https://bugs.llvm.org/show_bug.cgi?id=38792
Bug ID: 38792
Summary: [InstCombine] Fails to combine three shufflevectors
produced by LoopVectorizer
Product: libraries
Version: trunk
Hardware: PC
OS: Linux
Status: NEW
Severity: enhancement
Priority: P
Component: Scalar Optimizations
Assignee: unassignedb...@nondot.org
Reporter: pauls...@linux.vnet.ibm.com
CC: llvm-bugs@lists.llvm.org
Created attachment 20815
--> https://bugs.llvm.org/attachment.cgi?id=20815&action=edit
reduced testcase
The LoopVectorizer has interleaved loads and stores. Basically, the loaded
elements should pairwise be reversed like
[0 1 2 3] -> [1 0 3 2]
The Vectorizer does not understand this but generates from two interleave
groups first a result for the load group, and then makes another shuffle for
the store group
%tmp6 = load <4 x i64>, <4 x i64>* %tmp5, align 8
%tmp7 = shufflevector <4 x i64> %tmp6, <4 x i64> undef, <2 x i32> <i32 0, i32
2>
%tmp8 = shufflevector <4 x i64> %tmp6, <4 x i64> undef, <2 x i32> <i32 1, i32
3>
%tmp9 = shufflevector <2 x i64> %tmp8, <2 x i64> %tmp7, <4 x i32> <i32 0, i32
1, i32 2, i32 3>
%tmp10 = shufflevector <4 x i64> %tmp9, <4 x i64> undef, <4 x i32> <i32 0,
i32 2, i32 1, i32 3>
store <4 x i64> %tmp10, <4 x i64>* undef, align 8
This results in [1 0 3 2], and I would have hoped that this would become a
single shufflevector after instcombine, but this does not happen.
There are comments in InstCombine that this is purposely done very
conservatively. It is however clear that this does not give good code on
SystemZ.
I wonder if anyone has any idea if InstCombiner should handle this case, or if
not, where should this be done. A custom DAGCombine by the target?
bin/opt ./tc_instcombine.ll -mtriple=systemz-unknown -mcpu=z13 -S -o
out.opt.ll -instcombine
--
You are receiving this mail because:
You are on the CC list for the bug._______________________________________________
llvm-bugs mailing list
llvm-bugs@lists.llvm.org
http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-bugs
