You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
[X86] Use AVX512 BROADCAST from GPR to materialize uniform vector constants #239963
AVX512F provides VPBROADCASTD/Q and AVX512BW provides VPBROADCASTB/W - FP constants could be handled from bitcast integers. If we don't have VLX we'd need to always broadcast to ZMM.
We might be better off attempting other vector only patterns first, to avoid GPR->FPU traffic - much of it depends on how much register pressure we have.
Details
AVX512F provides VPBROADCASTD/Q and AVX512BW provides VPBROADCASTB/W - FP constants could be handled from bitcast integers. If we don't have VLX we'd need to always broadcast to ZMM.
We might be better off attempting other vector only patterns first, to avoid GPR->FPU traffic - much of it depends on how much register pressure we have.
Activity
llvmorg-github-actions commentedon Nov 7, 2026
@llvm/issue-subscribers-backend-x86
Author: Simon Pilgrim (RKSimon)
Details
AVX512F provides VPBROADCASTD/Q and AVX512BW provides VPBROADCASTB/W - FP constants could be handled from bitcast integers. If we don't have VLX we'd need to always broadcast to ZMM.We might be better off attempting other vector only patterns first, to avoid GPR->FPU traffic - much of it depends on how much register pressure we have.