[CIR] Improved cir::CastOp verifier to allow bitcasts between types of the same size (#1728)

The `cir::CastOp::verify` method was overly conservative, and would fail
on any `bitcast` from vector to scalar or scalar to vector.

Change List:
- Extends the `cir::CastOp::verify` method to check if the source and
result types are the same size using the `mlir::DataLayout` of the
current scope, and succeeds if the sizes match.
- Extends the CodeGen vectype tests with vector to scalar, scalar to
vector and vector to vector conversions.
- Extends the IR invalid tests with vector to scalar and scalar to
vector conversions with different source and result sizes.

Author: tommymcm

clang/lib/CIR/Dialect/IR/CIRDialect.cpp

@@ -35,6 +35,7 @@
 #include "mlir/IR/Location.h"
 #include "mlir/IR/OpDefinition.h"
 #include "mlir/IR/OpImplementation.h"
+#include "mlir/IR/Operation.h"
 #include "mlir/IR/StorageUniquerSupport.h"
 #include "mlir/IR/TypeUtilities.h"
 #include "mlir/Interfaces/DataLayoutInterfaces.h"
@@ -570,6 +571,17 @@ LogicalResult cir::CastOp::verify() {
         mlir::isa<cir::MethodType>(resType))
       return success();
 
+    // Handle scalar to vector and vector to scalar conversions.
+    if (mlir::isa<cir::VectorType>(getSrc().getType()) !=
+        mlir::isa<cir::VectorType>(getType())) {
+      // The source and result must be the same size.
+      mlir::DataLayout dataLayout(
+          getOperation()->getParentOfType<mlir::DataLayoutOpInterface>());
+      if (dataLayout.getTypeSize(getSrc().getType()) ==
+          dataLayout.getTypeSize(getType()))
+        return success();
+    }
+
     // This is the only cast kind where we don't want vector types to decay
     // into the element type.
     if ((!mlir::isa<cir::VectorType>(getSrc().getType()) ||

clang/test/CIR/CodeGen/vectype.cpp

@@ -120,19 +120,31 @@ void vector_int_test(int x, unsigned short usx) {
 
   // Shifts
   vi4 w = a << b;
-  // CHECK: %{{[0-9]+}} = cir.shift(left, {{%.*}} : !cir.vector<!s32i x 4>, 
+  // CHECK: %{{[0-9]+}} = cir.shift(left, {{%.*}} : !cir.vector<!s32i x 4>,
   // CHECK-SAME: {{%.*}} : !cir.vector<!s32i x 4>) -> !cir.vector<!s32i x 4>
   vi4 y = a >> b;
-  // CHECK: %{{[0-9]+}} = cir.shift(right, {{%.*}} : !cir.vector<!s32i x 4>, 
+  // CHECK: %{{[0-9]+}} = cir.shift(right, {{%.*}} : !cir.vector<!s32i x 4>,
   // CHECK-SAME: {{%.*}} : !cir.vector<!s32i x 4>) -> !cir.vector<!s32i x 4>
 
-  vus2 z = { usx, usx };  
+  vus2 z = { usx, usx };
   // CHECK: %{{[0-9]+}} = cir.vec.create(%{{[0-9]+}}, %{{[0-9]+}} : !u16i, !u16i) : !cir.vector<!u16i x 2>
   vus2 zamt = { 3, 4 };
   // CHECK: %{{[0-9]+}} = cir.const #cir.const_vector<[#cir.int<3> : !u16i, #cir.int<4> : !u16i]> : !cir.vector<!u16i x 2>
   vus2 zzz = z >> zamt;
-  // CHECK: %{{[0-9]+}} = cir.shift(right, {{%.*}} : !cir.vector<!u16i x 2>, 
-  // CHECK-SAME: {{%.*}} : !cir.vector<!u16i x 2>) -> !cir.vector<!u16i x 2> 
+  // CHECK: %{{[0-9]+}} = cir.shift(right, {{%.*}} : !cir.vector<!u16i x 2>,
+  // CHECK-SAME: {{%.*}} : !cir.vector<!u16i x 2>) -> !cir.vector<!u16i x 2>
+
+  // Vector to scalar conversion
+  unsigned int zi = (unsigned int)z;
+  // CHECK: %{{[0-9]+}} = cir.cast(bitcast, {{%.*}} : !cir.vector<!u16i x 2>), !u32i
+
+  // Scalar to vector conversion
+  vus2 zz = (vus2)zi;
+  // CHECK: %{{[0-9]+}} = cir.cast(bitcast, {{%.*}} : !u32i), !cir.vector<!u16i x 2>
+
+  // Vector to vector conversion
+  vll2 aaa = (vll2)a;
+  // CHECK: %{{[0-9]+}} = cir.cast(bitcast, {{%.*}} : !cir.vector<!s32i x 4>), !cir.vector<!s64i x 2>
 }
 
 void vector_double_test(int x, double y) {

clang/test/CIR/IR/invalid.cir

@@ -1410,7 +1410,35 @@ module {
     // expected-error@+1 {{'cir.cast' op result type address space does not match the address space of the operand}}
     %1 = cir.cast(bitcast, %0 : !cir.ptr<!s32i>), !cir.ptr<!s32i, addrspace(offload_local)>
   }
+}
+
+// -----
+
+!s16i = !cir.int<s, 16>
+!s64i = !cir.int<s, 64>
 
+module {
+  cir.func @test_bitcast_vec2scalar_diff_size() {
+    %0 = cir.const #cir.int<1> : !s16i
+    %1 = cir.vec.create(%0, %0 : !s16i, !s16i) : !cir.vector<!s16i x 2>
+    // expected-error@+1 {{'cir.cast' op requires !cir.ptr or !cir.vector type for source and result}}
+    %2 = cir.cast(bitcast, %1 : !cir.vector<!s16i x 2>), !s64i
+    cir.return
+  }
+}
+
+// -----
+
+!s32i = !cir.int<s, 32>
+!s64i = !cir.int<s, 64>
+
+module {
+  cir.func @test_bitcast_scalar2vec_diff_size() {
+    %0 = cir.const #cir.int<1> : !s64i
+    // expected-error@+1 {{'cir.cast' op requires !cir.ptr or !cir.vector type for source and result}}
+    %1 = cir.cast(bitcast, %0 : !s64i), !cir.vector<!s32i x 4>
+    cir.return
+  }
 }
 
 // -----