast [16 x double]* [[ALLOCA]] to <16 x double>*
+  // CHECK-NEXT:    [[VAL:%.+]] = load <16 x double>, <16 x double>* [[PTR2]], align 8
+  // CHECK-NEXT:    call void asm sideeffect "", "=*r|m,0,~{memory},~{dirflag},~{fpsr},~{flags}"(<16 x double>* [[PTR1]], <16 x double> [[VAL]])
   // CHECK-NEXT:    ret void
 
   dx4x4_t m;
   asm volatile(""
                : "+r,m"(m)
                :
                : "memory");
 }