https://gcc.gnu.org/bugzilla/show_bug.cgi?id=82533
Bug ID: 82533
Summary: inefficient code generation for copy loop on falkor
Product: gcc
Version: 8.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: target
Assignee: unassigned at gcc dot gnu.org
Reporter: wilson at gcc dot gnu.org
Target Milestone: ---
Target: aarch64
Created attachment 42348
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=42348&action=edit
testcase, to reproduce compile with -mcpu=falkor -O2 -ftree-vectorize
When lmbench stream copy is compiled with -O2 -ftree-vectorize -mcpu=falkor,
the inner loop gets compiled to
.L4:
ldr q0, [x2, x3]
str q0, [x1, x3]
add x3, x3, 16
cmp x3, x4
bne .L4
The str qX [reg+reg] instruction is very inefficient on Falkor. We get a 16%
performance increase if we disable use of str qX [r+r], to get instead
.L4:
ldr q0, [x2, x3]
add x5, x1, x3
add x3, x3, 16
cmp x3, x4
str q0, [x5]
bne .L4
A proposed patch was posted to gcc-patches here
https://gcc.gnu.org/ml/gcc-patches/2017-09/msg01547.html
