On 12/10/25 5:35 AM, Konstantinos Eleftheriou wrote:
On Tue, Dec 9, 2025 at 9:38 PM Andrew Pinski
<[email protected] <mailto:[email protected]>>
wrote:
On Tue, Dec 9, 2025 at 8:55 AM Konstantinos Eleftheriou
<[email protected]
<mailto:[email protected]>> wrote:
>
> From: kelefth <[email protected]
<mailto:[email protected]>>
>
> This patch enables the avoid-store-forwarding patch by default at
O2 or
> higher on AArch64.
I am not 100% convinced why this should be enabled specific to
aarch64. It seems like a generic issue. and looking into the testcases
all should happen on the gimple level much earlier than RTL.
We just don't have enough data to support that this would be
beneficial on other architectures. There have been some talks about
introducing a cost-function so that the targets can decide (see
threads of previous versions), but there was no clear decision, so
we went with this for now.
Right. There is an expectation that it may help other architectures and
I've got a TODO to test it on our RISC-V design -- enough with our
memory dependence predictor it may not help much in practice.
As for gimple vs RTL. It seems to me that the transformation will
ultimately be driven by cost modeling and (IMHO) would introduce more
target dependent decisions into the gimple pipeline than we likely want.
jeff
