Optimize miri engine performance for sanity checks of arrays

The knowledge about all the elements in an array can be used to run a single check instead of running the checks on each element one at a time (suggested in #53671 (comment)).

This is currently only done for strings, but could just as well be done for arrays of various types with Scalar layout.

The E-easy part of this issue is to check for arrays/slices of the builtin integer types and simply verify that there are no relocations in the entire array and that the entire array has no undefined bytes. Should be possible by simply calling

rust/src/librustc_mir/interpret/memory.rs

Line 740 in 1114ab6

pub fn read_bytes(&self, ptr: Scalar, size: Size) -> EvalResult<'tcx, &[u8]> {

and checking whether it returned Ok. No need to actually check the value.

The E-medium part is to refactor

rust/src/librustc_mir/interpret/validity.rs

Line 99 in 1114ab6

fn validate_scalar(

to take a bunch of closures which do the final checks. Instead of doing the check for a single memory location, run it for the entire array at once. This means the function won't have

rust/src/librustc_mir/interpret/validity.rs

Line 101 in 1114ab6

value: ScalarMaybeUndef,

as an argument anymore, but instead the closures pass around some generic value. E.g

rust/src/librustc_mir/interpret/validity.rs

Lines 110 to 113 in 1114ab6

    
           let value = match value { 
        
               ScalarMaybeUndef::Scalar(scalar) => scalar, 
        
               ScalarMaybeUndef::Undef => return validation_failure!("undefined bytes", path), 
        
           };

would be replaced by a closure call, and its result would be fed into the next closure e.g. in

rust/src/librustc_mir/interpret/validity.rs

Line 166 in 1114ab6

let in_range = |bound: RangeInclusive<u128>| bound.contains(&bits);

. Note that #53826 should be done before this, otherwise you'll go crazy while attempting to fiddle this refactoring into the current code

take a bunch of closures which do the final checks

I have no idea what you are suggesting here. The array field operand iteration is already quite optimized, so for which cases do you expect performance gains here? For non-integer scalars, we still have to actually check every single element for being valid.

Also we have again the question of when to look for the layout vs. the type.

E.g. for [&i32] it seems faster to do a single call to Memory::relocations than one for each field. I don't expect the iteration to be faster, I expect it to be faster to check N elements via an optimized operation instead of reading a Scalar at each index and then checking that.

For non-integer scalars, we still have to actually check every single element for being valid.

Even scalars with a valid range less than the full range require doing the access checks (defined, not pointer, aligned) on the memory, when one could just get the entire byte memory of the array at once and then just read integers out of it.

Even scalars with a valid range less than the full range require doing the access checks

Oh, I thought you'd only do this for "full-range scalars".

For anything else, the perf win will be smaller, and I am afraid we'll have to start tearing down those beautiful abstractions and make the code ugly. :/

Navigation Menu

Repository navigation

Optimize miri engine performance for sanity checks of arrays #53845

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	let value = match value {
	ScalarMaybeUndef::Scalar(scalar) => scalar,
	ScalarMaybeUndef::Undef => return validation_failure!("undefined bytes", path),
	};

Optimize miri engine performance for sanity checks of arrays #53845

Description

Activity

RalfJung commented on Aug 31, 2018

oli-obk commented on Aug 31, 2018

RalfJung commented on Aug 31, 2018

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions