interpreter: fix bug with ssair phi-node handling #29262
Conversation
|
@Keno This is failing the 32-bit build. Can you tell me how to fix that? |
vtjnash
force-pushed
the
jn/interpret-phi
branch
from
2aaf7c5 to
4d504aa
Compare
|
For adding tests, I plan to run the core.jl tests with the interpreter (takes a couple minutes). There's just one test that needs to be fixed. |
|
I would rather have an example of a specific function that fails. If an existing test in core.jl is sufficient that's fine, we should just copy it into a new test that runs it with --compile=min. |
vtjnash
force-pushed
the
jn/interpret-phi
branch
from
4d504aa to
a33675d
Compare
|
I know the approximate forms of IR that would fail, but I think it's also useful to show that our interpreter is an accurate representation of our compiler. |
|
Yes, I agree it's useful to run some subset of the tests both in the interpreter and compiler, but we also need an example that's known to trigger this problem. Otherwise e.g. moving tests out of core.jl or other test changes could inadvertently stop testing for this. |
vtjnash
force-pushed
the
jn/interpret-phi
branch
from
a33675d to
ed13540
Compare
somehow these were missed in the last round of `nospecialize` annotations in this file, but these are also very important for compiler performance (especially as we have now been adding more places where we inspect tuples with this method)
it is not valid to treat these like statements, since they are actually arguments to the basic block it turns out to be a bit hard to deal with this particular representation of phi nodes in an interpreter, since the information is spread out across multiple statements and could pop up at any time, but with some care, it is possible. we just need to be careful to process phi nodes as a block, and taking care to look for implicit edges. fix #29342
vtjnash
force-pushed
the
jn/interpret-phi
branch
from
ed13540 to
39f2561
Compare
| @@ -38,7 +38,10 @@ _setindex(v, i::Integer) = () | |||
|
|
|||
| ## iterating ## | |||
|
|
|||
| iterate(t::Tuple, i::Int=1) = 1 <= i <= length(t) ? (@inbounds t[i], i+1) : nothing | |||
| function iterate(@nospecialize(t::Tuple), i::Int=1) | |||
KristofferC
added
the
needs nanosoldier run
same as the bug Keno fixed in https://github.com/JuliaLang/julia/pull/28978/files#diff-6d4d21428a67320600faf5a1a9f3a16a