diff options
| author | Richard Sandiford <richard.sandiford@arm.com> | 2021-11-30 09:52:25 +0000 |
|---|---|---|
| committer | Richard Sandiford <richard.sandiford@arm.com> | 2021-11-30 09:52:25 +0000 |
| commit | e32b9eb32d7cd2d39bf9c70497890ac61b9ee14c (patch) | |
| tree | 3f4188ad4a2fc980f1b134c4f0137c3aba7947d9 /gcc/tree-vect-loop.c | |
| parent | 30213ae9a2eb53f6bc0913919457ceae2572b019 (diff) | |
| download | gcc-e32b9eb32d7cd2d39bf9c70497890ac61b9ee14c.zip gcc-e32b9eb32d7cd2d39bf9c70497890ac61b9ee14c.tar.gz gcc-e32b9eb32d7cd2d39bf9c70497890ac61b9ee14c.tar.bz2 | |
vect: Add support for fmax and fmin reductions
This patch adds support for reductions involving calls to fmax*()
and fmin*(), without the -ffast-math flags that allow them to be
converted to MAX_EXPR and MIN_EXPR.
gcc/
* doc/md.texi (reduc_fmin_scal_@var{m}): Document.
(reduc_fmax_scal_@var{m}): Likewise.
* optabs.def (reduc_fmax_scal_optab): New optab.
(reduc_fmin_scal_optab): Likewise
* internal-fn.def (REDUC_FMAX, REDUC_FMIN): New functions.
* tree-vect-loop.c (reduction_fn_for_scalar_code): Handle
CASE_CFN_FMAX and CASE_CFN_FMIN.
(neutral_op_for_reduction): Likewise.
(needs_fold_left_reduction_p): Likewise.
* config/aarch64/iterators.md (FMAXMINV): New iterator.
(fmaxmin): Handle UNSPEC_FMAXNMV and UNSPEC_FMINNMV.
* config/aarch64/aarch64-simd.md (reduc_<optab>_scal_<mode>): Fix
unspec mode.
(reduc_<fmaxmin>_scal_<mode>): New pattern.
* config/aarch64/aarch64-sve.md (reduc_<fmaxmin>_scal_<mode>):
Likewise.
gcc/testsuite/
* gcc.dg/vect/vect-fmax-1.c: New test.
* gcc.dg/vect/vect-fmax-2.c: Likewise.
* gcc.dg/vect/vect-fmax-3.c: Likewise.
* gcc.dg/vect/vect-fmin-1.c: New test.
* gcc.dg/vect/vect-fmin-2.c: Likewise.
* gcc.dg/vect/vect-fmin-3.c: Likewise.
* gcc.target/aarch64/fmaxnm_1.c: Likewise.
* gcc.target/aarch64/fmaxnm_2.c: Likewise.
* gcc.target/aarch64/fminnm_1.c: Likewise.
* gcc.target/aarch64/fminnm_2.c: Likewise.
* gcc.target/aarch64/sve/fmaxnm_2.c: Likewise.
* gcc.target/aarch64/sve/fmaxnm_3.c: Likewise.
* gcc.target/aarch64/sve/fminnm_2.c: Likewise.
* gcc.target/aarch64/sve/fminnm_3.c: Likewise.
Diffstat (limited to 'gcc/tree-vect-loop.c')
| -rw-r--r-- | gcc/tree-vect-loop.c | 45 |
1 files changed, 38 insertions, 7 deletions
diff --git a/gcc/tree-vect-loop.c b/gcc/tree-vect-loop.c index b1198e1..841da78 100644 --- a/gcc/tree-vect-loop.c +++ b/gcc/tree-vect-loop.c @@ -3185,9 +3185,22 @@ reduction_fn_for_scalar_code (code_helper code, internal_fn *reduc_fn) return true; default: - break; - } - return false; + return false; + } + else + switch (combined_fn (code)) + { + CASE_CFN_FMAX: + *reduc_fn = IFN_REDUC_FMAX; + return true; + + CASE_CFN_FMIN: + *reduc_fn = IFN_REDUC_FMIN; + return true; + + default: + return false; + } } /* If there is a neutral value X such that a reduction would not be affected @@ -3223,9 +3236,18 @@ neutral_op_for_reduction (tree scalar_type, code_helper code, return initial_value; default: - break; + return NULL_TREE; + } + else + switch (combined_fn (code)) + { + CASE_CFN_FMIN: + CASE_CFN_FMAX: + return initial_value; + + default: + return NULL_TREE; } - return NULL_TREE; } /* Error reporting helper for vect_is_simple_reduction below. GIMPLE statement @@ -3255,9 +3277,18 @@ needs_fold_left_reduction_p (tree type, code_helper code) return false; default: - break; + return !flag_associative_math; + } + else + switch (combined_fn (code)) + { + CASE_CFN_FMIN: + CASE_CFN_FMAX: + return false; + + default: + return !flag_associative_math; } - return !flag_associative_math; } if (INTEGRAL_TYPE_P (type)) |