Failure to remove redundant assume given alignment on call

[preames]

Bugzilla Link	25068
Version	trunk
OS	Linux
CC	@hfinkel

Extended Description

This assume in the following test case should be removable.

target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
target triple = "x86_64-unknown-linux-gnu"

declare align 8 i8* @get()

; Function Attrs: nounwind
declare void @llvm.assume(i1) #0

define void @test() {
  %p = tail call align 8 i8* @get()
  %ptrint = ptrtoint i8* %p to i64
  %maskedptr = and i64 %ptrint, 7
  %maskcond = icmp eq i64 %maskedptr, 0
  tail call void @llvm.assume(i1 %maskcond)
  ret void
}

I believe the problem is that ValueTracking hasn't been taught that the low bits of an aligned pointer result are known to be zero.

Note that this example is particularly problematic since the inliner will introduce @​llvm.assume calls with this pattern. I haven't caught it producing something quite this trivial, but I have seen cases where the IR simplifies to something this simple.

[hfinkel]

I believe the problem is that ValueTracking hasn't been taught that the low bits of an aligned pointer result are known to be zero.

We do already have this:

// Aligned pointers have trailing zeros - refine KnownZero set
if (V->getType()->isPointerTy()) {
unsigned Align = getAlignment(V, DL);
if (Align)
KnownZero |= APInt::getLowBitsSet(BitWidth, countTrailingZeros(Align));
}

Actually, does D12706 fix this too?

[preames]

Hal, see http://reviews.llvm.org/D13470. Essentially, getAlignment had a couple of missing cases.

[nikic]

The assume does get removed nowadays (https://llvm.godbolt.org/z/3ET575YxK), so closing this.