https://gcc.gnu.org/bugzilla/show_bug.cgi?id=120996
Bug ID: 120996
Summary: [16 regression][AArch64] 15% regression in microBUDE
since r16-1108-gb7960a3f966a0f
Product: gcc
Version: 16.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: dhruvc at nvidia dot com
Target Milestone: ---
Source code: https://github.com/UoB-HPC/microBUDE
Compiler flags: -std=c++17 -O3 main.cpp -fopenmp -mcpu=neoverse-v2
The commit (ranger: Add support for float <-> float casts [PR120231]) causes
early jump threading to combine the two blocks created for the ternary
operations on the following line of the benchmark:
float chrg_e = chrg_init * ((zone1 ? ONE : (ONE - distbb * elcdst1))
* (distbb < elcdst ? ONE : ZERO));
This leads to 183t.ifcvt/184t.vect generating some additional instructions that
increase the size of the loop and the runtime. I wasn't able to get a min-repro
using creduce, so I attached the repository itself instead.
