[DAGCombine] Use `IsKnownNeverZero` to see if we need zero-check in is_pow2 setcc patern

[goldstein.w.n]

Summary

ctpop(X) eq/ne 1 is checking if X is a non-zero power of 2. Power of
2 check including zero is (X & (X-1)) eq/ne 0 and unfortunately
there is no good pattern for checking a power of 2 while excluding
zero. So, when lowering ctpop(X) eq/ne 1, explicitly check
IsKnownNeverZero(X) to maybe be able to optimize out the extra zero
check.

We need this explicitly as DAGCombiner does not re-analyze provable
setcc nodes, and the middle-end never finds it beneficially to broaden
ctpop(X) eq/ne 1 -> ctpop(X) ule/ugt 1 (power of 2 including
zero).

[RKSimon]

LGTM

llvm/lib/CodeGen/SelectionDAG/TargetLowering.cpp

if (TLI.isOperationLegal(ISD::CTPOP, CTVT))
       return SDValue();
 
     SDValue Zero = DAG.getConstant(0, dl, CTVT);
     SDValue NegOne = DAG.getAllOnesConstant(dl, CTVT);
     assert(CTVT.isInteger());
     ISD::CondCode InvCond = ISD::getSetCCInverse(Cond, CTVT);
     SDValue Add = DAG.getNode(ISD::ADD, dl, CTVT, CTOp, NegOne);
     SDValue And = DAG.getNode(ISD::AND, dl, CTVT, CTOp, Add);
-    SDValue LHS = DAG.getSetCC(dl, VT, CTOp, Zero, InvCond);
     SDValue RHS = DAG.getSetCC(dl, VT, And, Zero, Cond);
+    // Its not uncommon for known-never-zero X to exist in (ctpop X) eq/ne 1, so
+    // check before the emit a potentially unnecessary op.
+    if (DAG.isKnownNeverZero(CTOp))
+      return RHS;
+    SDValue LHS = DAG.getSetCC(dl, VT, CTOp, Zero, InvCond);
     unsigned LogicOpcode = Cond == ISD::SETEQ ? ISD::AND : ISD::OR;
     return DAG.getNode(LogicOpcode, dl, VT, LHS, RHS);
   }
 
   return SDValue();
 }
 
 static SDValue foldSetCCWithRotate(EVT VT, SDValue N0, SDValue N1,

llvm/test/CodeGen/X86/ispow2.ll

 ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown | FileCheck %s --check-prefixes=CHECK,CHECK-NOBMI
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx2,+bmi2,+bmi | FileCheck %s --check-prefixes=CHECK,CHECK-BMI2,CHECK-AVX2
 ; RUN: llc < %s -mtriple=x86_64-unknown-unknown -mattr=+avx512f,+avx512vl,+avx512vpopcntdq,+bmi2,+bmi | FileCheck %s --check-prefixes=CHECK,CHECK-BMI2,CHECK-AVX512
 
 declare i32 @llvm.ctpop.i32(i32)
 declare <4 x i64> @llvm.ctpop.v4i64(<4 x i64>)
 define i1 @is_pow2_non_zero(i32 %xin) {
 ; CHECK-NOBMI-LABEL: is_pow2_non_zero:
 ; CHECK-NOBMI:       # %bb.0:
 ; CHECK-NOBMI-NEXT:    # kill: def $edi killed $edi def $rdi
 ; CHECK-NOBMI-NEXT:    orl $256, %edi # imm = 0x100
 ; CHECK-NOBMI-NEXT:    leal -1(%rdi), %eax
 ; CHECK-NOBMI-NEXT:    testl %eax, %edi
-; CHECK-NOBMI-NEXT:    sete %cl
-; CHECK-NOBMI-NEXT:    testl %edi, %edi
-; CHECK-NOBMI-NEXT:    setne %al
-; CHECK-NOBMI-NEXT:    andb %cl, %al
+; CHECK-NOBMI-NEXT:    sete %al
 ; CHECK-NOBMI-NEXT:    retq
 ;
 ; CHECK-BMI2-LABEL: is_pow2_non_zero:
 ; CHECK-BMI2:       # %bb.0:
 ; CHECK-BMI2-NEXT:    orl $256, %edi # imm = 0x100
-; CHECK-BMI2-NEXT:    setne %cl
 ; CHECK-BMI2-NEXT:    blsrl %edi, %eax
 ; CHECK-BMI2-NEXT:    sete %al
-; CHECK-BMI2-NEXT:    andb %cl, %al
 ; CHECK-BMI2-NEXT:    retq
   %x = or i32 %xin, 256
   %cnt = call i32 @llvm.ctpop.i32(i32 %x)
   %r = icmp eq i32 %cnt, 1
   ret i1 %r
 }
 
 define i1 @is_pow2_non_zero_x_maybe_z(i32 %x) {

 
 define i1 @neither_pow2_non_zero(i32 %xin) {
 ; CHECK-NOBMI-LABEL: neither_pow2_non_zero:
 ; CHECK-NOBMI:       # %bb.0:
 ; CHECK-NOBMI-NEXT:    # kill: def $edi killed $edi def $rdi
 ; CHECK-NOBMI-NEXT:    orl $256, %edi # imm = 0x100
 ; CHECK-NOBMI-NEXT:    leal -1(%rdi), %eax
 ; CHECK-NOBMI-NEXT:    testl %eax, %edi
-; CHECK-NOBMI-NEXT:    setne %cl
-; CHECK-NOBMI-NEXT:    testl %edi, %edi
-; CHECK-NOBMI-NEXT:    sete %al
-; CHECK-NOBMI-NEXT:    orb %cl, %al
+; CHECK-NOBMI-NEXT:    setne %al
 ; CHECK-NOBMI-NEXT:    retq
 ;
 ; CHECK-BMI2-LABEL: neither_pow2_non_zero:
 ; CHECK-BMI2:       # %bb.0:
 ; CHECK-BMI2-NEXT:    orl $256, %edi # imm = 0x100
-; CHECK-BMI2-NEXT:    sete %cl
 ; CHECK-BMI2-NEXT:    blsrl %edi, %eax
 ; CHECK-BMI2-NEXT:    setne %al
-; CHECK-BMI2-NEXT:    orb %cl, %al
 ; CHECK-BMI2-NEXT:    retq
   %x = or i32 %xin, 256
   %cnt = call i32 @llvm.ctpop.i32(i32 %x)
   %r = icmp ne i32 %cnt, 1
   ret i1 %r
 }
 
 define <4 x i1> @is_pow2_non_zero_4xv64(<4 x i64> %xin) {
 ; CHECK-NOBMI-LABEL: is_pow2_non_zero_4xv64:
 ; CHECK-NOBMI:       # %bb.0:
 ; CHECK-NOBMI-NEXT:    movdqa {{.*#+}} xmm2 = [256,256]
 ; CHECK-NOBMI-NEXT:    por %xmm2, %xmm0
 ; CHECK-NOBMI-NEXT:    por %xmm2, %xmm1
 ; CHECK-NOBMI-NEXT:    pcmpeqd %xmm2, %xmm2
 ; CHECK-NOBMI-NEXT:    movdqa %xmm1, %xmm3
 ; CHECK-NOBMI-NEXT:    paddq %xmm2, %xmm3
 ; CHECK-NOBMI-NEXT:    pand %xmm1, %xmm3
-; CHECK-NOBMI-NEXT:    pxor %xmm4, %xmm4
-; CHECK-NOBMI-NEXT:    pcmpeqd %xmm4, %xmm3
-; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[1,0,3,2]
-; CHECK-NOBMI-NEXT:    pand %xmm3, %xmm5
-; CHECK-NOBMI-NEXT:    pcmpeqd %xmm4, %xmm1
-; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[1,0,3,2]
-; CHECK-NOBMI-NEXT:    pand %xmm1, %xmm3
-; CHECK-NOBMI-NEXT:    pandn %xmm5, %xmm3
+; CHECK-NOBMI-NEXT:    pxor %xmm1, %xmm1
+; CHECK-NOBMI-NEXT:    pcmpeqd %xmm1, %xmm3
+; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,0,3,2]
+; CHECK-NOBMI-NEXT:    pand %xmm3, %xmm4
 ; CHECK-NOBMI-NEXT:    paddq %xmm0, %xmm2
-; CHECK-NOBMI-NEXT:    pand %xmm0, %xmm2
-; CHECK-NOBMI-NEXT:    pcmpeqd %xmm4, %xmm2
-; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm1 = xmm2[1,0,3,2]
-; CHECK-NOBMI-NEXT:    pand %xmm2, %xmm1
-; CHECK-NOBMI-NEXT:    pcmpeqd %xmm4, %xmm0
-; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm2 = xmm0[1,0,3,2]
 ; CHECK-NOBMI-NEXT:    pand %xmm2, %xmm0
-; CHECK-NOBMI-NEXT:    pandn %xmm1, %xmm0
-; CHECK-NOBMI-NEXT:    packssdw %xmm3, %xmm0
+; CHECK-NOBMI-NEXT:    pcmpeqd %xmm1, %xmm0
+; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
+; CHECK-NOBMI-NEXT:    pand %xmm1, %xmm0
+; CHECK-NOBMI-NEXT:    packssdw %xmm4, %xmm0
 ; CHECK-NOBMI-NEXT:    retq
 ;
 ; CHECK-AVX2-LABEL: is_pow2_non_zero_4xv64:
 ; CHECK-AVX2:       # %bb.0:
 ; CHECK-AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [256,256,256,256]
 ; CHECK-AVX2-NEXT:    vpor %ymm1, %ymm0, %ymm0
 ; CHECK-AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
 ; CHECK-AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm1

 ; CHECK-NOBMI:       # %bb.0:
 ; CHECK-NOBMI-NEXT:    movdqa {{.*#+}} xmm2 = [256,256]
 ; CHECK-NOBMI-NEXT:    por %xmm2, %xmm0
 ; CHECK-NOBMI-NEXT:    por %xmm2, %xmm1
 ; CHECK-NOBMI-NEXT:    pcmpeqd %xmm2, %xmm2
 ; CHECK-NOBMI-NEXT:    movdqa %xmm1, %xmm3
 ; CHECK-NOBMI-NEXT:    paddq %xmm2, %xmm3
 ; CHECK-NOBMI-NEXT:    pand %xmm1, %xmm3
-; CHECK-NOBMI-NEXT:    pxor %xmm4, %xmm4
-; CHECK-NOBMI-NEXT:    pcmpeqd %xmm4, %xmm3
-; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm5 = xmm3[1,0,3,2]
-; CHECK-NOBMI-NEXT:    pand %xmm3, %xmm5
-; CHECK-NOBMI-NEXT:    pxor %xmm2, %xmm5
-; CHECK-NOBMI-NEXT:    pcmpeqd %xmm4, %xmm1
-; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm3 = xmm1[1,0,3,2]
-; CHECK-NOBMI-NEXT:    pand %xmm1, %xmm3
-; CHECK-NOBMI-NEXT:    por %xmm5, %xmm3
-; CHECK-NOBMI-NEXT:    movdqa %xmm0, %xmm1
-; CHECK-NOBMI-NEXT:    paddq %xmm2, %xmm1
-; CHECK-NOBMI-NEXT:    pand %xmm0, %xmm1
-; CHECK-NOBMI-NEXT:    pcmpeqd %xmm4, %xmm1
-; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm5 = xmm1[1,0,3,2]
-; CHECK-NOBMI-NEXT:    pand %xmm1, %xmm5
-; CHECK-NOBMI-NEXT:    pxor %xmm2, %xmm5
-; CHECK-NOBMI-NEXT:    pcmpeqd %xmm4, %xmm0
+; CHECK-NOBMI-NEXT:    pxor %xmm1, %xmm1
+; CHECK-NOBMI-NEXT:    pcmpeqd %xmm1, %xmm3
+; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm4 = xmm3[1,0,3,2]
+; CHECK-NOBMI-NEXT:    pand %xmm3, %xmm4
+; CHECK-NOBMI-NEXT:    pxor %xmm2, %xmm4
+; CHECK-NOBMI-NEXT:    movdqa %xmm0, %xmm3
+; CHECK-NOBMI-NEXT:    paddq %xmm2, %xmm3
+; CHECK-NOBMI-NEXT:    pand %xmm3, %xmm0
+; CHECK-NOBMI-NEXT:    pcmpeqd %xmm1, %xmm0
 ; CHECK-NOBMI-NEXT:    pshufd {{.*#+}} xmm1 = xmm0[1,0,3,2]
 ; CHECK-NOBMI-NEXT:    pand %xmm1, %xmm0
-; CHECK-NOBMI-NEXT:    por %xmm5, %xmm0
-; CHECK-NOBMI-NEXT:    packssdw %xmm3, %xmm0
+; CHECK-NOBMI-NEXT:    pxor %xmm2, %xmm0
+; CHECK-NOBMI-NEXT:    packssdw %xmm4, %xmm0
 ; CHECK-NOBMI-NEXT:    retq
 ;
 ; CHECK-AVX2-LABEL: neither_pow2_non_zero_4xv64:
 ; CHECK-AVX2:       # %bb.0:
 ; CHECK-AVX2-NEXT:    vpbroadcastq {{.*#+}} ymm1 = [256,256,256,256]
 ; CHECK-AVX2-NEXT:    vpor %ymm1, %ymm0, %ymm0
 ; CHECK-AVX2-NEXT:    vpcmpeqd %ymm1, %ymm1, %ymm1
 ; CHECK-AVX2-NEXT:    vpaddq %ymm1, %ymm0, %ymm2