[llvm] [InstCombine] Try the flipped strictness of predicate in `foldICmpShlConstant` (PR #92773)

[via llvm-commits]

================
@@ -2414,14 +2414,37 @@ Instruction *InstCombinerImpl::foldICmpShlConstant(ICmpInst &Cmp,
   // free on the target. It has the additional benefit of comparing to a
   // smaller constant that may be more target-friendly.
   unsigned Amt = ShiftAmt->getLimitedValue(TypeBits - 1);
-  if (Shl->hasOneUse() && Amt != 0 && C.countr_zero() >= Amt &&
-      DL.isLegalInteger(TypeBits - Amt)) {
-    Type *TruncTy = IntegerType::get(Cmp.getContext(), TypeBits - Amt);
-    if (auto *ShVTy = dyn_cast<VectorType>(ShType))
-      TruncTy = VectorType::get(TruncTy, ShVTy->getElementCount());
-    Constant *NewC =
-        ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
-    return new ICmpInst(Pred, Builder.CreateTrunc(X, TruncTy), NewC);
+  if (Shl->hasOneUse() && Amt != 0 && DL.isLegalInteger(TypeBits - Amt)) {
+    auto FoldICmpShlToICmpTrunc = [&](ICmpInst::Predicate Pred,
+                                      const APInt &C) -> Instruction * {
+      if (C.countr_zero() < Amt)
+        return nullptr;
+      Type *TruncTy = ShType->getWithNewBitWidth(TypeBits - Amt);
+      Constant *NewC =
+          ConstantInt::get(TruncTy, C.ashr(*ShiftAmt).trunc(TypeBits - Amt));
+      return new ICmpInst(
+          Pred, Builder.CreateTrunc(X, TruncTy, "", Shl->hasNoSignedWrap()),
----------------
goldsteinn wrote:

> > If you don't have `nsw` and have `nuw`, you can just use `lshr` and propagate `nuw`? If you have both you have to drop either `nuw` or `nsw`. https://alive2.llvm.org/ce/z/9tjTDb
> 
> Yeah, I don't know which one is better when we have both flags :( 

I guess knowledge about 1 extra bit vs more knowledge about 1 less bit. For my money id stick w/ `nuw`.

> BTW, as the motivation cases (see [dtcxzyw/llvm-opt-benchmark#616](https://github.com/dtcxzyw/llvm-opt-benchmark/pull/616)) have no flags on shl, I think it is ok to just keep it simple.

IMO its not a big increase in complexity, and since we regularly have missed-optimizations related to throwing away info, think its better to just handle right from the start.

https://github.com/llvm/llvm-project/pull/92773