[clang] [llvm] Match bitsin(typeof(x)) - popcnt(x) to s_bcnt0_i32 on AMDGPU (PR #164847)

[Patrick Simmons via cfe-commits]

================
@@ -1977,6 +1989,37 @@ Value *AMDGPUCodeGenPrepareImpl::applyFractPat(IRBuilder<> &Builder,
   return insertValues(Builder, FractArg->getType(), ResultVals);
 }
 
+bool AMDGPUCodeGenPrepareImpl::visitCtpop(IntrinsicInst &I) {
+  uint32_t BitWidth, DestinationWidth, IntrinsicWidth;
+  if (!I.hasOneUse() || !I.getType()->isIntegerTy() ||
+      !ST.hasBCNT(BitWidth = I.getType()->getIntegerBitWidth()))
+    return false;
+
+  BinaryOperator *MustBeSub = dyn_cast<BinaryOperator>(I.user_back());
+  if (!MustBeSub || MustBeSub->getOpcode() != BinaryOperator::Sub)
+    return false;
+
+  ConstantInt *FirstOperand = dyn_cast<ConstantInt>(MustBeSub->getOperand(0));
+  if (!FirstOperand || FirstOperand->getZExtValue() != BitWidth)
+    return false;
+
+  IRBuilder<> Builder(MustBeSub);
+  Instruction *TransformedIns =
+      Builder.CreateIntrinsic(BitWidth > 32 ? Intrinsic::amdgcn_bcnt64_lo
+                                            : Intrinsic::amdgcn_bcnt32_lo,
+                              {}, {I.getArgOperand(0)});
+
+  if ((DestinationWidth = MustBeSub->getType()->getIntegerBitWidth()) !=
+      (IntrinsicWidth = TransformedIns->getType()->getIntegerBitWidth()))
+    TransformedIns = cast<Instruction>(Builder.CreateZExtOrTrunc(
----------------
linuxrocks123 wrote:

That's true, but it should always be an Instruction, so I like the cast and the internal assertion it provides.

https://github.com/llvm/llvm-project/pull/164847