[InstCombine] Extra uses of truncated value prevent icmp_eq(trunc(x),c) -> icmp_eq(x,c) fold

[RKSimon]

define i1 @src(i8 %a0) {
  %tz = tail call range(i8 0, 9) i8 @llvm.cttz.i8(i8 %a0, i1 false)
  %tr = trunc i8 %tz to i4
  %c = icmp eq i4 %tr, 8
  call void @use(i4 %tr)
  ret i1 %c
}
define i1 @tgt(i8 %a0) {
  %tz = tail call range(i8 0, 9) i8 @llvm.cttz.i8(i8 %a0, i1 false)
  %c = icmp eq i8 %a0, 0
  %tr = trunc i8 %tz to i4
  call void @use(i4 %tr)
  ret i1 %c
}
declare void @use(i4)

[kper]

Stupid question, but what's the difference here between src and tgt? I just noticed that the %tr was sunk below the the icmp

[RKSimon]

we're not performing the comparison on the i8 %tz result directly, which prevents a later fold from converting the comparison into icmp eq i8 %a0 ,0

[dtcxzyw]

We can try to remove the one-use constraint if the masking instruction is unnecessary (with trunc nuw).

llvm-project/llvm/lib/Transforms/InstCombine/InstCombineCompares.cpp

Lines 1484 to 1506 in cf1cdde

	if (Cmp.isEquality() && Trunc->hasOneUse()) {
	// Canonicalize to a mask and wider compare if the wide type is suitable:
	// (trunc X to i8) == C --> (X & 0xff) == (zext C)
	if (!SrcTy->isVectorTy() && shouldChangeType(DstBits, SrcBits)) {
	Constant *Mask =
	ConstantInt::get(SrcTy, APInt::getLowBitsSet(SrcBits, DstBits));
	Value *And = Builder.CreateAnd(X, Mask);
	Constant *WideC = ConstantInt::get(SrcTy, C.zext(SrcBits));
	return new ICmpInst(Pred, And, WideC);
	}
	// Simplify icmp eq (trunc x to i8), 42 -> icmp eq x, 42|highbits if all
	// of the high bits truncated out of x are known.
	KnownBits Known = computeKnownBits(X, &Cmp);
	// If all the high bits are known, we can do this xform.
	if ((Known.Zero | Known.One).countl_one() >= SrcBits - DstBits) {
	// Pull in the high bits from known-ones set.
	APInt NewRHS = C.zext(SrcBits);
	NewRHS |= Known.One & APInt::getHighBitsSet(SrcBits, SrcBits - DstBits);
	return new ICmpInst(Pred, X, ConstantInt::get(SrcTy, NewRHS));
	}
	}