Hi,
On 01/23/2018 07:40 AM, Kyrill Tkachov wrote:
Hi Luis,
On 22/01/18 13:46, Luis Machado wrote:
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.
I've regression-tested and bootstrapped it on aarch64-linux. No
problems found.
Ok?
This also looks like a sensible approach to me with a caveat inline.
The same general comments as for patch [1/2] apply.
diff --git a/gcc/config/aarch64/aarch64-protos.h
b/gcc/config/aarch64/aarch64-protos.h
index 8736bd9..22bd9ae 100644
--- a/gcc/config/aarch64/aarch64-protos.h
+++ b/gcc/config/aarch64/aarch64-protos.h
@@ -230,6 +230,9 @@ struct cpu_prefetch_tune
const int l1_cache_size;
const int l1_cache_line_size;
const int l2_cache_size;
+ /* Whether software prefetch hints should be issued for non-constant
+ strides. */
+ const unsigned int prefetch_dynamic_strides;
I understand that the midend PARAMs are defined as integers, but I think
the backend tuning option here is better represented as a boolean as it
really
is just a yes/no decision.
I started off with a boolean to be honest. Then i noticed the midend
only used integers, which i restricted to the range of 0..1.
I'll change this locally to use booleans again.
Thanks!
Luis