https://gcc.gnu.org/bugzilla/show_bug.cgi?id=126544
Bug ID: 126544
Summary: Quadraticness in SLP vectorisation
Product: gcc
Version: 17.0
Status: UNCONFIRMED
Keywords: compile-time-hog
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: ktkachov at gcc dot gnu.org
Target Milestone: ---
This needs a Python script to generate the offending form:
#!/usr/bin/env python3
"""Emit the basic block that makes BB SLP quadratic.
python3 04-slp-quadratic.py 3200 > t.c
gcc -O1 -ftree-slp-vectorize -fvect-cost-model=unlimited \
-march=armv8.2-a+sve -ftime-report -S -o /dev/null t.c
At n = 3200 the "tree slp vectorization" line reads 16.73 s and 1261 MB.
Doubling n multiplies both by four. Inserting an
__asm__ volatile ("" ::: "memory") every 64 stores splits the SLP region and
makes both perfectly linear, which is what identifies the region length as
the quadratic term.
"""
import sys
n = int(sys.argv[1]) if len(sys.argv) > 1 else 3200
print("int a[%d], b[%d], c[%d];" % (n + 8, n + 8, n + 8))
print("void f (void) {")
for i in range(n):
print(" a[%d] = b[%d] * c[%d] + %d;" % (i, i, i, i))
print("}")
Doing this on an aarch64 machine gives this line in -ftime-report:
tree slp vectorization : 14.57 ( 66%) 0.09 ( 47%) 14.66 ( 65%)
1883M ( 95%)
