On 5/20/2026 4:15 AM, Michiel Derhaeg wrote:
Hi Jin,
Thanks for the review.
Know that I'm not a RISC-V port reviewer. I can't approve your patches.
I'm just providing opinions, you're free to disagree with me :p
Perhaps true. But reviewing patches is a good path having those kind of
privileges, especially when paired with consistent contributions. Even
if you don't want to move into that kind of position with the project,
feedback from other developers who care about the quality of the code is
always appreciated.
Since the core logic remains largely unchanged (aside from these fixes), I
initially believed the existing test suite would be sufficient.
If you found bugs, they're clearly insufficient. That's a bit the problem,
most of these fusion checks are lacking tests.
Testsuite coverage for fusion is weak at best. It's doubly so because a
bug in this space doesn't typically result in a functional failure, but
just substandard performance.
That's part of what got Daniel and I looking at some of these problems a
while back. While we made significant progress and ultimately were able
to get fusion to improve performance, particularly after Artimey's work,
I'm far from convinced we're doing a good job.
One of the tests we never were able to set up and run was to run spec
under perf record with the right raw counters that exposed fusion
events. That ultimately gives us a histogram of where fusion was
occurring and what specific kinds of fusions. We then needed to
correlate that to the compiler's view of fusion. I don't remember how
we were going to map the compiler's view of fusion against the perf
output. Regardless, that's was the general direction we were heading.
Jeff
