[llvm] [GlobalISel] Handle div-by-pow2 (PR #83155)

[Jay Foad via llvm-commits]

================
@@ -5270,6 +5270,96 @@ MachineInstr *CombinerHelper::buildSDivUsingMul(MachineInstr &MI) {
   return MIB.buildMul(Ty, Res, Factor);
 }
 
+bool CombinerHelper::matchDivByPow2(MachineInstr &MI, bool IsSigned) {
+  assert((MI.getOpcode() == TargetOpcode::G_SDIV ||
+          MI.getOpcode() == TargetOpcode::G_UDIV) &&
+         "Expected SDIV or UDIV");
+  auto &Div = cast<GenericMachineInstr>(MI);
+  Register RHS = Div.getReg(2);
+  auto MatchPow2 = [&](const Constant *C) {
+    auto *CI = dyn_cast<ConstantInt>(C);
+    return CI && (CI->getValue().isPowerOf2() ||
+                  (IsSigned && CI->getValue().isNegatedPowerOf2()));
+  };
+  return matchUnaryPredicate(MRI, RHS, MatchPow2, /*AllowUndefs=*/false);
+}
+
+void CombinerHelper::applySDivByPow2(MachineInstr &MI) {
+  assert(MI.getOpcode() == TargetOpcode::G_SDIV && "Expected SDIV");
+  auto &SDiv = cast<GenericMachineInstr>(MI);
+  Register Dst = SDiv.getReg(0);
+  Register LHS = SDiv.getReg(1);
+  Register RHS = SDiv.getReg(2);
+  LLT Ty = MRI.getType(Dst);
+  LLT ShiftAmtTy = getTargetLowering().getPreferredShiftAmountTy(Ty);
+
+  Builder.setInstrAndDebugLoc(MI);
+
+  // Effectively we want to lower G_SDIV %lhs, %rhs, where %rhs is a power of 2,
+  // to the following version:
+  //
+  // %c1 = G_CTTZ %rhs
+  // %inexact = G_SUB $bitwidth, %c1
+  // %sign = %G_ASHR %lhs, $(bitwidth - 1)
+  // %srl = G_SHR %sign, %inexact
+  // %add = G_ADD %lhs, %srl
+  // %sra = G_ASHR %add, %c1
+  // %sra = G_SELECT, %isoneorallones, %lhs, %sra
+  // %zero = G_CONSTANT $0
+  // %neg = G_NEG %sra
+  // %isneg = G_ICMP SLT %lhs, %zero
+  // %res = G_SELECT %isneg, %neg, %sra
+
+  unsigned Bitwidth = Ty.getScalarSizeInBits();
+  auto Zero = Builder.buildConstant(Ty, 0);
+
+  auto Bits = Builder.buildConstant(ShiftAmtTy, Bitwidth);
+  auto C1 = Builder.buildCTTZ(ShiftAmtTy, RHS);
+  auto Inexact = Builder.buildSub(ShiftAmtTy, Bits, C1);
+  auto Sign = Builder.buildAShr(
+      Ty, LHS, Builder.buildConstant(ShiftAmtTy, Bitwidth - 1));
+
+  // Add (LHS < 0) ? abs2 - 1 : 0;
+  auto Lshr = Builder.buildLShr(Ty, Sign, Inexact);
+  auto Add = Builder.buildAdd(Ty, LHS, Lshr);
+  auto Shr = Builder.buildAShr(Ty, Add, C1);
+
+  LLT CCVT =
+      Ty.isVector() ? LLT::vector(Ty.getElementCount(), 1) : LLT::scalar(1);
+
+  auto One = Builder.buildConstant(Ty, 1);
+  auto AllOnes =
+      Builder.buildConstant(Ty, APInt::getAllOnes(Ty.getScalarSizeInBits()));
+  auto IsOne = Builder.buildICmp(CmpInst::Predicate::ICMP_EQ, CCVT, RHS, One);
+  auto IsAllOnes =
+      Builder.buildICmp(CmpInst::Predicate::ICMP_EQ, CCVT, RHS, AllOnes);
+  auto IsOneOrAllOnes = Builder.buildOr(CCVT, IsOne, IsAllOnes);
+  Shr = Builder.buildSelect(Ty, IsOneOrAllOnes, LHS, Shr);
+
+  // If dividing by a positive value, we're done. Otherwise, the result must
+  // be negated.
+  auto Neg = Builder.buildNeg(Ty, Shr);
+  auto IsNeg = Builder.buildICmp(CmpInst::Predicate::ICMP_SLT, CCVT, LHS, Zero);
----------------
jayfoad wrote:

```suggestion
  auto IsNeg = Builder.buildICmp(CmpInst::Predicate::ICMP_SLT, CCVT, RHS, Zero);
```

https://github.com/llvm/llvm-project/pull/83155