[PATCH] D13680: x86: preserve flags when folding atomic operations

[JF Bastien via llvm-commits]

This revision was automatically updated to reflect the committed changes.
Closed by commit rL250135: x86: preserve flags when folding atomic operations (authored by jfb).

Changed prior to commit:
  http://reviews.llvm.org/D13680?vs=37190&id=37200#toc

Repository:
  rL LLVM

http://reviews.llvm.org/D13680

Files:
  llvm/trunk/lib/Target/X86/X86InstrCompiler.td
  llvm/trunk/test/CodeGen/X86/atomic-flags.ll

Index: llvm/trunk/test/CodeGen/X86/atomic-flags.ll
===================================================================
--- llvm/trunk/test/CodeGen/X86/atomic-flags.ll
+++ llvm/trunk/test/CodeGen/X86/atomic-flags.ll
@@ -0,0 +1,38 @@
+; RUN: llc < %s -mtriple=x86_64-unknown-unknown -verify-machineinstrs | FileCheck %s
+; RUN: llc < %s -mtriple=i686-unknown-unknown -verify-machineinstrs | FileCheck %s
+
+; Make sure that flags are properly preserved despite atomic optimizations.
+
+define i32 @atomic_and_flags(i8* %p, i32 %a, i32 %b) {
+; CHECK-LABEL: atomic_and_flags:
+
+  ; Generate flags value, and use it.
+  ; CHECK:      cmpl
+  ; CHECK-NEXT: jne
+  %cmp = icmp eq i32 %a, %b
+  br i1 %cmp, label %L1, label %L2
+
+L1:
+  ; The following pattern will get folded.
+  ; CHECK: addb
+  %1 = load atomic i8, i8* %p seq_cst, align 1
+  %2 = add i8 %1, 2
+  store atomic i8 %2, i8* %p release, align 1
+
+  ; Use the comparison result again. We need to rematerialize the comparison
+  ; somehow. This test checks that cmpl gets emitted again, but any
+  ; rematerialization would work (the optimizer used to clobber the flags with
+  ; the add).
+  ; CHECK-NEXT: cmpl
+  ; CHECK-NEXT: jne
+  br i1 %cmp, label %L3, label %L4
+
+L2:
+  ret i32 2
+
+L3:
+  ret i32 3
+
+L4:
+  ret i32 4
+}
Index: llvm/trunk/lib/Target/X86/X86InstrCompiler.td
===================================================================
--- llvm/trunk/lib/Target/X86/X86InstrCompiler.td
+++ llvm/trunk/lib/Target/X86/X86InstrCompiler.td
@@ -774,12 +774,14 @@
         [(atomic_store_64 addr:$dst, (op
             (atomic_load_64 addr:$dst), GR64:$src))]>;
 }
-defm RELEASE_ADD : RELEASE_BINOP_MI<add>;
-defm RELEASE_AND : RELEASE_BINOP_MI<and>;
-defm RELEASE_OR  : RELEASE_BINOP_MI<or>;
-defm RELEASE_XOR : RELEASE_BINOP_MI<xor>;
-// Note: we don't deal with sub, because substractions of constants are
-// optimized into additions before this code can run
+let Defs = [EFLAGS] in {
+  defm RELEASE_ADD : RELEASE_BINOP_MI<add>;
+  defm RELEASE_AND : RELEASE_BINOP_MI<and>;
+  defm RELEASE_OR  : RELEASE_BINOP_MI<or>;
+  defm RELEASE_XOR : RELEASE_BINOP_MI<xor>;
+  // Note: we don't deal with sub, because substractions of constants are
+  //       optimized into additions before this code can run.
+}
 
 // Same as above, but for floating-point.
 // FIXME: imm version.


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D13680.37200.patch
Type: text/x-patch
Size: 2349 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20151013/8da52705/attachment-0001.bin>