ARM: set preferred aggregate alignment to 32 universally.
authorTim Northover <tnorthover@apple.com>
Tue, 14 Oct 2014 20:57:26 +0000 (20:57 +0000)
committerTim Northover <tnorthover@apple.com>
Tue, 14 Oct 2014 20:57:26 +0000 (20:57 +0000)
Before, ARM and Thumb mode code had different preferred alignments, which could
lead to some rather unexpected results. There's justification for reducing it
from the default 64-bits (wasted space), but I don't think there is for going
below 32-bits.

There's no actual ABI change here, just to reassure people.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@219719 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/ARM/ARMSubtarget.cpp
test/CodeGen/ARM/aggregate-align.ll [new file with mode: 0644]

index a3ade98..bc04f37 100644 (file)
@@ -123,10 +123,9 @@ static std::string computeDataLayout(ARMSubtarget &ST) {
   else
     Ret += "-v128:64:128";
 
-  // On thumb and APCS, only try to align aggregates to 32 bits (the default is
-  // 64 bits).
-  if (ST.isThumb() || ST.isAPCS_ABI())
-    Ret += "-a:0:32";
+  // Try to align aggregates to 32 bits (the default is 64 bits, which has no
+  // particular hardware support on 32-bit ARM).
+  Ret += "-a:0:32";
 
   // Integer registers are 32 bits.
   Ret += "-n32";
diff --git a/test/CodeGen/ARM/aggregate-align.ll b/test/CodeGen/ARM/aggregate-align.ll
new file mode 100644 (file)
index 0000000..22fc57c
--- /dev/null
@@ -0,0 +1,6 @@
+; RUN: llc -mtriple=armv7-linux-gnueabi %s -o - | FileCheck %s
+
+@var = global {i8, i8} zeroinitializer
+
+; CHECK: .globl var
+; CHECK-NEXT: .align 2