LATX: optimize AVX and IR2 instruction generation - #438

Open

luzeng87 wants to merge 17 commits into

lat-opensource:masterfrom

luzeng87:gb603-avx-patterns

Commits on Sep 3, 2026

LATX, AVX: defer redundant VEX.128 upper-half clears
luzeng87
committed
5d2f2d7
LATX: keep mixed XMM/YMM TBs out of SHBR
luzeng87
committed
26cd7c1
LATX, AVX: restore YMM mask and blend writeback
luzeng87
committed
7c3667e
LATX, AVX: cover blend source-destination aliases
luzeng87
committed
70d46cf
LATX, AVX: avoid temporary YMM mask and blend copies
luzeng87
committed
90cac72
LATX: optimize AVX and scalar instruction sequences
luzeng87
committed
ff9a284
LATX, AVX: write packed conversions directly to the destination
luzeng87
committed
38e5f48
LATX, AVX: streamline integer narrowing and rounded multiply
luzeng87
committed
a095205
LATX, AVX: avoid redundant vector copies and lane inserts
luzeng87
committed
66ff1d3
LATX, AVX: write VPBLENDVB directly to the destination
luzeng87
committed
85efe63
LATX, AVX: use direct widening multiply-add sequences
luzeng87
committed
a55596d
LATX, AVX: write shuffle and byte-shift results directly
luzeng87
committed
8abd793
LATX: eliminate redundant LASX immediate loads
luzeng87
committed
18d469b
LATX: fold RIP-relative displacement into JIT loads
luzeng87
committed
2782c98
tests: cover consolidated LATX AVX optimizations
luzeng87
committed
5672cc0
LATX, AVX: defer YMM high-half clears per register
luzeng87
committed
2d7322c

Commits on Sep 4, 2026

LATX: fold integer and YMM reduction sequences
luzeng87
committed
b382b0f