[CIR][ABI][AArch64][Lowering] support for calling struct types in range (64, 128) (#1141)

This PR adds support for the lowering of AArch64 calls with structs
having sizes greater than 64 and less than 128.

The idea is from the original
[CodeGen](https://github.com/llvm/clangir/blob/da601b374deea6665f710f7e432dfa82f457059e/clang/lib/CodeGen/CGCall.cpp#L1329),
where we perform a coercion through memory for these type of calls.

I have added a test for this.

Author: bruteforceboy

clang/lib/CIR/Dialect/Transforms/TargetLowering/LowerFunction.cpp

@@ -336,6 +336,13 @@ mlir::Value createCoercedValue(mlir::Value Src, mlir::Type Ty,
     return CGF.buildAggregateBitcast(Src, Ty);
   }
 
+  if (auto alloca = findAlloca(Src.getDefiningOp())) {
+    auto tmpAlloca = createTmpAlloca(CGF, alloca.getLoc(), Ty);
+    createMemCpy(CGF, tmpAlloca, alloca, SrcSize.getFixedValue());
+    return CGF.getRewriter().create<LoadOp>(alloca.getLoc(),
+                                            tmpAlloca.getResult());
+  }
+
   cir_cconv_unreachable("NYI");
 }
 

clang/test/CIR/CallConvLowering/AArch64/aarch64-cc-structs.c

@@ -206,6 +206,31 @@ GT_128 call_and_get_gt_128() {
 // LLVM:   call void @llvm.memcpy.p0.p0.i64(ptr %[[#V1]], ptr %[[#V2]], i64 12, i1 false)
 void passS(S s) {}
 
+// CHECK: @callS()
+// CHECK: %[[#V0:]] = cir.alloca !ty_S, !cir.ptr<!ty_S>, ["s"] {alignment = 4 : i64}
+// CHECK: %[[#V1:]] = cir.alloca !cir.array<!u64i x 2>, !cir.ptr<!cir.array<!u64i x 2>>, ["tmp"] {alignment = 8 : i64}
+// CHECK: %[[#V2:]] = cir.load %[[#V0]] : !cir.ptr<!ty_S>, !ty_S
+// CHECK: %[[#V3:]] = cir.cast(bitcast, %[[#V0]] : !cir.ptr<!ty_S>), !cir.ptr<!void>
+// CHECK: %[[#V4:]] = cir.cast(bitcast, %[[#V1]] : !cir.ptr<!cir.array<!u64i x 2>>), !cir.ptr<!void>
+// CHECK: %[[#V5:]] = cir.const #cir.int<12> : !u64i
+// CHECK: cir.libc.memcpy %[[#V5]] bytes from %[[#V3]] to %[[#V4]] : !u64i, !cir.ptr<!void> -> !cir.ptr<!void>
+// CHECK: %[[#V6:]] = cir.load %[[#V1]] : !cir.ptr<!cir.array<!u64i x 2>>, !cir.array<!u64i x 2>
+// CHECK: cir.call @passS(%[[#V6]]) : (!cir.array<!u64i x 2>) -> ()
+// CHECK: cir.return
+
+// LLVM: @callS()
+// LLVM: %[[#V1:]] = alloca %struct.S, i64 1, align 4
+// LLVM: %[[#V2:]] = alloca [2 x i64], i64 1, align 8
+// LLVM: %[[#V3:]] = load %struct.S, ptr %[[#V1]], align 4
+// LLVM: call void @llvm.memcpy.p0.p0.i64(ptr %[[#V2]], ptr %[[#V1]], i64 12, i1 false)
+// LLVM: %[[#V4:]] = load [2 x i64], ptr %[[#V2]], align 8
+// LLVM: call void @passS([2 x i64] %[[#V4]])
+// LLVM: ret void
+void callS() {
+  S s;
+  passS(s);
+}
+
 typedef struct {
   uint8_t a;
   uint16_t b;