temp work-around for 32-bit inference failure (#29923)

[StefanKarpinski]

No description provided.

[KristofferC]

@test_broken maybe.

Won't we just merge this and then forget about it?

[KlausC]

Shouldn't if Sys.WORD_SIZE == 64 || S === Int128 be if Sys.WORD_SIZE == 32 && S == Int128 ?

[StefanKarpinski]

Yes, it should. I had the logic inverted previously and only half fixed it!

[StefanKarpinski]

@test_broken maybe.

I'm not aware of a good way to do that for @inferred which is what's failing.

Won't we just merge this and then forget about it?

Possibly but it's been a week with all 32-bit CI failing, which is not acceptable. I've pinged the entire compiler team on Slack but gotten no response. @JeffBezanson, @vtjnash, @Keno, can someone at least comment if they're looking at #29923 or not?

[KristofferC]

Possibly but it's been a week with all 32-bit CI failing, which is not acceptable.

Agreed, this is probably for the best to not waste everyones time looking at CI logs.

[JeffBezanson]

I agree we should temporarily disable this. We don't know how to fix this. It has come up before and we've never fully figured it out, so there's no estimate of how long it will take to fix.

[martinholters]

Alternatively, we keep the test and do

--- a/base/int.jl
+++ b/base/int.jl
@@ -775,7 +775,7 @@ if Core.sizeof(Int) == 4
         return Int128(rem(BigInt(x), BigInt(y)))
     end
     function rem(x::UInt128, y::UInt128)
-        return UInt128(rem(BigInt(x), BigInt(y)))
+        return UInt128(rem(BigInt(x), BigInt(y)))::UInt128
     end
 
     function mod(x::Int128, y::Int128)

Here, I'd like to point out that (on 32bit systems), we have two definitions in int.jl which for operations on UInt128 go via BigInt:

julia/base/int.jl

Lines 770 to 772 in e209c3d

	function div(x::UInt128, y::UInt128)
	return UInt128(div(BigInt(x), BigInt(y)))::UInt128
	end

julia/base/int.jl

Lines 777 to 779 in e209c3d

	function rem(x::UInt128, y::UInt128)
	return UInt128(rem(BigInt(x), BigInt(y)))
	end

So by adding the type assertion to the second one, too, we then have covered "all" of them. Therefore, one may hope for these test failures not to show up again. It is still more of a work-around than a fix, though.

[martinholters]

Even though I'm experimenting with an alternative over in #30050, I'm in favour of merging this ASAP to get CI back up and then investigate further.

[fredrikekre]

Unrelated (?) Profile test failure on AV 32-bit.

[StefanKarpinski]

Restarting AV given that the whole point of this PR is to fix CI, so I'd like to see it actually pass.

[musm]

all green 👍