[WIP][X86] lowerBuildVectorAsBroadcast - don't convert constant vectors to broadcasts on AVX512VL targets #73509
+92,997
−96,561
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
RKSimon
requested review from
phoebewang,
KanRobert,
goldsteinn and
yubingex007-a11y
|
✅ With the latest revision this PR passed the C/C++ code formatter. |
phoebewang
reviewed
Nov 27, 2023
phoebewang
reviewed
Nov 27, 2023
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
8 times, most recently
from
21fad09
to
dae6506
Compare
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
5 times, most recently
from
fc410b2
to
c5db884
Compare
|
Added basic handling for non-VLX AVX512 targets when dealing with 512-bit constant vectors |
goldsteinn
reviewed
Dec 7, 2023
goldsteinn
reviewed
Dec 7, 2023
| @@ -948,7 +948,8 @@ define void @load_i8_stride4_vf32(ptr %in.vec, ptr %out.vec0, ptr %out.vec1, ptr | |||
| ; AVX512F-NEXT: vpshufb %ymm0, %ymm1, %ymm2 | |||
| ; AVX512F-NEXT: vmovdqa 64(%rdi), %ymm3 | |||
| ; AVX512F-NEXT: vpshufb %ymm0, %ymm3, %ymm0 | |||
| ; AVX512F-NEXT: vmovdqa {{.*#+}} ymm4 = [0,4,0,4,0,4,8,12] | |||
| ; AVX512F-NEXT: vbroadcasti128 {{.*#+}} ymm4 = [0,4,8,12,0,4,8,12] | |||
There was a problem hiding this comment.
Yes - still addressing regressions, that's why its still a draft :)
|
Can't memory ops microfuse on all targets? Why is this avx512vl only? |
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
from
c5db884
to
4fedfe0
Compare
RKSimon
added a commit
that referenced
this pull request
Dec 8, 2023
We were using VPTERNLOGQ for everything but i32 types, which made broadcasts wider than necessary Noticed in #73509
I don't understand what you're asking - we already load-fold for all targets. This patch is about improving broadcast-load-fold. By prematurely converting to constant broadcasts in DAG we're hindering later optimizations - MachineLICM is a good example (we end up hoisting the broadcast which then often spills the full width broadcasted vector......). By keeping to full vector width until X86FixupVectorConstants we avoid a lot of this. I will eventually be disabling constant broadcasting in lowerBuildVectorAsBroadcast for all AVX targets later but theres a lot of regressions to still deal with - AVX512VL (and AVX512F for 512-bit vectors) is the first step. |
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
2 times, most recently
from
adc89f0
to
3af0810
Compare
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
7 times, most recently
from
0c80ea8
to
6b2809b
Compare
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
from
6b2809b
to
5eff513
Compare
RKSimon
added a commit
that referenced
this pull request
Jan 3, 2024
…ndi(z,w,c1)) to AVX512BW mask select Yet another yak shaving regression fix for #73509
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
from
5eff513
to
f738150
Compare
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
from
f738150
to
6d1519e
Compare
RKSimon
added a commit
that referenced
this pull request
Feb 5, 2024
Handle masked predicated load/broadcasts in addConstantComments now that we can generically handle the destination + mask register This will more significantly help improve 'fixup constant' comments from #73509
RKSimon
added a commit
that referenced
this pull request
Feb 5, 2024
Handle masked predicated movss/movsd in addConstantComments now that we can generically handle the destination + mask register This will more significantly help improve 'fixup constant' comments from #73509
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
2 times, most recently
from
86ed907
to
925a8d0
Compare
agozillon
pushed a commit
to agozillon/llvm-project
that referenced
this pull request
Feb 5, 2024
Handle masked predicated load/broadcasts in addConstantComments now that we can generically handle the destination + mask register This will more significantly help improve 'fixup constant' comments from llvm#73509
agozillon
pushed a commit
to agozillon/llvm-project
that referenced
this pull request
Feb 5, 2024
Handle masked predicated movss/movsd in addConstantComments now that we can generically handle the destination + mask register This will more significantly help improve 'fixup constant' comments from llvm#73509
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
from
925a8d0
to
77629d5
Compare
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
from
77629d5
to
e8d60f1
Compare
… targets On AVX512 targets we're better off keeping constant vector at full width to ensure that they can be load folded into vector instructions, reducing register pressure. If a vector constant remains as a basic load, X86FixupVectorConstantsPass will still convert this to a broadcast instruction for us. Non-VLX targets are still seeing some regressions due to these being implicitly widened to 512-bit ops in isel patterns and not in the DAG, so I've limited this to just 512-bit vectors for now.
RKSimon
force-pushed
the
perf/broadcast-avx512
branch
from
e8d60f1
to
27c0a8a
Compare
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.
On AVX512VL targets we're better off keeping constant vectors at full width to ensure that they can be load folded into vector instructions, reducing register pressure. If a vector constant remains as a basic load, X86FixupVectorConstantsPass will still convert this to a broadcast instruction for us.
This is still a WIP patch - as can be seen by the changes to X86InstrFoldTables.cpp, we have very poor coverage for the BroadcastFoldTables (Issue #66360). I don't know whether just to continue manually extending these tables or to wait for #66360 to be done.
Non-VLX AVX512 targets are still seeing some regressions due to main instructions being implicitly widened to 512-bit ops in isel patterns and not in the DAG, so for now lets keep them as it is (same for AVX1/AVX2 targets). For AVX1/AVX2, broadcasting constants via lowerBuildVectorAsBroadcast helps a lot, as long as we don't cause register spills, which is major problem on larger vectorized hot loops. I'm currently thinking we should add a x86 pass, similar to MachineLICM, that unfolds broadcastable constant loads as long as we have spare registers; we could then remove the remaining lowerBuildVectorAsBroadcast constant handling entirely - any thoughts?
My goal is to improve AVX1/AVX2 vector constant handling but getting AVX512 out of the way appears to be an easier first step.