[llvm] [SROA] Avoid unnecessary stack realignment when rewriting allocas (PR #222967)

[Nikita Popov via llvm-commits]

================
@@ -0,0 +1,102 @@
+; RUN: opt -data-layout="e-p:64:64-i64:64-n32:64-S128" -passes='sroa<aggregate-to-vector>' -S %s | FileCheck %s --check-prefixes=COMMON,STACK128
+; RUN: opt -data-layout="e-p:64:64-i64:64-n32:64-S64" -passes='sroa<aggregate-to-vector>' -S %s | FileCheck %s --check-prefixes=COMMON,STACK64
+; RUN: opt -data-layout="e-p:64:64-i64:64-n32:64" -passes='sroa<aggregate-to-vector>' -S %s | FileCheck %s --check-prefixes=COMMON,UNSPECIFIED
+
+; SROA preserves inherited alignment and only raises it to the preferred
+; alignment when the natural stack alignment permits.
+target datalayout = "e-p:64:64-i64:64-n32:64-S128"
----------------
nikic wrote:

You're already passing a data layout to opt, so omit this one to avoid confusion.

https://github.com/llvm/llvm-project/pull/222967