Repository navigation
[Tests][AArch64] Make SVE codegen assertions robust across LLVM versions - #19752
Merged
Merged
Conversation
tests/python/codegen/test_target_codegen_aarch64.py cross-compiles AArch64 SVE kernels and regex-matches the generated assembly for specific instruction forms and counts. Several of those assertions encode the exact code shape produced by the LLVM versions used in CI (15-17). On a TVM built against LLVM 20 the tests fail, even though the emitted IR is correct (+sve target-features and vscale_range(1,16) are present) -- the difference is entirely inside LLVM's loop vectorizer / cost model, not in TVM's codegen. This is the same class of issue as apache#19744 (which gated the +v9a vscale_range expectation); that fix landed separately, and this change covers the remaining test_muladd / test_eq / test_neq failures. LLVM 20 behavior changes observed: - Floating-point mul-add: LLVM 20 prefers a fixed-width NEON main loop (fmla v0.4s, ...) over a scalable SVE loop (mla z0.s, ...) for float / float16 fmuladd on generic AArch64 +sve, so the SVE mla/mad count drops to 0. - Equality / inequality compares: LLVM <= 17 interleaves the scalable loop by two (2 SVE compares), while LLVM 20 emits a fixed-width NEON main loop with a single predicated SVE epilogue (1 SVE compare). Changes (test-only, no codegen change): - test_muladd: for LLVM >= 18 on float/float16, also accept the NEON ldr/ldp q* + fmla/fmls v*.Ns form in addition to the SVE form (a union, so whichever form a given LLVM emits keeps the assertion meaningful). The NEON regex requires a vector register (v0.4s), so a scalar tail fmadd s0 cannot satisfy it. - test_muladd: group the mad|mla alternation as (?:mad|mla)\t... . Without the group, a bare mad could match anywhere in the assembly (e.g. inside a scalar fmadd), inflating the count. - test_eq / test_neq: relax the compare-count assertion from > 1 to > 0, since the interleave factor is an LLVM cost-model detail, not a codegen guarantee. CI's LLVM 15-17 still produces >= 1 SVE compare. The assertions still require real AArch64 vector code -- they are not weakened to vacuous checks, and the strict SVE path is unchanged for the LLVM versions used in CI.
Contributor
|
Warning You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again! |
tqchen
approved these changes
Jun 13, 2026
MasterJH5574
pushed a commit
to MasterJH5574/tvm
that referenced
this pull request
Jun 15, 2026
…ons (apache#19752) `tests/python/codegen/test_target_codegen_aarch64.py` cross-compiles AArch64 SVE kernels and regex-matches the generated assembly for specific instruction forms and counts. Several of those assertions encode the exact code shape produced by the LLVM versions used in CI (15-17). On a TVM built against LLVM 20 the tests fail, even though the emitted IR is correct (+sve target-features and vscale_range(1,16) are present) -- the difference is entirely inside LLVM's loop vectorizer / cost model, not in TVM's codegen. (cherry picked from commit 383c446)
MasterJH5574
pushed a commit
to MasterJH5574/tvm
that referenced
this pull request
Jun 15, 2026
…ons (apache#19752) `tests/python/codegen/test_target_codegen_aarch64.py` cross-compiles AArch64 SVE kernels and regex-matches the generated assembly for specific instruction forms and counts. Several of those assertions encode the exact code shape produced by the LLVM versions used in CI (15-17). On a TVM built against LLVM 20 the tests fail, even though the emitted IR is correct (+sve target-features and vscale_range(1,16) are present) -- the difference is entirely inside LLVM's loop vectorizer / cost model, not in TVM's codegen. (cherry picked from commit 383c446)
tqchen
pushed a commit
that referenced
this pull request
Oct 16, 2026
…ons (#19752) `tests/python/codegen/test_target_codegen_aarch64.py` cross-compiles AArch64 SVE kernels and regex-matches the generated assembly for specific instruction forms and counts. Several of those assertions encode the exact code shape produced by the LLVM versions used in CI (15-17). On a TVM built against LLVM 20 the tests fail, even though the emitted IR is correct (+sve target-features and vscale_range(1,16) are present) -- the difference is entirely inside LLVM's loop vectorizer / cost model, not in TVM's codegen.
tqchen
pushed a commit
that referenced
this pull request
Oct 18, 2026
…ons (#19752) `tests/python/codegen/test_target_codegen_aarch64.py` cross-compiles AArch64 SVE kernels and regex-matches the generated assembly for specific instruction forms and counts. Several of those assertions encode the exact code shape produced by the LLVM versions used in CI (15-17). On a TVM built against LLVM 20 the tests fail, even though the emitted IR is correct (+sve target-features and vscale_range(1,16) are present) -- the difference is entirely inside LLVM's loop vectorizer / cost model, not in TVM's codegen.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
tests/python/codegen/test_target_codegen_aarch64.pycross-compiles AArch64 SVE kernels and regex-matches the generated assembly for specific instruction forms and counts. Several of those assertions encode the exact code shape produced by the LLVM versions used in CI (15-17). On a TVM built against LLVM 20 the tests fail, even though the emitted IR is correct (+sve target-features and vscale_range(1,16) are present) -- the difference is entirely inside LLVM's loop vectorizer / cost model, not in TVM's codegen.