Reintroduce ext-based implementations for shift intrinsics
#543
Conversation
This file contains 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
Use a vector extract to implement _mm_s[l,r]li_si128 instead of storing re-loading from memory. This is very similar to the implementation prior to DLTcollab#484. Avoid compiler errors due to invalid immediates from macro expansion by guarding the immediate arguments with extra redundant ternary expressions (evaluated at compile time). This performs significantly better than the memory approach (a single vector instruction vs stores, loads & potential stack management). Example codegen for _mm_srli_si128(a, 11) with GCC 11.2.0 (-O3): Prior to commit: str q0, [sp] stp xzr, xzr, [sp, 16] ldur q0, [sp, 11] After this commit: movi v1.4s, 0x0 ext v0.16b, v0.16b, v1.16b, 11
Use a vector extract to perform _mm_alignr_epi8 instead of storing and loading from memory. Add additional ternary operators to safeguard against compiler errors at lower optimisation levels (prevents recurrence of DLTcollab#482). This performs significantly better than the memory approach. Example codegen for _mm_alignr_epi8(a, b, 11) with GCC 11.2.0 (-O3): Prior to commit: stp q1, q0, [sp] ldur q0, [sp, 11] After commit: ext v0.16b, v1.16b, v0.16b, 11
|
Thank @AymenQ for contributing! I appreciate the efforts from Arm. |
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Outdated suggestions cannot be applied.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.
Use the vector extract instruction (ext) to implement
_mm_srli_si128,_mm_slli_si128and_mm_alignr_epi8, significantly improving performanceby avoiding memory.
These were originally changed in #483 and #484 to instead use a store
followed by a shifted load. This was done in order to resolve issue
#482, where the compiler would throw an error if there were an invalid
immediate in the intrinsic arguments after the macros are expanded.
These commits avoid the recurrence of this issue by guarding the
immediate arguments with additional (redundant) ternary expressions.
These are evaluated at compile time and should cause no performance
loss; they only prevent the compilers from attempting to propagate
invalid immediates. This does now require that the immediate argument
to these intrinsics be a compile-time constant expression, but this
matches the requirement imposed by the original SSE intrinsics.
Sample codegen diffs are included in the individual commit messages.