[X86] Fold AND(SRL(X,Y),1) -> SETCC(BT(X,Y)) (RECOMMITTED)

As noticed on PR39174, if we're extracting a single non-constant bit index, then try to use BT+SETCC instead to avoid messing around moving the shift amount to the ECX register, using slow x86 shift ops etc.

Recommitted with a fix to ensure we zext/trunc the SETCC result to the original type.

Differential Revision: https://reviews.llvm.org/D122891

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -47329,6 +47329,19 @@ static SDValue combineAnd(SDNode *N, SelectionDAG &DAG,
   if (SDValue R = combineAndLoadToBZHI(N, DAG, Subtarget))
     return R;
 
+  // Fold AND(SRL(X,Y),1) -> SETCC(BT(X,Y), COND_B) iff Y is not a constant
+  // avoids slow variable shift (moving shift amount to ECX etc.)
+  if (isOneConstant(N1) && N0->hasOneUse()) {
+    SDValue Src = N0;
+    while ((Src.getOpcode() == ISD::ZERO_EXTEND ||
+            Src.getOpcode() == ISD::TRUNCATE) &&
+           Src.getOperand(0)->hasOneUse())
+      Src = Src.getOperand(0);
+    if (Src.getOpcode() == ISD::SRL && !isa<ConstantSDNode>(Src.getOperand(1)))
+      if (SDValue BT = getBT(Src.getOperand(0), Src.getOperand(1), dl, DAG))
+        return DAG.getZExtOrTrunc(getSETCC(X86::COND_B, BT, dl, DAG), dl, VT);
+  }
+
   if (VT.isVector() && (VT.getScalarSizeInBits() % 8) == 0) {
     // Attempt to recursively combine a bitmask AND with shuffles.
     SDValue Op(N, 0);

llvm/test/CodeGen/X86/setcc.ll

@@ -139,19 +139,17 @@ define zeroext i1 @t6(i32 %a) #0 {
 define zeroext i1 @t7(i32 %0) {
 ; X86-LABEL: t7:
 ; X86:       ## %bb.0:
-; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    movb $19, %al
+; X86-NEXT:    movl $19, %ecx
-; X86-NEXT:    shrb %cl, %al
+; X86-NEXT:    btl %eax, %ecx
-; X86-NEXT:    andb $1, %al
+; X86-NEXT:    setb %al
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: t7:
 ; X64:       ## %bb.0:
-; X64-NEXT:    movl %edi, %ecx
+; X64-NEXT:    movl $19, %eax
-; X64-NEXT:    movb $19, %al
+; X64-NEXT:    btl %edi, %eax
-; X64-NEXT:    ## kill: def $cl killed $cl killed $ecx
+; X64-NEXT:    setb %al
-; X64-NEXT:    shrb %cl, %al
-; X64-NEXT:    andb $1, %al
 ; X64-NEXT:    retq
   %2 = trunc i32 %0 to i5
   %3 = lshr i5 -13, %2
@@ -163,20 +161,16 @@ define zeroext i1 @t7(i32 %0) {
 define zeroext i1 @t8(i8 %0, i8 %1) {
 ; X86-LABEL: t8:
 ; X86:       ## %bb.0:
-; X86-NEXT:    movb {{[0-9]+}}(%esp), %cl
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %eax
-; X86-NEXT:    movb {{[0-9]+}}(%esp), %al
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
-; X86-NEXT:    shrb %cl, %al
+; X86-NEXT:    btl %eax, %ecx
-; X86-NEXT:    andb $1, %al
+; X86-NEXT:    setb %al
 ; X86-NEXT:    retl
 ;
 ; X64-LABEL: t8:
 ; X64:       ## %bb.0:
-; X64-NEXT:    movl %esi, %ecx
+; X64-NEXT:    btl %esi, %edi
-; X64-NEXT:    movl %edi, %eax
+; X64-NEXT:    setb %al
-; X64-NEXT:    ## kill: def $cl killed $cl killed $ecx
-; X64-NEXT:    shrb %cl, %al
-; X64-NEXT:    andb $1, %al
-; X64-NEXT:    ## kill: def $al killed $al killed $eax
 ; X64-NEXT:    retq
   %3 = lshr i8 %0, %1
   %4 = and i8 %3, 1
@@ -184,6 +178,30 @@ define zeroext i1 @t8(i8 %0, i8 %1) {
   ret i1 %5
 }
 
+define i64 @t9(i32 %0, i32 %1) {
+; X86-LABEL: t9:
+; X86:       ## %bb.0:
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %ecx
+; X86-NEXT:    movl {{[0-9]+}}(%esp), %edx
+; X86-NEXT:    xorl %eax, %eax
+; X86-NEXT:    btl %edx, %ecx
+; X86-NEXT:    setb %al
+; X86-NEXT:    xorl %edx, %edx
+; X86-NEXT:    retl
+;
+; X64-LABEL: t9:
+; X64:       ## %bb.0:
+; X64-NEXT:    xorl %eax, %eax
+; X64-NEXT:    btl %esi, %edi
+; X64-NEXT:    setb %al
+; X64-NEXT:    retq
+  %3 = lshr i32 %0, %1
+  %4 = and i32 %3, 1
+  %5 = icmp ne i32 %4, 0
+  %6 = zext i1 %5 to i64
+  ret i64 %6
+}
+
 define i16 @shift_and(i16 %a) {
 ; X86-LABEL: shift_and:
 ; X86:       ## %bb.0: