You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Removes ValueIds from SSA to reduce memory usage and hopefully improve compiler performance as well.
ValueIds are replaced by the Value enum itself, which is now thinner and implements Copy.
Several functions for iterating over values (block_parameters, instruction_results) now no longer require references to the DFG, so we don't need as much cloning in several passes.
Related to the previous point, values can be created much more often without referencing the DFG at all now, e.g. Value::constant just needs a field and a type since we don't need the separate map to keep each valueid unique to each constant value any more. Similarly, even Value::Instruction only needs an instruction id and a position. No need to store instruction id values in an external map in the dfg. This means we can get rid of quite a few fields on the DFG to save memory.
DFG::set_value_from_id can no longer mutate the value a ValueId points to, and instead relies solely on the internal map between values.
Additional Context
The integration tests are currently all passing, but there is a stack overflow somewhere within the test runner when running the unit tests, in between running the actual tests.
I have not started migrating all the SSA parser tests in the evaluator, these are expected to be failing currently since the display of valueids has changed.
Documentation*
Check one:
No documentation needed.
Documentation included in this PR.
[For Experimental Features] Documentation to be submitted in a separate PR.
PR Checklist*
I have tested the changes locally.
I have formatted the changes with Prettier and/or cargo fmt on default settings.
Initial results are very ungood, I think this is because FieldElement is 256bits causing Value to be 320 bits.
I'll re-add constant ids which should bring Value down to 96 bits. If we could embed the enum tag in the value somehow we could get it to 64.
Value's size has been reduced, although I'm unsure why the changes aren't more dramatic. I'd think removing the instruction_results map alone would have been a decent improvement. We're also cloning Vecs less throughout. I think I'll have to profile this more before fixing the ssa parser but as is this may not be worth merging with the regression in ram_blowup_regression.
After some more perf testing, I'm going to close this PR. Testing it against master on rollup-base-private, this PR is compiling in ~90s compared to master's ~60s. Unsure if this is from the same regression as conditional_1 but this PR has been enough of a rabbit hole already that I'd rather close it.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
jfecher
force-pushed
the
jf/no-valueids
branch
from
Description
Problem*
Summary*
Removes
ValueIds from SSA to reduce memory usage and hopefully improve compiler performance as well.ValueIds are replaced by theValueenum itself, which is now thinner and implementsCopy.block_parameters,instruction_results) now no longer require references to the DFG, so we don't need as much cloning in several passes.Value::constantjust needs a field and a type since we don't need the separate map to keep each valueid unique to each constant value any more. Similarly, evenValue::Instructiononly needs an instruction id and a position. No need to store instruction id values in an external map in the dfg. This means we can get rid of quite a few fields on the DFG to save memory.DFG::set_value_from_idcan no longer mutate the value a ValueId points to, and instead relies solely on the internal map between values.Additional Context
The integration tests are currently all passing, but there is a stack overflow somewhere within the test runner when running the unit tests, in between running the actual tests.
I have not started migrating all the SSA parser tests in the evaluator, these are expected to be failing currently since the display of valueids has changed.
Documentation*
Check one:
PR Checklist*
cargo fmton default settings.