SSE: incorrect code generated for i8 vector shuffles

[llvmbot]

Bugzilla Link	11184
Resolution	FIXED
Resolved on	Oct 20, 2011 08:39
Version	trunk
OS	All
Attachments	bitcode, test program
Reporter	LLVM Bugzilla Contributor

Extended Description

The attached test case generates an <8xi8> vector with values <1 2 3 4 5 6 7 8> using a shuffle of 2 <4xi8> vectors and stores it to memory. With top-of-tree, it seems that incorrect values are being stored; the attached test program should print out "1 2 3 4 5 6 7 8" but instead prints the last 4 values, repeated twice:

% llc -filetype=obj bug.ll -o bug.o && clang bug.cpp bug.o && ./a.out
5 6 7 8 5 6 7 8
%

Some archaeology in the checkins indicates that the enabling of element promotion type legalization broke this. If I check out the version immediately before that one, I get the expected result:

% llc -filetype=obj bug.ll -o bug.o && clang bug.cpp bug.o && ./a.out
1 2 3 4 5 6 7 8
%

Author: Nadav Rotem nadav.rotem@intel.com
Date: Sun Oct 16 20:31:33 2011 +0000

Enable element promotion type legalization by deafault.
Changed tests which assumed that vectors are legalized by widening them.



git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@142152 91177308-0d34-0410-b5e6-96231b3b80d8

[llvmbot]

Hi Matt,

I fixed a bug in the trunc-store implementation in r142434. Did you test this program after r142434 ?

Thanks,
Nadav

[llvmbot]

Yes. I saw this with r142537.

[llvmbot]

I am able to reproduce it on ToT. Hope to look at it later tonight.

[llvmbot]

Thanks! Let me know if there's anything I can do to help.

[llvmbot]

Fixed in 142579.