[Mlir-commits] [clang] [llvm] [mlir] [AMDGPU] fix alignment of i128 (PR #213523)

[Matt Arsenault]

================
@@ -0,0 +1,82 @@
+; RUN: llc -mtriple=amdgcn-amd-amdhsa -mcpu=gfx1100 < %s | FileCheck %s
+
+; The AMDGPU data layout has to give i128 an ABI alignment of 16, matching the
+; ABI implemented by Clang, whose AMDGPUTargetInfo leaves Int128Align at its
+; 128-bit default. Without an explicit entry the alignment would be inherited
+; from the i64:64 entry, and the layout LLVM computes for an aggregate would
----------------
arsenm wrote:

I don't understand this. Surely the default datalayout should be 16-byte aligned, and the default should match the default in clang? This bug must exist in most backends?

https://github.com/llvm/llvm-project/pull/213523