[clang] [CIR] Add support for __builtin_assume_aligned (PR #152152)

[Andy Kaylor via cfe-commits]

================
@@ -460,6 +460,28 @@ mlir::LogicalResult CIRToLLVMAssumeOpLowering::matchAndRewrite(
   return mlir::success();
 }
 
+mlir::LogicalResult CIRToLLVMAssumeAlignedOpLowering::matchAndRewrite(
+    cir::AssumeAlignedOp op, OpAdaptor adaptor,
+    mlir::ConversionPatternRewriter &rewriter) const {
+  SmallVector<mlir::Value, 3> opBundleArgs{adaptor.getPointer()};
+
+  auto alignment = mlir::LLVM::ConstantOp::create(
+      rewriter, op.getLoc(), rewriter.getI64Type(), adaptor.getAlignment());
+  opBundleArgs.push_back(alignment);
+
+  if (mlir::Value offset = adaptor.getOffset())
+    opBundleArgs.push_back(offset);
+
+  auto cond = mlir::LLVM::ConstantOp::create(rewriter, op.getLoc(),
+                                             rewriter.getI1Type(), 1);
+  mlir::LLVM::AssumeOp::create(rewriter, op.getLoc(), cond, "align",
+                               opBundleArgs);
+  rewriter.replaceAllUsesWith(op, op.getPointer());
----------------
andykaylor wrote:

```suggestion
  // The LLVM assume doesn't return a value, so we need to update any users of the op result.
  rewriter.replaceAllUsesWith(op, adaptor.getPointer());
```
It took me a bit to recognize why this was here, so a comment would be useful. Also, I think using `adaptor.getPointer()` rather than `op.getPointer()` avoids re-lowering the replaced argument since `adaptor.getPointer()` already has it in lowered form.

https://github.com/llvm/llvm-project/pull/152152