https://gcc.gnu.org/bugzilla/show_bug.cgi?id=128790
Bug ID: 128790
Summary: reassoc splits each multiply-add chain of a vectorised
loop in two without regard to how many chains are live
Product: gcc
Version: 16.2.0
Status: UNCONFIRMED
Severity: normal
Priority: P3
Component: tree-optimization
Assignee: unassigned at gcc dot gnu.org
Reporter: freddie at witherden dot org
Target Milestone: ---
Consider the following (I can't believe it's not APL) C snippet:
#define F(M, k) M(k, 0) M(k, 1) M(k, 2) M(k, 3) M(k, 4)
#define L(k, i) double a##k##i = s[r][k*5 + i][j];
#define S(k, i) s[r][k*5 + i][j] = a##k##i;
#define W(k, c) double w##k = e*wg[(k*8 + c)*8 + j];
#define T(k, c, i) a##k##i += w##k*u[(c*5 + i)*8 + j];
#define R(k, c) T(k, c, 0) T(k, c, 1) T(k, c, 2) T(k, c, 3) T(k, c, 4)
#define C(c) { double e = op[(c*8 + r)*8 + j]; \
W(0, c) W(1, c) W(2, c) R(0, c) R(1, c) R(2, c) }
void tile(const double *restrict op, const double *restrict wg,
const double *restrict u, double (*restrict s)[15][8])
{
for (int r = 0; r < 8; r++)
#pragma omp simd
for (int j = 0; j < 8; j++)
{
F(L, 0) F(L, 1) F(L, 2)
C(0) C(1) C(2) C(3) C(4) C(5) C(6) C(7)
F(S, 0) F(S, 1) F(S, 2)
}
}
Compiling this gcc 16.2.0 at -O3 -ffast-math -fopenmp-simd
-mcpu=apple-m1 gives an inner loop with ~400 instructions with a lot of them
being stack spills. This is due to the reassoc path cracking the FMA chains in
two, which creates register pressure. Passing -fno-tree-reassoc clears the
issue up and gives an inner loop with ~250 instructions, no stack spills, and
~2x better performance on my M4.
Thus, it seems as if the heuristics for reassoc need tweaking to ensure that it
does not result in too many accumulators.