[X86][APX] Convert store(cmov(load(x), y), x) to cstore(y, x) (#118946)

Author: phoebewang

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -52820,6 +52820,44 @@ static SDValue combineStore(SDNode *N, SelectionDAG &DAG,
     }
   }
 
+  // Convert store(cmov(load(p), x, CC), p) to cstore(x, p, CC)
+  //         store(cmov(x, load(p), CC), p) to cstore(x, p, InvertCC)
+  if ((VT == MVT::i16 || VT == MVT::i32 || VT == MVT::i64) &&
+      Subtarget.hasCF() && St->isSimple()) {
+    SDValue Cmov;
+    if (StoredVal.getOpcode() == X86ISD::CMOV)
+      Cmov = StoredVal;
+    else if (StoredVal.getOpcode() == ISD::TRUNCATE &&
+             StoredVal.getOperand(0).getOpcode() == X86ISD::CMOV)
+      Cmov = StoredVal.getOperand(0);
+    else
+      return SDValue();
+
+    auto *Ld = dyn_cast<LoadSDNode>(St->getChain());
+    if (!Ld || !Ld->isSimple() || Ld->getBasePtr() != St->getBasePtr())
+      return SDValue();
+
+    bool InvertCC = false;
+    SDValue V = SDValue(Ld, 0);
+    if (V == Cmov.getOperand(1))
+      InvertCC = true;
+    else if (V != Cmov.getOperand(0))
+      return SDValue();
+
+    SDVTList Tys = DAG.getVTList(MVT::Other);
+    SDValue CC = Cmov.getOperand(2);
+    SDValue Src = DAG.getAnyExtOrTrunc(Cmov.getOperand(!InvertCC), dl, VT);
+    if (InvertCC)
+      CC = DAG.getTargetConstant(
+          GetOppositeBranchCondition(
+              (X86::CondCode)Cmov.getConstantOperandVal(2)),
+          dl, MVT::i8);
+    SDValue Ops[] = {St->getChain(), Src, St->getBasePtr(), CC,
+                     Cmov.getOperand(3)};
+    return DAG.getMemIntrinsicNode(X86ISD::CSTORE, dl, Tys, Ops, VT,
+                                   St->getMemOperand());
+  }
+
   // Turn load->store of MMX types into GPR load/stores.  This avoids clobbering
   // the FP state in cases where an emms may be missing.
   // A preferable solution to the general problem is to figure out the right

llvm/test/CodeGen/X86/apx/cfcmov.ll

@@ -93,3 +93,91 @@ define i64 @cfcmov64rr_inv(i64 %0) {
   %3 = select i1 %2, i64 0, i64 %0
   ret i64 %3
 }
+
+define void @cfcmov16mr(ptr %p, i16 %0) {
+; CHECK-LABEL: cfcmov16mr:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movzwl (%rdi), %eax
+; CHECK-NEXT:    cmpw %ax, %si
+; CHECK-NEXT:    cfcmovlew %si, (%rdi)
+; CHECK-NEXT:    retq
+  %2 = load i16, ptr %p, align 2
+  %3 = icmp sgt i16 %0, %2
+  %4 = select i1 %3, i16 %2, i16 %0
+  store i16 %4, ptr %p, align 2
+  ret void
+}
+
+define void @cfcmov32mr(ptr %p, i32 %0) {
+; CHECK-LABEL: cfcmov32mr:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    cmpl (%rdi), %esi
+; CHECK-NEXT:    cfcmovgl %esi, (%rdi)
+; CHECK-NEXT:    retq
+  %2 = load i32, ptr %p, align 4
+  %3 = call i32 @llvm.smax.i32(i32 %0, i32 %2)
+  store i32 %3, ptr %p, align 4
+  ret void
+}
+
+define void @cfcmov64mr(ptr %p, i64 %0) {
+; CHECK-LABEL: cfcmov64mr:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    cmpq (%rdi), %rsi
+; CHECK-NEXT:    cfcmovgq %rsi, (%rdi)
+; CHECK-NEXT:    retq
+  %2 = load i64, ptr %p, align 8
+  %3 = icmp sgt i64 %0, %2
+  %4 = select i1 %3, i64 %0, i64 %2
+  store i64 %4, ptr %p, align 8
+  ret void
+}
+
+define void @volatileload(ptr %p, i32 %0) {
+; CHECK-LABEL: volatileload:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movl (%rdi), %eax
+; CHECK-NEXT:    cmpl %eax, %esi
+; CHECK-NEXT:    cmovbl %esi, %eax
+; CHECK-NEXT:    movl %eax, (%rdi)
+; CHECK-NEXT:    retq
+  %2 = load volatile i32, ptr %p, align 4
+  %3 = call i32 @llvm.umin.i32(i32 %0, i32 %2)
+  store i32 %3, ptr %p, align 4
+  ret void
+}
+
+define void @atomicstore(ptr %p, i64 %0) {
+; CHECK-LABEL: atomicstore:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movq (%rdi), %rax
+; CHECK-NEXT:    cmpq %rax, %rsi
+; CHECK-NEXT:    cmovaq %rsi, %rax
+; CHECK-NEXT:    movq %rax, (%rdi)
+; CHECK-NEXT:    retq
+  %2 = load i64, ptr %p, align 8
+  %3 = icmp ugt i64 %0, %2
+  %4 = select i1 %3, i64 %0, i64 %2
+  store atomic i64 %4, ptr %p unordered, align 8
+  ret void
+}
+
+define void @loadstorediffptr(ptr %p, i32 %0) {
+; CHECK-LABEL: loadstorediffptr:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    movl (%rdi), %eax
+; CHECK-NEXT:    cmpl %eax, %esi
+; CHECK-NEXT:    cmovbel %eax, %esi
+; CHECK-NEXT:    movl %esi, 4(%rdi)
+; CHECK-NEXT:    retq
+  %2 = getelementptr [2 x i32], ptr %p, i32 0, i32 0
+  %3 = load i32, ptr %2, align 4
+  %4 = icmp ule i32 %0, %3
+  %5 = select i1 %4, i32 %3, i32 %0
+  %6 = getelementptr [2 x i32], ptr %p, i32 0, i32 1
+  store i32 %5, ptr %6, align 4
+  ret void
+}
+
+declare i32 @llvm.smax.i32(i32, i32)
+declare i32 @llvm.umin.i32(i32, i32)