On 8/4/2026 4:01 AM, [email protected] wrote:
From: Kyrylo Tkachov <[email protected]>
Two reductions feeding the matching scalar operation are one reduction of
the elementwise operation:
REDUC (a) op REDUC (b) -> REDUC (a op b)
The two sides accumulate the same multiset of lanes, just grouped
differently, so for the associative and commutative reductions the result
is identical. This trades a lane-crossing reduction, which is the
expensive part, for an elementwise operation.
MIN, MAX, AND, IOR and XOR need no flag. The sum needs one only for
floating point, where the regrouping is a reassociation; the integer sum
is done in the unsigned type, since the reduction wraps while the signed
vector type has undefined overflow.
#include <arm_neon.h>
int f (int32x4_t a, int32x4_t b) { return vaddvq_s32 (a) + vaddvq_s32 (b); }
aarch64 -O3 before:
addv s1, v1.4s
addv s0, v0.4s
fmov w1, s1
fmov w0, s0
add w0, w0, w1
after:
add v0.4s, v0.4s, v1.4s
addv s0, v0.4s
fmov w0, s0
A pair of maxima goes from 8 instructions to 3. On a target without a
horizontal reduce instruction the reduction is open coded as a chain of
shuffles, so the saving is larger there, never smaller.
For integers, preserve trapping and sanitized scalar overflow.
For floating point, require reassociation, insignificant zero signs, and
non-trapping exceptions.
Restrict the reassociative alternative to floating-point types.
Floating-point options must not license fixed-point saturation changes.
Bootstrapped and tested on aarch64-none-linux-gnu.
Ok for trunk?
Thanks,
Kyrill
gcc/ChangeLog:
* match.pd (REDUC (a) op REDUC (b)): New simplifications combining
two reductions into one.
gcc/testsuite/ChangeLog:
* gcc.dg/tree-ssa/vec-reduc-pair-1.c: New test.
* gcc.dg/tree-ssa/vec-reduc-pair-2.c: New test.
* gcc.dg/tree-ssa/vec-reduc-pair-3.c: New test.
* gcc.dg/tree-ssa/vec-reduc-pair-4.c: New test.
Signed-off-by: Kyrylo Tkachov <[email protected]>
Nice. Given the often multi-cycle nature of reductions, particularly on
vectors with many elements, this seems particularly helpful.
OK
jeff
