[llvm-bugs] [Bug 39235] New: [loop vectorize] failure to vectorize after GVN load-pre
via llvm-bugs
llvm-bugs at lists.llvm.org
Tue Oct 9 13:42:56 PDT 2018
https://bugs.llvm.org/show_bug.cgi?id=39235
Bug ID: 39235
Summary: [loop vectorize] failure to vectorize after GVN
load-pre
Product: libraries
Version: trunk
Hardware: PC
OS: All
Status: NEW
Severity: enhancement
Priority: P
Component: Loop Optimizer
Assignee: unassignedbugs at nondot.org
Reporter: mpfergu at gmail.com
CC: llvm-bugs at lists.llvm.org
Created attachment 20981
--> https://bugs.llvm.org/attachment.cgi?id=20981&action=edit
test showing the problem
I'm seeing an issue where adding !noalias metadata seems to prevents loop
vectorization from occurring. The issue seems to be that GVN does a PRE of a
load and this prevents the loop vectorizer from functioning.
See the attached .ll file. Supposing it's named opt-me.ll, to reproduce you can
do this:
opt -O3 -loop-vectorize -force-vector-width=4 -force-vector-interleave=1
opt-me.ll -S
Result:
remark: <unknown>:0:0: loop not vectorized: value that could not be identified
as reduction is used outside the loop
remark: <unknown>:0:0: loop not vectorized: value that could not be identified
as reduction is used outside the loop
-> loop not vectorized
Compare with
opt -O3 -loop-vectorize -force-vector-width=4 -force-vector-interleave=1
-enable-load-pre=false opt-me.ll -S
-> loop is vectorized, I see 4 x i32, extractelements, and no loop not
vectorized remarks.
If you remove the !noalias metadata from opt-me.ll, vectorization occurs
whether or not -enable-load-pre=false is passed. This seems unfortunate, as
front-ends would like to provide !noalias metadata where possible to improve
(and not hinder!) optimization.
I believe this problem is present in LLVM 6 as well as master.
--
You are receiving this mail because:
You are on the CC list for the bug.
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-bugs/attachments/20181009/5cce0cd2/attachment-0001.html>
More information about the llvm-bugs
mailing list