[flang-commits] [flang] [Flang][Fir] Set default alignment to 64 on inlineAllocation for arrays (PR #208781)

[Slava Zakharin via flang-commits]

================
@@ -1373,6 +1437,68 @@ struct AllocMemOpConversion : public fir::FIROpConversion<fir::AllocMemOp> {
         mlir::IntegerType::get(rewriter.getContext(), mallocTyWidth);
     if (mallocTyWidth != ity.getIntOrFloatBitWidth())
       size = integerCast(loc, rewriter, mallocTy, size);
+
+    std::optional<uint64_t> alignment = heap.getAlignment();
+    if (alignment && *alignment > 16) {
+      auto mod = heap->getParentOfType<mlir::ModuleOp>();
+      llvm::Triple triple = mod ? fir::getTargetTriple(mod) : llvm::Triple{};
+      bool isGpu = heap->getParentOfType<mlir::gpu::GPUModuleOp>() != nullptr ||
+                   triple.isNVPTX() || triple.isAMDGPU() || triple.isSPIRV();
+
+      if (!isGpu && !triple.isOSWindows()) {
+        mlir::Value alignVal = fir::genConstantIndex(
+            loc, mallocTy, rewriter, static_cast<std::int64_t>(*alignment));
+
+        if (triple.isOSDarwin()) {
+          // aligned_alloc requires macOS 10.15+, so use posix_memalign instead
+          mlir::Type ptrTy = ::getLlvmPtrType(heap.getContext());
+          mlir::Value memptr;
+          {
+            mlir::OpBuilder::InsertionGuard guard(rewriter);
+            mlir::Operation *parentOp =
+                rewriter.getInsertionBlock()->getParentOp();
+            mlir::Region *parentRegion =
+                rewriter.getInsertionBlock()->getParent();
+            mlir::Block *insertBlock =
+                getBlockForAllocaInsert(parentOp, parentRegion);
+            rewriter.setInsertionPointToStart(insertBlock);
+            mlir::Value one = fir::genConstantIndex(loc, mallocTy, rewriter, 1);
+            memptr =
+                mlir::LLVM::AllocaOp::create(rewriter, loc, ptrTy, ptrTy, one);
+          }
+          mlir::Value nullPtr =
+              mlir::LLVM::ZeroOp::create(rewriter, loc, ptrTy);
+          mlir::LLVM::StoreOp::create(rewriter, loc, nullPtr, memptr);
+          heap->setAttr("callee", getPosixMemalign(heap, rewriter, mallocTy));
+          mlir::LLVM::CallOp::create(
+              rewriter, loc,
+              mlir::TypeRange{
+                  mlir::IntegerType::get(rewriter.getContext(), 32)},
+              mlir::ValueRange{memptr, alignVal, size},
+              addLLVMOpBundleAttrs(rewriter, heap->getAttrs(), 3));
+          mlir::Value newPtr =
+              mlir::LLVM::LoadOp::create(rewriter, loc, ptrTy, memptr);
+          rewriter.replaceOp(heap, newPtr);
+          return mlir::success();
+        }
+
+        mlir::Value alignMinusOne = fir::genConstantIndex(
+            loc, mallocTy, rewriter, static_cast<std::int64_t>(*alignment - 1));
+        mlir::Value sizePlus = mlir::LLVM::AddOp::create(
+            rewriter, loc, mallocTy, size, alignMinusOne);
+        mlir::Value div = mlir::LLVM::UDivOp::create(rewriter, loc, mallocTy,
+                                                     sizePlus, alignVal);
+        mlir::Value roundedSize =
+            mlir::LLVM::MulOp::create(rewriter, loc, mallocTy, div, alignVal);
----------------
vzakhari wrote:

Can you generate `(size + alignMinusOne) & ~alignMinusOne` instead?

https://github.com/llvm/llvm-project/pull/208781