https://gcc.gnu.org/g:5a5ad4c11dbdfacd8d236134e37e2d5812f52147
commit r17-5950-g5a5ad4c11dbdfacd8d236134e37e2d5812f52147 Author: Onalan, T <[email protected]> Date: Tue Nov 3 17:22:15 2026 +0000 arm: Add expander for `negv2di2` with Neon Although Neon has a VSUB.i64 insn, it lacks VNEG for this mode. This means we can scalarize negation unnecessarily. gcc/ChangeLog: * config/arm/neon.md (negv2di2): New expander. gcc/testsuite/ChangeLog: * gcc.target/arm/neon-neg64.c: New test. * gcc.target/arm/neon-neg64-autovec.c: Likewise. Assisted-by: GPT-6.1 Sol Diff: --- gcc/config/arm/neon.md | 13 +++++++++ gcc/testsuite/gcc.target/arm/neon-neg64-autovec.c | 15 ++++++++++ gcc/testsuite/gcc.target/arm/neon-neg64.c | 34 +++++++++++++++++++++++ 3 files changed, 62 insertions(+) diff --git a/gcc/config/arm/neon.md b/gcc/config/arm/neon.md index d8f5c56ea6bd..ffda37babaec 100644 --- a/gcc/config/arm/neon.md +++ b/gcc/config/arm/neon.md @@ -812,6 +812,19 @@ (const_string "neon_abs<q>")))] ) +;; Not handled by "<absneg_str><mode>2" due to the lack of a direct +;; 64-bit vneg. +(define_expand "negv2di2" + [(set (match_operand:V2DI 0 "s_register_operand") + (neg:V2DI (match_operand:V2DI 1 "s_register_operand")))] + "TARGET_NEON" + { + rtx zero = force_reg (V2DImode, CONST0_RTX (V2DImode)); + emit_insn (gen_subv2di3 (operands[0], zero, operands[1])); + DONE; + } +) + (define_insn "neon_neg<mode>2" [(set (match_operand:VDQW 0 "s_register_operand" "=w") (neg:VDQW (match_operand:VDQW 1 "s_register_operand" "w")))] diff --git a/gcc/testsuite/gcc.target/arm/neon-neg64-autovec.c b/gcc/testsuite/gcc.target/arm/neon-neg64-autovec.c new file mode 100644 index 000000000000..171cf42a3f57 --- /dev/null +++ b/gcc/testsuite/gcc.target/arm/neon-neg64-autovec.c @@ -0,0 +1,15 @@ +/* { dg-do compile } */ +/* { dg-require-effective-target arm_neon_ok } */ +/* { dg-options "-O3 -fvect-cost-model=unlimited -fdump-tree-vect-details" } */ +/* { dg-add-options arm_neon } */ + +void +negate_array (unsigned long long *__restrict output, + const unsigned long long *__restrict input, unsigned int count) +{ + for (unsigned int i = 0; i < count; ++i) + output[i] = -input[i]; +} + +/* { dg-final { scan-assembler {vsub\.i64} } } */ +/* { dg-final { scan-tree-dump "vectorized 1 loops" "vect" } } */ diff --git a/gcc/testsuite/gcc.target/arm/neon-neg64.c b/gcc/testsuite/gcc.target/arm/neon-neg64.c new file mode 100644 index 000000000000..be707a8acefa --- /dev/null +++ b/gcc/testsuite/gcc.target/arm/neon-neg64.c @@ -0,0 +1,34 @@ +/* { dg-do assemble } */ +/* { dg-require-effective-target arm_neon_ok } */ +/* { dg-options "-O2 -save-temps -fno-ipa-icf" } */ +/* { dg-add-options arm_neon } */ + +#include <arm_neon.h> + +int64x2_t +negate_intrinsic (int64x2_t value) +{ + return vsubq_s64 (vdupq_n_s64 (0), value); +} + +int64x2_t +negate_signed (int64x2_t value) +{ + return -value; +} + +uint64x2_t +negate_unsigned (uint64x2_t value) +{ + return -value; +} + +int64x2_t +negate_memset (int64x2_t value) +{ + int64x2_t zero; + __builtin_memset (&zero, 0, sizeof (zero)); + return vsubq_s64 (zero, value); +} + +/* { dg-final { scan-assembler-times {vsub\.i64} 4 } } */
