x86: preserve flags when folding atomic operations

D4796 taught LLVM to fold some atomic integer operations into a single
instruction. The pattern was unaware that the instructions clobbered
flags. I fixed some of this issue in D13680 but had missed INC/DEC.

This patch adds the missing EFLAGS definition.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@250438 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86InstrCompiler.td b/lib/Target/X86/X86InstrCompiler.td
index 4c793e9e4d8e7d4238006964d25a54da3a864f03..3e16eedfe70e2dc27753cae821bb02f8499d180a 100644 (file)
--- a/lib/Target/X86/X86InstrCompiler.td
+++ b/lib/Target/X86/X86InstrCompiler.td
@@ -821,24 +821,29 @@ multiclass RELEASE_UNOP<dag dag8, dag dag16, dag dag32, dag dag64> {
         [(atomic_store_64 addr:$dst, dag64)]>;
 }
 
-defm RELEASE_INC : RELEASE_UNOP<
-    (add (atomic_load_8  addr:$dst), (i8 1)),
-    (add (atomic_load_16 addr:$dst), (i16 1)),
-    (add (atomic_load_32 addr:$dst), (i32 1)),
-    (add (atomic_load_64 addr:$dst), (i64 1))>, Requires<[NotSlowIncDec]>;
-defm RELEASE_DEC : RELEASE_UNOP<
-    (add (atomic_load_8  addr:$dst), (i8 -1)),
-    (add (atomic_load_16 addr:$dst), (i16 -1)),
-    (add (atomic_load_32 addr:$dst), (i32 -1)),
-    (add (atomic_load_64 addr:$dst), (i64 -1))>, Requires<[NotSlowIncDec]>;
+let Defs = [EFLAGS] in {
+  defm RELEASE_INC : RELEASE_UNOP<
+      (add (atomic_load_8  addr:$dst), (i8 1)),
+      (add (atomic_load_16 addr:$dst), (i16 1)),
+      (add (atomic_load_32 addr:$dst), (i32 1)),
+      (add (atomic_load_64 addr:$dst), (i64 1))>, Requires<[NotSlowIncDec]>;
+  defm RELEASE_DEC : RELEASE_UNOP<
+      (add (atomic_load_8  addr:$dst), (i8 -1)),
+      (add (atomic_load_16 addr:$dst), (i16 -1)),
+      (add (atomic_load_32 addr:$dst), (i32 -1)),
+      (add (atomic_load_64 addr:$dst), (i64 -1))>, Requires<[NotSlowIncDec]>;
+}
 /*
 TODO: These don't work because the type inference of TableGen fails.
 TODO: find a way to fix it.
-defm RELEASE_NEG : RELEASE_UNOP<
-    (ineg (atomic_load_8  addr:$dst)),
-    (ineg (atomic_load_16 addr:$dst)),
-    (ineg (atomic_load_32 addr:$dst)),
-    (ineg (atomic_load_64 addr:$dst))>;
+let Defs = [EFLAGS] in {
+  defm RELEASE_NEG : RELEASE_UNOP<
+      (ineg (atomic_load_8  addr:$dst)),
+      (ineg (atomic_load_16 addr:$dst)),
+      (ineg (atomic_load_32 addr:$dst)),
+      (ineg (atomic_load_64 addr:$dst))>;
+}
+// NOT doesn't set flags.
 defm RELEASE_NOT : RELEASE_UNOP<
     (not (atomic_load_8  addr:$dst)),
     (not (atomic_load_16 addr:$dst)),

diff --git a/test/CodeGen/X86/atomic-flags.ll b/test/CodeGen/X86/atomic-flags.ll
index 141a7690dba2004cb7aac32e3767d2d82f395ca4..e0c4a915965c8bafbb5def08d5730d133dd8d97e 100644 (file)
--- a/test/CodeGen/X86/atomic-flags.ll
+++ b/test/CodeGen/X86/atomic-flags.ll
@@ -3,8 +3,8 @@
 
 ; 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:
+define i32 @atomic_and_flags_1(i8* %p, i32 %a, i32 %b) {
+; CHECK-LABEL: atomic_and_flags_1:
 
   ; Generate flags value, and use it.
   ; CHECK:      cmpl
@@ -14,9 +14,9 @@ define i32 @atomic_and_flags(i8* %p, i32 %a, i32 %b) {
 
 L1:
   ; The following pattern will get folded.
-  ; CHECK: addb
+  ; CHECK: incb
   %1 = load atomic i8, i8* %p seq_cst, align 1
-  %2 = add i8 %1, 2
+  %2 = add i8 %1, 1 ; This forces the INC instruction to be generated.
   store atomic i8 %2, i8* %p release, align 1
 
   ; Use the comparison result again. We need to rematerialize the comparison
@@ -36,3 +36,26 @@ L3:
 L4:
   ret i32 4
 }
+
+; Same as above, but using 2 as immediate to avoid the INC instruction.
+define i32 @atomic_and_flags_2(i8* %p, i32 %a, i32 %b) {
+; CHECK-LABEL: atomic_and_flags_2:
+  ; CHECK:      cmpl
+  ; CHECK-NEXT: jne
+  %cmp = icmp eq i32 %a, %b
+  br i1 %cmp, label %L1, label %L2
+L1:
+  ; 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
+  ; 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
+}