Repository navigation
[meta][x86] Vector Constant Materialization #239959
Description
This was raised at the 2026 devmtg x86 roundtable.
X86 still doesn't make any attempt to materialize any vector constant patterns apart from all-zero and all-ones, we have no mechanism to determine the cost of materialization (vs a folded load) or when they can be hoisted (LICM etc.) into spare registers (vs cheap rematerialization).
@preames has proposed he raise an initial pathfinder patch that we can then add additional patterns to for specific ISAs etc.
Possible patterns - see Agner Fog's "Generating Constants" chapter in "Optimizing subroutines in assembly language".
Activity
llvmorg-github-actions commentedon Nov 7, 2026
@llvm/issue-subscribers-backend-x86
Author: Simon Pilgrim (RKSimon)
Details
This was raised at the 2026 devmtg x86 roundtable.X86 still doesn't make any attempt to materialize any vector constant patterns apart from all-zero and all-ones, we have no mechanism to determine the cost of materialization or when they can be hoisted (LICM etc.) into spare registers (vs cheap rematerialization).
@preames has proposed he raise an initial pathfinder patch that we can then add additional patterns to for specific ISAs etc.