Introduce prefetch-dynamic-strides option.
authorLuis Machado <luis.machado@linaro.org>
Mon, 7 May 2018 14:12:54 +0000 (14:12 +0000)
committerLuis Machado <luisgpm@gcc.gnu.org>
Mon, 7 May 2018 14:12:54 +0000 (14:12 +0000)
commit005e72b9ddb6b2d9d954b7a48c3d3c28972fff4d
treeaab4e836bba4385064c122da97ae201b12f14fcf
parent57e2d1175cf972b0a352be46903d87633304ce4e
Introduce prefetch-dynamic-strides option.

The following patch adds an option to control software prefetching of memory
references with non-constant/unknown strides.

Currently we prefetch these references if the pass thinks there is benefit to
doing so. But, since this is all based on heuristics, it's not always the case
that we end up with better performance.

For Falkor there is also the problem of conflicts with the hardware prefetcher,
so we need to be more conservative in terms of what we issue software prefetch
hints for.

This also aligns GCC with what LLVM does for Falkor.

Similarly to the previous patch, the defaults guarantee no change in behavior
for other targets and architectures.

2018-05-07  Luis Machado  <luis.machado@linaro.org>

gcc/
* config/aarch64/aarch64-protos.h (cpu_prefetch_tune)
<prefetch_dynamic_strides>: New const bool field.
* config/aarch64/aarch64.c (generic_prefetch_tune): Update to include
prefetch_dynamic_strides.
(exynosm1_prefetch_tune): Likewise.
(thunderxt88_prefetch_tune): Likewise.
(thunderx_prefetch_tune): Likewise.
(thunderx2t99_prefetch_tune): Likewise.
(qdf24xx_prefetch_tune): Likewise. Set prefetch_dynamic_strides to false.
(aarch64_override_options_internal): Update to set
PARAM_PREFETCH_DYNAMIC_STRIDES.
* doc/invoke.texi (prefetch-dynamic-strides): Document new option.
* params.def (PARAM_PREFETCH_DYNAMIC_STRIDES): New.
* params.h (PARAM_PREFETCH_DYNAMIC_STRIDES): Define.
* tree-ssa-loop-prefetch.c (should_issue_prefetch_p): Account for
prefetch-dynamic-strides setting.

From-SVN: r259996
gcc/ChangeLog
gcc/config/aarch64/aarch64-protos.h
gcc/config/aarch64/aarch64.c
gcc/doc/invoke.texi
gcc/params.def
gcc/params.h
gcc/tree-ssa-loop-prefetch.c