[X86][APX] Use TEST instruction for CLOAD/CSTORE

[phoebewang]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-x86

Author: Phoebe Wang (phoebewang)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/151160.diff

2 Files Affected:

• (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+2-4)
• (modified) llvm/test/CodeGen/X86/apx/cf.ll (+2-2)

diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index 11ab8dc685ea8..7244a6d4d805a 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -58074,11 +58074,9 @@ static SDValue combineX86CloadCstore(SDNode *N, SelectionDAG &DAG) {
     // res, flags2 = sub 0, (and X, Y)
     // cload/cstore ..., cond_ne, flag2
     // ->
-    // res, flags2 = and X, Y
+    // res, flags2 = cmp (and X, Y), 0
     // cload/cstore ..., cond_ne, flag2
-    Ops[4] = DAG.getNode(X86ISD::AND, DL, Sub->getVTList(), Op1.getOperand(0),
-                         Op1.getOperand(1))
-                 .getValue(1);
+    Ops[4] = DAG.getNode(X86ISD::CMP, DL, MVT::i32, Op1, Sub.getOperand(0));
   } else {
     return SDValue();
   }
diff --git a/llvm/test/CodeGen/X86/apx/cf.ll b/llvm/test/CodeGen/X86/apx/cf.ll
index 1e4ac3f419314..b111ae542d93a 100644
--- a/llvm/test/CodeGen/X86/apx/cf.ll
+++ b/llvm/test/CodeGen/X86/apx/cf.ll
@@ -162,7 +162,7 @@ entry:
 define void @load_zext(i1 %cond, ptr %b, ptr %p) {
 ; CHECK-LABEL: load_zext:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    andb $1, %dil
+; CHECK-NEXT:    testb $1, %dil
 ; CHECK-NEXT:    cfcmovnew (%rsi), %ax
 ; CHECK-NEXT:    movzwl %ax, %eax
 ; CHECK-NEXT:    cfcmovnel %eax, (%rdx)
@@ -180,7 +180,7 @@ entry:
 define void @load_sext(i1 %cond, ptr %b, ptr %p) {
 ; CHECK-LABEL: load_sext:
 ; CHECK:       # %bb.0: # %entry
-; CHECK-NEXT:    andb $1, %dil
+; CHECK-NEXT:    testb $1, %dil
 ; CHECK-NEXT:    cfcmovnel (%rsi), %eax
 ; CHECK-NEXT:    cltq
 ; CHECK-NEXT:    cfcmovneq %rax, (%rdx)

[KanRobert]

Why it's not optimized to test by eflags-related peephole optimization?

I remember we used and instead of test to give more chance for CSE.

[phoebewang]

I only see cmp <-> test conversions. I think test is always preferred when only eflags of an and is used, hence pattern:

def X86testpat : PatFrag<(ops node:$lhs, node:$rhs),
                         (X86cmp (and_su node:$lhs, node:$rhs), 0)>;

[KanRobert]

@phoebewang Then it seems clang generates the worse code than gcc about this.

https://www.godbolt.org/z/T7djcxv5e

[phoebewang]

@phoebewang Then it seems clang generates the worse code than gcc about this.

https://www.godbolt.org/z/T7djcxv5e

Why worse? It reduces a branch.

[KanRobert]

My fault. It's an incorrect example.