con-kitty
Add a DynFlags plugin.
This only has an effect on GHC 8.10 and later. Older versions need to manually set the relevant flags.
TODO: warn when we had to override flags set by the user (ideally, we'd only
warn when they explicitly differ, not when we're overriding behavior set by
-O2 or something.
This also currently doesn't work, because some of our tests break when we don't ignore interface pragmas (I was hoping this would go away when we removed the presimplifier, but it didn't).
Fixes #43.
It is not entirely clear to me that this fixes the problem we had. We never set -funbox-strict-fields; it's the linear package that sets it.
Could you test this on Heavisoft? We re-implemented cross to avoid that problem, and you can remove that implementation and turn on this plugin and see what happens.
Also, any idea why some tests require ignoring interface pragmas? I thought interface pragmas should never be ignored in practice (otherwise the plugin wouldn't be able to inline anything).
It is not entirely clear to me that this fixes the problem we had. We never set
-funbox-strict-fields; it's thelinearpackage that sets it.Could you test this on Heavisoft? We re-implemented
crossto avoid that problem, and you can remove that implementation and turn on this plugin and see what happens.
Yeah, I think you're right, this is unlikely to help here. Shouldn't have added it. Was mostly trying to rescue various WIPs from Gerrit. And added that flag in vain hope.
Also, any idea why some tests require ignoring interface pragmas? I thought interface pragmas should never be ignored in practice (otherwise the plugin wouldn't be able to inline anything).
I haven't touched this change in like a year, so I don't recall. I know that initially I fixed some of the issues that were getting in the way of -no-ignore-interface-pragmas, but clearly we never got past them all.
The tests that fail are all ones where we expect to interpret the expression directly, so presumably the interface pragma is resulting in some premature inlining, specialization, or something.
That's weird. With the removal of presimplifier, now no simplifier runs before the plugin, so there shouldn't be any premature simplification. I'll look into it
On Sat, Apr 16, 2022, 3:06 PM Greg Pfeil @.***> wrote:
Also, any idea why some tests require ignoring interface pragmas? I thought interface pragmas should never be ignored in practice (otherwise the plugin wouldn't be able to inline anything).
I haven't touched this change in like a year, so I don't recall. I know that initially I fixed some of the issues that were getting in the way of -no-ignore-interface-pragmas, but clearly we never got past them all.
The tests that fail are all ones where we expect to interpret the expression directly, so presumably the interface pragma is resulting in some premature inlining, specialization, or something.
— Reply to this email directly, view it on GitHub https://github.com/con-kitty/categorifier/pull/56#issuecomment-1100762141, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABQMPCTP32UZHOTABTH4XJLVFM2X5ANCNFSM5TRN32TQ . You are receiving this because you commented.Message ID: @.***>