On 7/6/2026 8:14 AM, [email protected] wrote:
From: Kyrylo Tkachov <[email protected]>
A double-word count of trailing or leading zeros written as
lo ? ctz (lo) : 32 + ctz (hi)
with lo and hi the low and high halves of X equals a single wide ctzll (X),
and likewise for clz, but GCC kept the branch and two narrow counts.
Recognize the idiom in phiopt diamond handling and emit one .CTZ/.CLZ on the
double-word value. As __builtin_c[lt]z is undefined at zero, so is the
result, so a one-argument .CTZ/.CLZ is used.
Based on a prototype by Andrea, refactored into helpers and extended
to handle clz as well as ctz, with testcases added.
Bootstrapped and tested on aarch64-none-linux-gnu and x86_64-linux.
PR tree-optimization/126035
gcc/ChangeLog:
* tree-ssa-phiopt.cc (clzctz_match_arm): New function.
(clzctz_decode_half): New function.
(clzctz_combine_replacement): New function.
(pass_phiopt::execute): Call it for the diamond case.
gcc/testsuite/ChangeLog:
* gcc.target/aarch64/pr126035.c: New test.
* gcc.target/aarch64/pr126035-clz.c: New test.
* gcc.target/aarch64/pr126035-run.c: New test.
* gcc.target/aarch64/pr126035-debug.c: New test.
* gcc.target/aarch64/pr126035-extra-pred.c: New test.
* gcc.target/aarch64/pr126035-narrow.c: New test.
Co-authored-by: Andrew Pinski <[email protected]>
Signed-off-by: Andrew Pinski <[email protected]>
Signed-off-by: Kyrylo Tkachov <[email protected]>
So do we need review on this, or are you and Andrea iterating on a newer
version? (Andrea mentioned this problem in another context)?
jeff
