[llvm-bugs] [Bug 31286] New: Instcombine(?) should factor broadcasts across math
via llvm-bugs
llvm-bugs at lists.llvm.org
Mon Dec 5 17:31:05 PST 2016
https://llvm.org/bugs/show_bug.cgi?id=31286
Bug ID: 31286
Summary: Instcombine(?) should factor broadcasts across math
Product: libraries
Version: trunk
Hardware: PC
OS: Linux
Status: NEW
Severity: normal
Priority: P
Component: Scalar Optimizations
Assignee: unassignedbugs at nondot.org
Reporter: mkuper at google.com
CC: llvm-bugs at lists.llvm.org
Classification: Unclassified
Not actually sure "factor" is the right term here, what I mean is that we're
missing:
bcast(x) + bcast(y) -> bcast(x + y)
Right now, for:
#include <immintrin.h>
__m128 bar(float f, float g, float h, float i) {
__m128 v0 = _mm_set_ps1(f);
__m128 v1 = _mm_set_ps1(g);
__m128 v2 = _mm_set_ps1(h);
__m128 v3 = _mm_set_ps1(i);
return v0 + v1 + v2 + v3;
}
We get:
bar(float, float, float, float): # @bar(float,
float, float, float)
shufps $0, %xmm0, %xmm0 # xmm0 = xmm0[0,0,0,0]
shufps $0, %xmm1, %xmm1 # xmm1 = xmm1[0,0,0,0]
shufps $0, %xmm2, %xmm2 # xmm2 = xmm2[0,0,0,0]
shufps $0, %xmm3, %xmm3 # xmm3 = xmm3[0,0,0,0]
addps %xmm1, %xmm0
addps %xmm2, %xmm0
addps %xmm3, %xmm0
retq
Instead of:
bar(float, float, float, float):
addss %xmm1, %xmm0
addss %xmm2, %xmm0
addss %xmm3, %xmm0
shufps $0, %xmm0, %xmm0
ret
There's another odd thing here - in IR we represent each broadcast as a
sequence of insertelements, instead of an insert + shuffle, which, I believe,
is the canonical form (at least, this is what the vectorizer produces)
That should probably be fixed first, so we have only one canonical broadcast to
match.
--
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/20161206/b0c6170c/attachment-0001.html>
More information about the llvm-bugs
mailing list