Rewrite the `Visitor` for `non_ssa_locals` #73854

New issue

Have a question about this project? # for a free GitHub account to open an issue and contact its maintainers and the community.

#

By clicking “#”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? # to your account

Jump to bottom

Closed

ecstatic-morse wants to merge 8 commits into rust-lang:master from ecstatic-morse:rewrite-local-analyzer

Contributor

ecstatic-morse commented Jun 28, 2020 •

edited

Loading

Resolves #72931. Seems like we're not going to do #73732 anytime soon, so might as well clean this up while it's fresh in my head. This tries to implement what I perceived as the intent of the old code, since there were several bugs which I mention in #72931. Also, we now only look for ZSTs in the final type of the mir::Place, not in all the types of its projections besides the base local. I couldn't see a good reason for the old behavior.

rust-highfive assigned varkor

Contributor

rust-highfive commented Jun 28, 2020

(rust_highfive has picked a reviewer for you, use r? to override)

rust-highfive added the S-waiting-on-review label

Contributor Author

ecstatic-morse commented Jun 28, 2020

rust-highfive assigned eddyb and unassigned varkor

eddyb reviewed

View reviewed changes

src/librustc_codegen_ssa/mir/analyze.rs Outdated Show resolved Hide resolved

eddyb reviewed

View reviewed changes

src/librustc_codegen_ssa/mir/analyze.rs Outdated Show resolved Hide resolved

eddyb reviewed

View reviewed changes

src/librustc_codegen_ssa/mir/analyze.rs Outdated Show resolved Hide resolved

eddyb reviewed

View reviewed changes

src/librustc_codegen_ssa/mir/analyze.rs Outdated Show resolved Hide resolved

eddyb added S-waiting-on-author and removed S-waiting-on-review labels

ecstatic-morse force-pushed the rewrite-local-analyzer branch from 35cd09f to ccd63ea Compare

July 19, 2020 19:27

Collaborator

bors commented Jul 21, 2020

☔ The latest upstream changes (presumably #69749) made this pull request unmergeable. Please resolve the merge conflicts.

ecstatic-morse force-pushed the rewrite-local-analyzer branch 5 times, most recently from 8f82918 to 86bf63a Compare

July 22, 2020 20:53

Contributor Author

ecstatic-morse commented Jul 25, 2020

I made the requested changes @eddyb, can you take another look?

JohnCSimon added S-waiting-on-review and removed S-waiting-on-author labels

eddyb reviewed

View reviewed changes

src/librustc_codegen_ssa/mir/analyze.rs Outdated

Comment on lines 167 to 168

    
                          // Projections like `(*x)[12]` are allowed but not `*(x[12])`, since a `Deref` of a

                          // local acts like a `Copy` of that local.

Member

eddyb Aug 13, 2020

This is a bit iffy in that it doesn't mention pointers at all, which is what is being "copied", and the arbitrary example of indexing had me confused for a while, although I see what it's saying now.

I would frankly just say that a Deref is a pointer/reference Copy and be done with it.

eddyb reviewed

View reviewed changes

src/librustc_codegen_ssa/mir/analyze.rs Outdated

    
                              continue;

                          }

                          // Ignoring `Deref`s, the only allowed projections are reads of scalar fields.

Member

eddyb Aug 13, 2020

s/allowed/SSA-capable (or similar phrasing) would help IMO.

eddyb reviewed

View reviewed changes

src/librustc_codegen_ssa/mir/analyze.rs

Comment on lines +33 to 35

    
                      if ty_requires_alloca(&analyzer.fx, layout) {

                          analyzer.not_ssa(local);

                      }

Member

eddyb Aug 13, 2020

Maybe we should do this first, and even keep the monomorphized TyAndLayouts around, to make it cheaper to check if "consume" field projections can be handled in an SSA way.

Member

eddyb Aug 13, 2020

Actually, huh, we have access to FunctionCx, I'm tempted to say we should create the FunctionCx with all of the local TyAndLayouts already computed (although we couldn't have placeholder LocalRefs easily I don't think?) - but maybe that's too much for this PR?

eddyb reviewed

View reviewed changes

src/librustc_codegen_ssa/mir/analyze.rs Outdated

Comment on lines 163 to 177

    
                      while let [ref proj_base @ .., elem] = *projection {

                          projection = proj_base;

                          self.super_projection_elem(local, proj_base, elem, context, location);

                          // Projections like `(*x)[12]` are allowed but not `*(x[12])`, since a `Deref` of a

                          // local acts like a `Copy` of that local.

                          if let mir::PlaceElem::Deref = elem {

                              has_disqualifying_projection = false;

                              context = PlaceContext::NonMutatingUse(NonMutatingUseContext::Copy);

                              continue;

                          }

                          // Ignoring `Deref`s, the only allowed projections are reads of scalar fields.

                          if is_consume(context) && matches!(elem, mir::ProjectionElem::Field(..)) {

                              let base_ty = mir::Place::ty_from(local, proj_base, self.fx.mir, self.fx.cx.tcx());

                              let base_ty = self.fx.monomorphize(&base_ty);

                              let span = self.fx.mir.local_decls[local].source_info.span;

                              let layout = self.fx.cx.spanned_layout_of(base_ty.ty, span);

                              if !ty_requires_alloca(self.fx, layout) {

                                  continue;

                              }

                          }

                          has_disqualifying_projection = true;

                      }

Member

eddyb Aug 13, 2020

Looking more at this loop, I would split it into several parts:

a loop to just call super_projection_elem uniformly on all the elems, to get it out of the way (it only serves to visit the index local of an Index projection, with a `Copy context)
find the first Deref, if there is one, e.g. with .position(|elem| matches!(elem, mir::PlaceElem::Deref))
loop over the "direct projections" (i.e. before any Deref, if any) and do self.not_ssa(local); break; at the first sign of trouble

I mostly want this for clarity, but that last part has some promise for making this more efficient, so here's a few more ideas:

context wouldn't change across it, meaning if is_consume(context) can be placed around the whole loop
there could be a quick check that all the elements are fields, to avoid time even looking at the layouts
we probably should check all the field projections (like today / this PR) but if we go forward, from the local to the innermost field, we should almost always hit the non-SSA layout first, making the rejection cheaper
we could call visit_local before the loop, and skip the loop if non_ssa_locals already contains the local (see also the comment at the top of this file, which would make this more useful)
if we keep the monomorphized TyAndLayout of the local (which we always have to compute anyway) around, we could start from it and just go layout = layout.field(self.fx, i) at every step instead of computing the whole type and monomorphizing every time

Contributor Author

ecstatic-morse Aug 22, 2020 •

edited

Loading

I tried implementing some of these ideas in the latest commit. It's more efficient, but seems more complex to me. If you have a clear idea for how this code should look, you should implement it on top of or in place of this PR. I don't think I'm going to keep working on this.

eddyb reviewed

View reviewed changes

src/librustc_codegen_ssa/mir/analyze.rs Outdated

    
                              let span = self.fx.mir.local_decls[local].source_info.span;

                              let layout = self.fx.cx.spanned_layout_of(base_ty.ty, span);

                              if !ty_requires_alloca(self.fx, layout) {

Member

eddyb Aug 13, 2020

Hmm, I feel like this doesn't capture the important aspect here: only for some layouts we can extract their fields as SSA values. So maybe instead of !ty_requires_alloca it could be layout_allows_ssa (i.e. with flipped meaning).

Contributor Author

ecstatic-morse Aug 22, 2020 •

edited

Loading

I wouldn't have done the renaming in 45fd741, since being assigned exactly once before any read (SSA-like) seems to be only one of several requirements for a value to be forced to memory. I've been using the term "requires alloca" for locals that fail this check, but it feels like there should be a good term for locals that do not require allocas that I'm not seeing.

eddyb added S-waiting-on-author and removed S-waiting-on-review labels


          Use helper method for what types need allocas

f0ed021

ecstatic-morse added 6 commits

August 22, 2020 11:34


          Use Option<Location> to indicate the first assignment

3e74727

This is the same size as `Location`. Using an invalid location is
premature optimization.


          Refactor LocalAnalyzer to use visit_projection_elem

e0fdf7c


          Use loop instead of visitor to check for invalid projections

da03a98


          first_assignment no longer needs a getter

7fcd4cb


          Ignore VarDebugInfo when forcing locals on the stack

a4d0bed


          Remove outdated comment

f74e65c

ecstatic-morse force-pushed the rewrite-local-analyzer branch 2 times, most recently from c0fb44c to 99badcf Compare

August 22, 2020 19:09


          Rewrite projection loop using notes from review

277d4ca

ecstatic-morse force-pushed the rewrite-local-analyzer branch from 99badcf to 277d4ca Compare

August 22, 2020 20:24

Collaborator

bors commented Aug 30, 2020

☔ The latest upstream changes (presumably #74862) made this pull request unmergeable. Please resolve the merge conflicts.

crlf0710 added S-waiting-on-author and removed S-waiting-on-author labels

Dylan-DPC-zz commented Sep 18, 2020

@ecstatic-morse if you can resolve the conflicts, we can get this merged

Dylan-DPC-zz commented Oct 6, 2020

Closing this due to inactivity. Thanks

Dylan-DPC-zz closed this

Dylan-DPC-zz added S-inactive and removed S-waiting-on-author labels

# for free to join this conversation on GitHub. Already have an account? # to comment

Labels