https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109240
Bug ID: 109240 Summary: Missed fneg/fsub optimization Product: gcc Version: 13.0 Status: UNCONFIRMED Keywords: wrong-code Severity: normal Priority: P3 Component: tree-optimization Assignee: unassigned at gcc dot gnu.org Reporter: jakub at gcc dot gnu.org CC: avieira at gcc dot gnu.org, burnus at gcc dot gnu.org, jakub at gcc dot gnu.org, marxin at gcc dot gnu.org, tnfchris at gcc dot gnu.org Depends on: 109230 Target Milestone: --- Host: aarch64-linux-gnu Target: aarch64-linux-gnu +++ This bug was initially created as a clone of Bug #109230 +++ On aarch64 we optimize at -O2 only half of the following routines: typedef float V __attribute__((vector_size (4 * sizeof (float)))); typedef int VI __attribute__((vector_size (4 * sizeof (float)))); __attribute__((noipa)) V foo (V x, V y) { V a = x - y; V b = y + x; return __builtin_shuffle (b, a, (VI) { 0, 5, 2, 7 }); } __attribute__((noipa)) V bar (V x, V y) { V a = x - y; V b = y + x; return __builtin_shuffle (a, b, (VI) { 4, 1, 6, 3 }); } __attribute__((noipa)) V baz (V x, V y) { V a = x - y; V b = y + x; return __builtin_shuffle (b, a, (VI) { 4, 1, 6, 3 }); } __attribute__((noipa)) V qux (V x, V y) { V a = x - y; V b = y + x; return __builtin_shuffle (a, b, (VI) { 0, 5, 2, 7 }); } __attribute__((noipa)) V boo (V x, V y) { V a = x + y; V b = y - x; return __builtin_shuffle (b, a, (VI) { 0, 5, 2, 7 }); } __attribute__((noipa)) V corge (V x, V y) { V a = x + y; V b = y - x; return __builtin_shuffle (a, b, (VI) { 4, 1, 6, 3 }); } __attribute__((noipa)) V fred (V x, V y) { V a = x + y; V b = y - x; return __builtin_shuffle (b, a, (VI) { 4, 1, 6, 3 }); } __attribute__((noipa)) V garply (V x, V y) { V a = x + y; V b = y - x; return __builtin_shuffle (a, b, (VI) { 0, 5, 2, 7 }); } starting with r13-4024-gb2bb611d90d01f64a24 (plus r13-4122-g1bc7efa948f751 bugfix). The other half could be handled similarly, just with fneg+fsub rather than fneg+fadd. Unfortunately, match.pd canonicalizes those, we still have 0, 5, 2, 7 permutations for all of them, but the two operations swapped. Unfortunately match.pd doesn't allow :c on vec_perm, and if we use (for op (plus minus) otherop (minus plus) then we couldn't add :c to the plus one. So, copy and paste the whole large simplification, swap (plus:c @0 @1) and (minus @0 @1) and replace (plus at the end with (minus? Or handle the commutativity manually? (for op (plus minus) otherop (minus plus) (simplify (vec_perm (op @0 @1) (otherop @2 @3) VECTOR_CST@4) and use operand_equal_p manually to allow all forms we want? Referenced Bugs: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=109230 [Bug 109230] [13 Regression] Maybe wrong code for opus package on aarch64 since r13-4122-g1bc7efa948f751