https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92130
Bug ID: 92130
Summary: Missed vectorization for iteration dependent loads and
simple multiplicative accumulators
Product: gcc
Version: 9.2.1
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: witold.baryluk+gcc at gmail dot com
Target Milestone: ---
Created attachment 47051
--> https://gcc.gnu.org/bugzilla/attachment.cgi?id=47051&action=edit
Perlin2D noise mesh generation
So,
I do have pretty complex multi level loop spread across many functions, but it
can be all vectorized, but under certain scenarios gcc does not vectorize it
with gcc 9.2.1
I am attaching somehow simplified code with few defines inside to play with it.
The one exposed by default present the biggest challenge to gcc, despite me
able to vectorize it manually.
I tested this on SSE2, AVX2 (cascadelake and znver2), AVX512 (-march=knm and
-march=skylake-avx512) and ARM SVE, with all same effects. I am using
associative math and other flags mentioned in the sourcefile at the top.
The high level overview is like this:
input: A, F, W, maxO, sufficiently aligned d.
foreach y:
foreach x:
float v = 0.0
float a = 1.0
float f = 1.0
foreach o in [0, maxO):
v += a * g(f * x, f * y, o, h(o, p))
a *= A
f *= F
d[y*W + x] = v
where both g and h are pure functions (relatively complex tho) with no control
flow or data dependent flow.
In some situations if a and f are replaced by a precomputed table of
coefficient for every o, and then used as v += a[o] * g(f[o] * x, f[o] * y,
h(o, p)), it does vectorize, but not always. h(o, p) could also be precomputed,
but I didn't bother as it appears to not have any bad effect on vectorizer.
Vectorizater should vectorize along the 'foreach x', and compute multiple x-s
per-lane completely independently. It is true that when updating a and f, each
lane need to be duplicated, but that can be done by computing it scalarly, and
then broadcasting, or by repeating same constants updates in each lane.
