Changed default value of slp-max-vf to 192

[d-smirnov]

This PR

1. Changed default value of slp-max-vf to 192
2. Minor performance fix: SmallSet -> SmallDenseSet

The PR fixes sharp compilation time increase noted in 527.cam4_r when patch https://reviews.llvm.org/D155689 applied.
Issue observed at LTO phase (link time increased from ~2 mins to ~62 mins) and caused by increased amount of lengthy instruction chains that SLP vectorizer tries to asses.

@vporpo

[llvmbot]

@llvm/pr-subscribers-llvm-transforms

Author: Dmitriy Smirnov (d-smirnov)

Changes

This PR

1. Changed default value of slp-max-vf to 192
2. Minor performance fix: SmallSet -> SmallDenseSet

The PR fixes sharp compilation time increase noted in 527.cam4_r when patch https://reviews.llvm.org/D155689 applied.
Issue observed at LTO phase and caused by increased amount of lengthy instruction chains that SLP vectorizer tries to asses.
@vporpo

---

Full diff: https://github.com/llvm/llvm-project/pull/70479.diff

1 Files Affected:

• (modified) llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp (+2-2)

diff --git a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
index bb4e743c1544a98..de7952b1839f5b3 100644
--- a/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
+++ b/llvm/lib/Transforms/Vectorize/SLPVectorizer.cpp
@@ -138,7 +138,7 @@ MaxVectorRegSizeOption("slp-max-reg-size", cl::init(128), cl::Hidden,
     cl::desc("Attempt to vectorize for this register size in bits"));
 
 static cl::opt<unsigned>
-MaxVFOption("slp-max-vf", cl::init(0), cl::Hidden,
+MaxVFOption("slp-max-vf", cl::init(192), cl::Hidden,
     cl::desc("Maximum SLP vectorization factor (0=unlimited)"));
 
 /// Limits the size of scheduling regions in a block.
@@ -4135,7 +4135,7 @@ static bool areTwoInsertFromSameBuildVector(
   // Go through the vector operand of insertelement instructions trying to find
   // either VU as the original vector for IE2 or V as the original vector for
   // IE1.
-  SmallSet<int, 8> ReusedIdx;
+  SmallDenseSet<int, 8> ReusedIdx;
   bool IsReusedIdx = false;
   do {
     if (IE2 == VU && !IE1)

[github-actions]

✅ With the latest revision this PR passed the C/C++ code formatter.

[alexey-bataev]

I don't like this change, need to take a closer look at the problematic case. This is a hacak that does not fix the issue, just hides it.

[kiranchandramohan]

Can the usage of SmallDenseSet instead of SmallSet (below) go in without further info or investigation?

Would you have suggestions on how to share a reproducer, since the issue manifests only at LTO on a large benchmark 528.cam4_r (spec2017)?

[d-smirnov]

@alexey-bataev, @vporpo Looks like I narrowed down "search space" a bit. The demonstrator IR is attached.
full.ll.gz
/usr/bin/time opt --passes="loop-vectorize,sroa,simplifycfg,slp-vectorizer" full.ll --disable-output --pass-remarks=slp-vectorize
Run takes about 30 minutes on Graviton 3 machine with unlimited chains and 13 sec with --slp-max-vf=192
Most of the time spent slp-vectorizing compute_uwshcu subroutine

[alexey-bataev]

I have a fix for this problem, hope to commit it later today after thorough testing. There are 2 problems, which can be easily fixed: 1. Better to use SmallBitVector instead of SmallSet. 2. Some particular trees, consisting only of phis and buildvector/gather nodes, can be dropped early as non-vectorizable in many cases.

Fixed opt results
time opt -passes=slp-vectorizer repro.ll -disable-output opt -passes=slp-vectorizer repro.ll 11.06s user 0.17s system 99% cpu 11.276 total

[alexey-bataev]

Fixed in d4cec1c