https://gcc.gnu.org/bugzilla/show_bug.cgi?id=69935
Bug ID: 69935
Summary: load not hoisted out of linked-list traversal loop
Product: gcc
Version: 5.3.0
Status: UNCONFIRMED
Keywords: missed-optimization
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: peter at cordes dot ca
Target Milestone: ---
(please check the component. I guessed tree-optimization since it's
cross-architecture.)
gcc doesn't hoist the p->a load out of the loop in this linked-list function
int traverse_loadloop(struct foo_head *ph)
{
int a = -1;
struct foo *p = ph->h;
while (p) {
a = p->a;
p = p->n;
}
return a;
}
I checked on godbolt with gcc 4.8 on ARM/PPC/ARM64, and gcc 4.5.3 for AVR.
For x86, gcc 5.3.0 -O3 on godbolt (http://goo.gl/r8vb5L) does this:
movq (%rdi), %rdx
movl $-1, %eax
testq %rdx, %rdx
je .L10
.L11:
movl 8(%rdx), %eax ; load p->a inside the loop, not hoisted
movq (%rdx), %rdx
testq %rdx, %rdx
jne .L11
.L10:
rep ret
This is nice and compact, but less hyperthreading-friendly than it could be.
(The mov reg,reg alternative doesn't even take an execution unit on recent
CPUs).
The load of p->a every time through the loop might also delay the p->n load by
a cycle on CPUs with only one load port, or when there's a cache-bank conflict.
This might take the loop from one iteration per 4c to one per 5c (if L1
load-use latency is 4c).
Clang hoists the load out of the loop, producing identical asm output for this
function and one with the load hoisted in the C source. (The godbolt link has
both versions. Also see bug 69933 which I just reported, since gcc showed a
separate branch-layout issue for the source-level hoisting version.)
