add option to print inlining costs #160
Conversation
This allows printing inlining costs alongside optimized IR similar to `Base.print_statement_costs`, which I found useful for debugging why some functions are not inlined. They are highlighted green, because I found them hard to distinguish otherwise, but I am open to suggestions regarding formatting. This relies on JuliaLang/julia#41257, which I hope can be backported to 1.7, so we don't have to worry about compatibility.
|
An example of what this looks like: |
Codecov Report
@@ Coverage Diff @@
## master #160 +/- ##
==========================================
- Coverage 44.61% 44.40% -0.21%
==========================================
Files 7 7
Lines 901 912 +11
==========================================
+ Hits 402 405 +3
- Misses 499 507 +8
Continue to review full report at Codecov.
|
This makes a minor changes to the README: - removes the part about accepting nominations (no time left...) - expands the installation documentation - proposes that we split the SnoopCompile part into "pre-Cthulhu" and "post-Cthulhu" This also adds a couple more packages (JET, SnoopCompile) and adds [compat] for long-term compatibility. There does not seem to be a suitable Cthulhu release; either we'll need to make a branch or we can get a release out before Saturday. I favor the latter, if possible. JuliaDebug/Cthulhu.jl#160 looks delicious and would be nice to merge if it wouldn't jeopardize the existing functionality.
* Update the README and add [compat] This makes a minor changes to the README: - removes the part about accepting nominations (no time left...) - expands the installation documentation - proposes that we split the SnoopCompile part into "pre-Cthulhu" and "post-Cthulhu" This also adds a couple more packages (JET, SnoopCompile) and adds [compat] for long-term compatibility. There does not seem to be a suitable Cthulhu release; either we'll need to make a branch or we can get a release out before Saturday. I favor the latter, if possible. JuliaDebug/Cthulhu.jl#160 looks delicious and would be nice to merge if it wouldn't jeopardize the existing functionality. * update Manifest to julia 1.7
| end | ||
| println(io) | ||
|
|
||
| if src isa IRCode && inline_cost |
There was a problem hiding this comment.
We can just allow the inlining cost printing for unoptimized source as well ?
| if src isa IRCode && inline_cost | |
| if inline_cost |
There was a problem hiding this comment.
I did actually experiment with that, but the issue is that any call that isn't an intrinsic is assumed to be a dynamic call, so it's probably more misleading than useful. I wonder whether we can get the recursive total costs after inlining for each call from the AbstractInterpreter, but that is probably a separate PR.
There was a problem hiding this comment.
Ah, I see.
For now maybe it'd be more kind if we print some warning message like:
elseif toggle === :optimize
optimize ⊻= true
if !is_cached(mi)
@warn "can't switch to post-optimization state, since this inference frame isn't cached"
optimize ⊻= true
end
continue
elseif toggle === :debuginfo
debuginfo = DebugInfo((Int(debuginfo) + 1) % 3)
elseif toggle === :inline_cost
if !optimize
@warn "enable optimization to see the inlining costs"
end
inline_cost ⊻= true
?
There was a problem hiding this comment.
Yeah, that sounds reasonable. Do you just want to push that to this branch?
There was a problem hiding this comment.
Okay, I will push the change and merge this.
This allows printing inlining costs alongside optimized IR similar to
Base.print_statement_costs, which I found useful for debugging whysome functions are not inlined. They are highlighted green, because I
found them hard to distinguish otherwise, but I am open to suggestions
regarding formatting. This relies on JuliaLang/julia#41257, which I hope
can be backported to 1.7, so we don't have to worry about compatibility.