[llvm-bugs] [Bug 25808] New: [InstCombine] canonicalize extractelement reductions with shuffles
via llvm-bugs
llvm-bugs at lists.llvm.org
Fri Dec 11 09:03:44 PST 2015
https://llvm.org/bugs/show_bug.cgi?id=25808
Bug ID: 25808
Summary: [InstCombine] canonicalize extractelement reductions
with shuffles
Product: libraries
Version: trunk
Hardware: PC
OS: All
Status: NEW
Severity: normal
Priority: P
Component: Scalar Optimizations
Assignee: unassignedbugs at nondot.org
Reporter: spatel+llvm at rotateright.com
CC: llvm-bugs at lists.llvm.org
Classification: Unclassified
Hal suggested something like this in D15250:
http://reviews.llvm.org/D15250
$ cat reduction.c
typedef int v4i __attribute__((vector_size(16)));
int foo(v4i x) {
return x[0] + x[1] + x[2] + x[3];
}
$ ./clang -O1 reduction.c -S -o - -emit-llvm
...
define i32 @foo(<4 x i32> %x) #0 {
entry:
%vecext = extractelement <4 x i32> %x, i32 0
%vecext1 = extractelement <4 x i32> %x, i32 1
%add = add nsw i32 %vecext, %vecext1
%vecext2 = extractelement <4 x i32> %x, i32 2
%add3 = add nsw i32 %add, %vecext2
%vecext4 = extractelement <4 x i32> %x, i32 3
%add5 = add nsw i32 %add3, %vecext4
ret i32 %add5
}
InstCombine should be able to transform this into vector math + shuffles to
reduce the instruction count for any vector with 4+ elements:
%shuf1 = shufflevector <4 x i32> %x, <4 x i32> undef, <4 x i32> <i32 2, i32
3, i32 undef, i32 undef> ; shift left by 2 elements
%rdx1 = add <4 x i32> %x, %shuf1 ; 0+2 : 1+3 : undef : undef
%shuf2 = shufflevector <4 x i32> %rdx1, <4 x i32> undef, <4 x i32> <i32 1,
i32 undef, i32 undef, i32 undef> ; shift left 1 element
%rdx2 = add <4 x i32> %rdx1, %shuf2 ; 0+1+2+3 in first vec elt
%31 = extractelement <4 x i32> %rdx2, i32 0
This pattern would then be recognized as a canonical reduction.
I was looking for a similar transform to help bug 23116.
--
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/20151211/19554c31/attachment.html>
More information about the llvm-bugs
mailing list