Failure to optimize popcount of a single bit #47952

GabrielRavier · 2020-12-27T23:56:05Z

Extended Description

int f(unsigned int a)
{
return __builtin_popcount(a >> (CHAR_BIT * sizeof(a) - 1));
}

This can be optimized to return (a >> (__CHAR_BIT__ * sizeof(a) - 1));. This transformation is done by GCC, but not by LLVM.

The text was updated successfully, but these errors were encountered:

GabrielRavier · 2020-12-27T23:56:05Z

rotateright · 2020-12-28T16:59:24Z

We miss both of these:

define i32 @pop_signbit(i32 %x) {
%b = lshr i32 %x, 31
%r = tail call i32 @llvm.ctpop.i32(i32 %b), !range !0
ret i32 %r
}

define i32 @pop_lowbit(i32 %x) {
%b = and i32 %x, 1
%r = tail call i32 @llvm.ctpop.i32(i32 %b), !range !0
ret i32 %r
}

We might be able to rely on the range metadata. I'll take a look.

rotateright · 2020-12-28T21:38:37Z

Metadata doesn't provide what we need. I used a ValueTracking call instead.
Should be fixed with:
https://reviews.llvm.org/rG236c4524a7cd

It looks like gcc still does more, so please open a new bug if more popcount optimizations are expected/needed.

llvmbot transferred this issue from llvm/llvm-bugzilla-archive Dec 11, 2021

This issue was closed.