Vigneshwar Jayakumar 0d67b2edac

[AMDGPU] Add f32 cost model for exp/exp2/exp10 intrinsics (#185369) ...

exp2 maps to a single v_exp_f32 (quarter rate), while exp and exp10 need
additional full-rate costs for the base conversion. Also added cost based on 
non-afn and denorm mode.

This enables SLP vectorization of exp2 on targets with packed f32 ops.

2026-03-10 11:08:39 -05:00

..

add_sub_sat-inseltpoison.ll

…

add_sub_sat.ll

…

address-space-ptr-sze-gep-index-assert.ll

…

bswap-inseltpoison.ll

…

bswap.ll

…

crash_extract_subvector_cost.ll

…

exp2-vectorize.ll

[AMDGPU] Add f32 cost model for exp/exp2/exp10 intrinsics (#185369)

2026-03-10 11:08:39 -05:00

external-shuffle.ll

…

extract-ordering.ll

…

horizontal-store.ll

…

invariant-load-no-alias-store.ll

…

lit.local.cfg

…

min_max.ll

…

packed-math.ll

…

phi-result-use-order.ll

…

reduction.ll

AMDGPU: Improve getShuffleCost accuracy for 8- and 16-bit shuffles (#168818)

2025-11-21 19:33:13 +00:00

round-inseltpoison.ll

…

round.ll

…

slp-v2f16.ll

AMDGPU: Improve getShuffleCost accuracy for 8- and 16-bit shuffles (#168818)

2025-11-21 19:33:13 +00:00

slp-v2f32.ll

…

vectorize-i8.ll

…

zext-duplicate-shift.ll

[SLP] Reject duplicate shift amounts in matchesShlZExt reorder path (#183627)

2026-02-27 13:00:58 -06:00