[X86] Handle BSF/BSF "zero-input fall through" behaviour

Intel docs have been updated to be similar to AMD and now describe BSF/BSF as not changing the destination register if the input value was zero, which allows us to support CTTZ/CTLZ zero-input cases by setting the destination to support a NumBits result (BSR is a bit messy as it has to XOR'd to create a CTLZ result). VIA/Zhaoxin x86_64 CPUs have also been confirmed to match this behaviour.

There are still some limits to this - its only supported for x86_64 capable processors (and I've only enabled it for x86_64 codegen), and there are some Intel CPUs that don't correctly zero the upper 32-bits of a pass through register when used for BSR32/BSF32 with a zero source value (i.e. the whole 64bits may get p[assed through).

Author: RKSimon

llvm/lib/Target/X86/X86ISelLowering.cpp

@@ -3386,15 +3386,19 @@ bool X86TargetLowering::shouldFormOverflowOp(unsigned Opcode, EVT VT,
 }
 
 bool X86TargetLowering::isCheapToSpeculateCttz(Type *Ty) const {
-  // Speculate cttz only if we can directly use TZCNT or can promote to i32/i64.
+  // Speculate cttz only if we can directly use TZCNT/CMOV, can promote to
+  // i32/i64 or can rely on BSF passthrough value.
   return Subtarget.hasBMI() || Subtarget.canUseCMOV() ||
+         Subtarget.hasBitScanPassThrough() ||
          (!Ty->isVectorTy() &&
           Ty->getScalarSizeInBits() < (Subtarget.is64Bit() ? 64u : 32u));
 }
 
 bool X86TargetLowering::isCheapToSpeculateCtlz(Type *Ty) const {
-  // Speculate ctlz only if we can directly use LZCNT.
-  return Subtarget.hasLZCNT() || Subtarget.canUseCMOV();
+  // Speculate ctlz only if we can directly use LZCNT/CMOV, or can rely on BSR
+  // passthrough value.
+  return Subtarget.hasLZCNT() || Subtarget.canUseCMOV() ||
+         Subtarget.hasBitScanPassThrough();
 }
 
 bool X86TargetLowering::ShouldShrinkFPConstant(EVT VT) const {
@@ -28694,11 +28698,18 @@ static SDValue LowerCTLZ(SDValue Op, const X86Subtarget &Subtarget,
     Op = DAG.getNode(ISD::ZERO_EXTEND, dl, OpVT, Op);
   }
 
+  // Check if we can safely pass a result though BSR for zero sources.
+  SDValue PassThru = DAG.getUNDEF(OpVT);
+  if (Opc == ISD::CTLZ && Subtarget.hasBitScanPassThrough() &&
+      !DAG.isKnownNeverZero(Op))
+    PassThru = DAG.getConstant(NumBits + NumBits - 1, dl, OpVT);
+
   // Issue a bsr (scan bits in reverse) which also sets EFLAGS.
   SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
-  Op = DAG.getNode(X86ISD::BSR, dl, VTs, Op);
+  Op = DAG.getNode(X86ISD::BSR, dl, VTs, PassThru, Op);
 
-  if (Opc == ISD::CTLZ) {
+  // Skip CMOV if we're using a pass through value.
+  if (Opc == ISD::CTLZ && PassThru.isUndef()) {
     // If src is zero (i.e. bsr sets ZF), returns NumBits.
     SDValue Ops[] = {Op, DAG.getConstant(NumBits + NumBits - 1, dl, OpVT),
                      DAG.getTargetConstant(X86::COND_E, dl, MVT::i8),
@@ -28721,16 +28732,22 @@ static SDValue LowerCTTZ(SDValue Op, const X86Subtarget &Subtarget,
   unsigned NumBits = VT.getScalarSizeInBits();
   SDValue N0 = Op.getOperand(0);
   SDLoc dl(Op);
+  bool NonZeroSrc = DAG.isKnownNeverZero(N0);
 
   assert(!VT.isVector() && Op.getOpcode() == ISD::CTTZ &&
          "Only scalar CTTZ requires custom lowering");
 
+  // Check if we can safely pass a result though BSF for zero sources.
+  SDValue PassThru = DAG.getUNDEF(VT);
+  if (!NonZeroSrc && Subtarget.hasBitScanPassThrough())
+    PassThru = DAG.getConstant(NumBits, dl, VT);
+
   // Issue a bsf (scan bits forward) which also sets EFLAGS.
   SDVTList VTs = DAG.getVTList(VT, MVT::i32);
-  Op = DAG.getNode(X86ISD::BSF, dl, VTs, N0);
+  Op = DAG.getNode(X86ISD::BSF, dl, VTs, PassThru, N0);
 
-  // If src is known never zero we can skip the CMOV.
-  if (DAG.isKnownNeverZero(N0))
+  // Skip CMOV if src is never zero or we're using a pass through value.
+  if (NonZeroSrc || !PassThru.isUndef())
     return Op;
 
   // If src is zero (i.e. bsf sets ZF), returns NumBits.
@@ -38193,12 +38210,34 @@ void X86TargetLowering::computeKnownBitsForTargetNode(const SDValue Op,
     Known = KnownBits::mul(Known, Known2);
     break;
   }
-  case X86ISD::BSR:
-    // BSR(0) is undef, but any use of BSR already accounts for non-zero inputs.
-    // Similar KnownBits behaviour to CTLZ_ZERO_UNDEF.
+  case X86ISD::BSF: {
+    Known.Zero.setBitsFrom(Log2_32(BitWidth));
+
+    KnownBits Known2;
+    Known2 = DAG.computeKnownBits(Op.getOperand(1), DemandedElts, Depth + 1);
+    if (Known2.isNonZero()) {
+      // If we have a known 1, its position is our upper bound.
+      unsigned PossibleTZ = Known2.countMaxTrailingZeros();
+      unsigned LowBits = llvm::bit_width(PossibleTZ);
+      Known.Zero.setBitsFrom(LowBits);
+    } else if (!Op.getOperand(0).isUndef()) {
+      Known2 = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
+      Known = Known.intersectWith(Known2);
+    }
+    break;
+  }
+  case X86ISD::BSR: {
     // TODO: Bound with input known bits?
     Known.Zero.setBitsFrom(Log2_32(BitWidth));
+
+    if (!Op.getOperand(0).isUndef() &&
+        !DAG.isKnownNeverZero(Op.getOperand(1), Depth + 1)) {
+      KnownBits Known2;
+      Known2 = DAG.computeKnownBits(Op.getOperand(0), DemandedElts, Depth + 1);
+      Known = Known.intersectWith(Known2);
+    }
     break;
+  }
   case X86ISD::SETCC:
     Known.Zero.setBitsFrom(1);
     break;
@@ -54243,7 +54282,7 @@ static SDValue combineXorSubCTLZ(SDNode *N, const SDLoc &DL, SelectionDAG &DAG,
   }
 
   SDVTList VTs = DAG.getVTList(OpVT, MVT::i32);
-  Op = DAG.getNode(X86ISD::BSR, DL, VTs, Op);
+  Op = DAG.getNode(X86ISD::BSR, DL, VTs, DAG.getUNDEF(OpVT), Op);
   if (VT == MVT::i8)
     Op = DAG.getNode(ISD::TRUNCATE, DL, MVT::i8, Op);
 

llvm/lib/Target/X86/X86InstrCompiler.td

@@ -2213,12 +2213,12 @@ def : Pat<(mul (loadi64 addr:$src1), i64immSExt32:$src2),
           (IMUL64rmi32 addr:$src1, i64immSExt32:$src2)>;
 
 // Bit scan instruction patterns to match explicit zero-undef behavior.
-def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr GR16:$src)>;
-def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr GR32:$src)>;
-def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr GR64:$src)>;
-def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm addr:$src)>;
-def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm addr:$src)>;
-def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm addr:$src)>;
+def : Pat<(cttz_zero_undef GR16:$src), (BSF16rr (i16 (IMPLICIT_DEF)), GR16:$src)>;
+def : Pat<(cttz_zero_undef GR32:$src), (BSF32rr (i32 (IMPLICIT_DEF)), GR32:$src)>;
+def : Pat<(cttz_zero_undef GR64:$src), (BSF64rr (i64 (IMPLICIT_DEF)), GR64:$src)>;
+def : Pat<(cttz_zero_undef (loadi16 addr:$src)), (BSF16rm (i16 (IMPLICIT_DEF)), addr:$src)>;
+def : Pat<(cttz_zero_undef (loadi32 addr:$src)), (BSF32rm (i32 (IMPLICIT_DEF)), addr:$src)>;
+def : Pat<(cttz_zero_undef (loadi64 addr:$src)), (BSF64rm (i64 (IMPLICIT_DEF)), addr:$src)>;
 
 // When HasMOVBE is enabled it is possible to get a non-legalized
 // register-register 16 bit bswap. This maps it to a ROL instruction.

llvm/lib/Target/X86/X86InstrFragments.td

@@ -134,8 +134,8 @@ def SDTX86Cmpccxadd : SDTypeProfile<1, 4, [SDTCisSameAs<0, 2>,
 def X86MFence : SDNode<"X86ISD::MFENCE", SDTNone, [SDNPHasChain]>;
 
 
-def X86bsf     : SDNode<"X86ISD::BSF",      SDTUnaryArithWithFlags>;
-def X86bsr     : SDNode<"X86ISD::BSR",      SDTUnaryArithWithFlags>;
+def X86bsf     : SDNode<"X86ISD::BSF",      SDTBinaryArithWithFlags>;
+def X86bsr     : SDNode<"X86ISD::BSR",      SDTBinaryArithWithFlags>;
 def X86fshl    : SDNode<"X86ISD::FSHL",     SDTIntShiftDOp>;
 def X86fshr    : SDNode<"X86ISD::FSHR",     SDTIntShiftDOp>;
 
@@ -685,16 +685,19 @@ def anyext_sdiv : PatFrag<(ops node:$lhs), (anyext node:$lhs),[{
 // register. Truncate can be lowered to EXTRACT_SUBREG. CopyFromReg may
 // be copying from a truncate. AssertSext/AssertZext/AssertAlign aren't saying
 // anything about the upper 32 bits, they're probably just qualifying a
-// CopyFromReg. FREEZE may be coming from a a truncate. Any other 32-bit
-// operation will zero-extend up to 64 bits.
+// CopyFromReg. FREEZE may be coming from a a truncate. BitScan fall through
+// values may not zero the upper bits correctly.
+// Any other 32-bit operation will zero-extend up to 64 bits.
 def def32 : PatLeaf<(i32 GR32:$src), [{
   return N->getOpcode() != ISD::TRUNCATE &&
          N->getOpcode() != TargetOpcode::EXTRACT_SUBREG &&
          N->getOpcode() != ISD::CopyFromReg &&
          N->getOpcode() != ISD::AssertSext &&
          N->getOpcode() != ISD::AssertZext &&
          N->getOpcode() != ISD::AssertAlign &&
-         N->getOpcode() != ISD::FREEZE;
+         N->getOpcode() != ISD::FREEZE &&
+         !((N->getOpcode() == X86ISD::BSF|| N->getOpcode() == X86ISD::BSR) &&
+           (!N->getOperand(0).isUndef() && !isa<ConstantSDNode>(N->getOperand(0))));
 }]>;
 
 // Treat an 'or' node is as an 'add' if the or'ed bits are known to be zero.

llvm/lib/Target/X86/X86InstrInfo.cpp

@@ -5221,42 +5221,42 @@ inline static bool isDefConvertible(const MachineInstr &MI, bool &NoSignFlag,
 }
 
 /// Check whether the use can be converted to remove a comparison against zero.
-static X86::CondCode isUseDefConvertible(const MachineInstr &MI) {
+static std::pair<X86::CondCode, unsigned> isUseDefConvertible(const MachineInstr &MI) {
   switch (MI.getOpcode()) {
   default:
-    return X86::COND_INVALID;
+    return std::make_pair(X86::COND_INVALID, ~0U);
   CASE_ND(NEG8r)
   CASE_ND(NEG16r)
   CASE_ND(NEG32r)
   CASE_ND(NEG64r)
-    return X86::COND_AE;
+    return std::make_pair(X86::COND_AE, 1U);
   case X86::LZCNT16rr:
   case X86::LZCNT32rr:
   case X86::LZCNT64rr:
-    return X86::COND_B;
+    return std::make_pair(X86::COND_B, 1U);
   case X86::POPCNT16rr:
   case X86::POPCNT32rr:
   case X86::POPCNT64rr:
-    return X86::COND_E;
+    return std::make_pair(X86::COND_E, 1U);
   case X86::TZCNT16rr:
   case X86::TZCNT32rr:
   case X86::TZCNT64rr:
-    return X86::COND_B;
+    return std::make_pair(X86::COND_B, 1U);
   case X86::BSF16rr:
   case X86::BSF32rr:
   case X86::BSF64rr:
   case X86::BSR16rr:
   case X86::BSR32rr:
   case X86::BSR64rr:
-    return X86::COND_E;
+    return std::make_pair(X86::COND_E, 2U);
   case X86::BLSI32rr:
   case X86::BLSI64rr:
-    return X86::COND_AE;
+    return std::make_pair(X86::COND_AE, 1U);
   case X86::BLSR32rr:
   case X86::BLSR64rr:
   case X86::BLSMSK32rr:
   case X86::BLSMSK64rr:
-    return X86::COND_B;
+    return std::make_pair(X86::COND_B, 1U);
     // TODO: TBM instructions.
   }
 }
@@ -5337,6 +5337,7 @@ bool X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
   bool ClearsOverflowFlag = false;
   bool ShouldUpdateCC = false;
   bool IsSwapped = false;
+  unsigned OpNo = 0;
   X86::CondCode NewCC = X86::COND_INVALID;
   int64_t ImmDelta = 0;
 
@@ -5392,9 +5393,9 @@ bool X86InstrInfo::optimizeCompareInstr(MachineInstr &CmpInstr, Register SrcReg,
         //      ...                 // EFLAGS not changed
         //      testl %eax, %eax    // <-- can be removed
         if (IsCmpZero) {
-          NewCC = isUseDefConvertible(Inst);
-          if (NewCC != X86::COND_INVALID && Inst.getOperand(1).isReg() &&
-              Inst.getOperand(1).getReg() == SrcReg) {
+          std::tie(NewCC, OpNo) = isUseDefConvertible(Inst);
+          if (NewCC != X86::COND_INVALID && Inst.getOperand(OpNo).isReg() &&
+              Inst.getOperand(OpNo).getReg() == SrcReg) {
             ShouldUpdateCC = true;
             MI = &Inst;
             break;

llvm/lib/Target/X86/X86InstrMisc.td

@@ -247,55 +247,55 @@ def BSWAP64r : RI<0xC8, AddRegFrm, (outs GR64:$dst), (ins GR64:$src),
 } // Constraints = "$src = $dst", SchedRW
 
 // Bit scan instructions.
-let Defs = [EFLAGS] in {
-def BSF16rr  : I<0xBC, MRMSrcReg, (outs GR16:$dst), (ins GR16:$src),
+let Defs = [EFLAGS], Constraints = "$fallback = $dst" in {
+def BSF16rr  : I<0xBC, MRMSrcReg, (outs GR16:$dst), (ins GR16:$fallback, GR16:$src),
                  "bsf{w}\t{$src, $dst|$dst, $src}",
-                 [(set GR16:$dst, EFLAGS, (X86bsf GR16:$src))]>,
+                 [(set GR16:$dst, EFLAGS, (X86bsf GR16:$fallback, GR16:$src))]>,
                   TB, OpSize16, Sched<[WriteBSF]>;
-def BSF16rm  : I<0xBC, MRMSrcMem, (outs GR16:$dst), (ins i16mem:$src),
+def BSF16rm  : I<0xBC, MRMSrcMem, (outs GR16:$dst), (ins GR16:$fallback, i16mem:$src),
                  "bsf{w}\t{$src, $dst|$dst, $src}",
-                 [(set GR16:$dst, EFLAGS, (X86bsf (loadi16 addr:$src)))]>,
+                 [(set GR16:$dst, EFLAGS, (X86bsf GR16:$fallback, (loadi16 addr:$src)))]>,
                  TB, OpSize16, Sched<[WriteBSFLd]>;
-def BSF32rr  : I<0xBC, MRMSrcReg, (outs GR32:$dst), (ins GR32:$src),
+def BSF32rr  : I<0xBC, MRMSrcReg, (outs GR32:$dst), (ins GR32:$fallback, GR32:$src),
                  "bsf{l}\t{$src, $dst|$dst, $src}",
-                 [(set GR32:$dst, EFLAGS, (X86bsf GR32:$src))]>,
+                 [(set GR32:$dst, EFLAGS, (X86bsf GR32:$fallback, GR32:$src))]>,
                  TB, OpSize32, Sched<[WriteBSF]>;
-def BSF32rm  : I<0xBC, MRMSrcMem, (outs GR32:$dst), (ins i32mem:$src),
+def BSF32rm  : I<0xBC, MRMSrcMem, (outs GR32:$dst), (ins GR32:$fallback, i32mem:$src),
                  "bsf{l}\t{$src, $dst|$dst, $src}",
-                 [(set GR32:$dst, EFLAGS, (X86bsf (loadi32 addr:$src)))]>,
+                 [(set GR32:$dst, EFLAGS, (X86bsf GR32:$fallback, (loadi32 addr:$src)))]>,
                  TB, OpSize32, Sched<[WriteBSFLd]>;
-def BSF64rr  : RI<0xBC, MRMSrcReg, (outs GR64:$dst), (ins GR64:$src),
+def BSF64rr  : RI<0xBC, MRMSrcReg, (outs GR64:$dst), (ins GR64:$fallback, GR64:$src),
                   "bsf{q}\t{$src, $dst|$dst, $src}",
-                  [(set GR64:$dst, EFLAGS, (X86bsf GR64:$src))]>,
+                  [(set GR64:$dst, EFLAGS, (X86bsf GR64:$fallback, GR64:$src))]>,
                   TB, Sched<[WriteBSF]>;
-def BSF64rm  : RI<0xBC, MRMSrcMem, (outs GR64:$dst), (ins i64mem:$src),
+def BSF64rm  : RI<0xBC, MRMSrcMem, (outs GR64:$dst), (ins GR64:$fallback, i64mem:$src),
                   "bsf{q}\t{$src, $dst|$dst, $src}",
-                  [(set GR64:$dst, EFLAGS, (X86bsf (loadi64 addr:$src)))]>,
+                  [(set GR64:$dst, EFLAGS, (X86bsf GR64:$fallback, (loadi64 addr:$src)))]>,
                   TB, Sched<[WriteBSFLd]>;
 
-def BSR16rr  : I<0xBD, MRMSrcReg, (outs GR16:$dst), (ins GR16:$src),
+def BSR16rr  : I<0xBD, MRMSrcReg, (outs GR16:$dst), (ins GR16:$fallback, GR16:$src),
                  "bsr{w}\t{$src, $dst|$dst, $src}",
-                 [(set GR16:$dst, EFLAGS, (X86bsr GR16:$src))]>,
+                 [(set GR16:$dst, EFLAGS, (X86bsr GR16:$fallback, GR16:$src))]>,
                  TB, OpSize16, Sched<[WriteBSR]>;
-def BSR16rm  : I<0xBD, MRMSrcMem, (outs GR16:$dst), (ins i16mem:$src),
+def BSR16rm  : I<0xBD, MRMSrcMem, (outs GR16:$dst), (ins GR16:$fallback, i16mem:$src),
                  "bsr{w}\t{$src, $dst|$dst, $src}",
-                 [(set GR16:$dst, EFLAGS, (X86bsr (loadi16 addr:$src)))]>,
+                 [(set GR16:$dst, EFLAGS, (X86bsr GR16:$fallback, (loadi16 addr:$src)))]>,
                  TB, OpSize16, Sched<[WriteBSRLd]>;
-def BSR32rr  : I<0xBD, MRMSrcReg, (outs GR32:$dst), (ins GR32:$src),
+def BSR32rr  : I<0xBD, MRMSrcReg, (outs GR32:$dst), (ins GR32:$fallback, GR32:$src),
                  "bsr{l}\t{$src, $dst|$dst, $src}",
-                 [(set GR32:$dst, EFLAGS, (X86bsr GR32:$src))]>,
+                 [(set GR32:$dst, EFLAGS, (X86bsr GR32:$fallback, GR32:$src))]>,
                  TB, OpSize32, Sched<[WriteBSR]>;
-def BSR32rm  : I<0xBD, MRMSrcMem, (outs GR32:$dst), (ins i32mem:$src),
+def BSR32rm  : I<0xBD, MRMSrcMem, (outs GR32:$dst), (ins GR32:$fallback, i32mem:$src),
                  "bsr{l}\t{$src, $dst|$dst, $src}",
-                 [(set GR32:$dst, EFLAGS, (X86bsr (loadi32 addr:$src)))]>,
+                 [(set GR32:$dst, EFLAGS, (X86bsr GR32:$fallback, (loadi32 addr:$src)))]>,
                  TB, OpSize32, Sched<[WriteBSRLd]>;
-def BSR64rr  : RI<0xBD, MRMSrcReg, (outs GR64:$dst), (ins GR64:$src),
+def BSR64rr  : RI<0xBD, MRMSrcReg, (outs GR64:$dst), (ins GR64:$fallback, GR64:$src),
                   "bsr{q}\t{$src, $dst|$dst, $src}",
-                  [(set GR64:$dst, EFLAGS, (X86bsr GR64:$src))]>,
+                  [(set GR64:$dst, EFLAGS, (X86bsr GR64:$fallback, GR64:$src))]>,
                   TB, Sched<[WriteBSR]>;
-def BSR64rm  : RI<0xBD, MRMSrcMem, (outs GR64:$dst), (ins i64mem:$src),
+def BSR64rm  : RI<0xBD, MRMSrcMem, (outs GR64:$dst), (ins GR64:$fallback, i64mem:$src),
                   "bsr{q}\t{$src, $dst|$dst, $src}",
-                  [(set GR64:$dst, EFLAGS, (X86bsr (loadi64 addr:$src)))]>,
+                  [(set GR64:$dst, EFLAGS, (X86bsr GR64:$fallback, (loadi64 addr:$src)))]>,
                   TB, Sched<[WriteBSRLd]>;
 } // Defs = [EFLAGS]
 

llvm/lib/Target/X86/X86Subtarget.h

@@ -263,6 +263,11 @@ class X86Subtarget final : public X86GenSubtargetInfo {
     return hasBWI() && useAVX512Regs();
   }
 
+  // Returns true if the destination register of a BSF/BSR instruction is
+  // not touched if the source register is zero.
+  // NOTE: i32->i64 implicit zext isn't guaranteed by BSR/BSF pass through.
+  bool hasBitScanPassThrough() const { return is64Bit(); }
+
   bool isXRaySupported() const override { return is64Bit(); }
 
   /// Use clflush if we have SSE2 or we're on x86-64 (even if we asked for

llvm/test/CodeGen/X86/bit_ceil.ll

@@ -10,9 +10,8 @@ define i32 @bit_ceil_i32(i32 %x) {
 ; NOBMI:       # %bb.0:
 ; NOBMI-NEXT:    # kill: def $edi killed $edi def $rdi
 ; NOBMI-NEXT:    leal -1(%rdi), %eax
-; NOBMI-NEXT:    bsrl %eax, %eax
 ; NOBMI-NEXT:    movl $63, %ecx
-; NOBMI-NEXT:    cmovnel %eax, %ecx
+; NOBMI-NEXT:    bsrl %eax, %ecx
 ; NOBMI-NEXT:    xorl $31, %ecx
 ; NOBMI-NEXT:    negb %cl
 ; NOBMI-NEXT:    movl $1, %edx
@@ -47,9 +46,8 @@ define i32 @bit_ceil_i32(i32 %x) {
 define i32 @bit_ceil_i32_plus1(i32 noundef %x) {
 ; NOBMI-LABEL: bit_ceil_i32_plus1:
 ; NOBMI:       # %bb.0: # %entry
-; NOBMI-NEXT:    bsrl %edi, %eax
 ; NOBMI-NEXT:    movl $63, %ecx
-; NOBMI-NEXT:    cmovnel %eax, %ecx
+; NOBMI-NEXT:    bsrl %edi, %ecx
 ; NOBMI-NEXT:    xorl $31, %ecx
 ; NOBMI-NEXT:    negb %cl
 ; NOBMI-NEXT:    movl $1, %edx
@@ -86,9 +84,8 @@ define i64 @bit_ceil_i64(i64 %x) {
 ; NOBMI-LABEL: bit_ceil_i64:
 ; NOBMI:       # %bb.0:
 ; NOBMI-NEXT:    leaq -1(%rdi), %rax
-; NOBMI-NEXT:    bsrq %rax, %rax
 ; NOBMI-NEXT:    movl $127, %ecx
-; NOBMI-NEXT:    cmovneq %rax, %rcx
+; NOBMI-NEXT:    bsrq %rax, %rcx
 ; NOBMI-NEXT:    xorl $63, %ecx
 ; NOBMI-NEXT:    negb %cl
 ; NOBMI-NEXT:    movl $1, %edx
@@ -122,9 +119,8 @@ define i64 @bit_ceil_i64(i64 %x) {
 define i64 @bit_ceil_i64_plus1(i64 noundef %x) {
 ; NOBMI-LABEL: bit_ceil_i64_plus1:
 ; NOBMI:       # %bb.0: # %entry
-; NOBMI-NEXT:    bsrq %rdi, %rax
 ; NOBMI-NEXT:    movl $127, %ecx
-; NOBMI-NEXT:    cmovneq %rax, %rcx
+; NOBMI-NEXT:    bsrq %rdi, %rcx
 ; NOBMI-NEXT:    xorl $63, %ecx
 ; NOBMI-NEXT:    negb %cl
 ; NOBMI-NEXT:    movl $1, %edx