[naga spv-out] Spill arrays and matrices for runtime indexing.

[jimblandy]

I tried to break this up into easy-to-review commits, but ymmv.

• [naga spv-out] Spill arrays and matrices for runtime indexing.

  Improve handling of Access expressions whose base is an array or
  matrix (not a pointer to such), and whose index is not known at
  compile time. SPIR-V does not have instructions that can do this
  directly, so spill such values to temporary variables, and perform the
  accesses using OpAccessChain instructions applied to the
  temporaries.

  When performing chains of accesses like a[i].x[j], do not reify
  intermediate values; generate a single OpAccessChain for the entire
  thing.

  Remove special cases for arrays; the same code now handles arrays and
  matrices.

  Update validation to permit dynamic indexing of matrices.

  For details, see the comments on the new tracking structures in
  naga::back::spv::Function.

  Add snapshot test index-by-value.wgsl.

  Fixes [naga spv-out] Indexing arrays by value causes unnecessary copies #6358.
  Fixes Support dynamic indexing of by-value matrices and arrays, per WGSL #4337.
  Alternative to [naga] Support dynamic indexing of by-value matrices #6362.

• [naga spv-out] Add some tracing output to Writer::write_function.

• [naga spv-out] Introduce Writer::get_resolution_pointer_id.

  Introduce a new helper function,
  naga::back::spv::Writer::get_resolution_pointer_id. Use it in
  BlockContext::write_expression_pointer.

• [naga] Add new function, GuardedIndex::from_expression.

  Pull out the code to build a naga::proc::index::GuardedIndex from a
  Handle<Expression> into its own function,
  GuardedIndex::from_expression. Use that function in
  GuardedIndex::try_resolve_to_constant.

• [naga spv-out] Abstract out NumericType::from_inner.

  Pull out the code for building a naga::back::spv::NumericType from a
  TypeInner into its own function, NumericType::from_inner. Use that
  in LocalType::from_inner.

• [naga spv-out] Use crate::proc::index::GuardedIndex.

  Replace qualified paths with a use directive.

• [naga spv-out] Gather array, matrix, and vector cases.

  This commit is just code motion.

• [naga] Test access to a member/element through a pointer.

• [naga spv-out] Clean up write_expression_pointer type adjustment.

  Replace the return_type_override argument of
  BlockContext::write_expression_pointer with an enum that says how to
  derive the return type from expr_handle's type.

  Introduce a new type, AccessTypeAdjustment, that covers possible
  derivation rules.

  This simplifies callers and the callee, in part by making the possible
  alternatives less general, and by giving them explicit names (the
  variants of the AccessTypeAdjustment enum).

• [naga spv-out] Move code to load a pointer into its own function.

  Introduce a new function,
  naga::back::spv::BlockContext::write_checked_load, that does the
  work of Expression::Load.

  This change is just code motion, and should have no effect on
  behavior. The new function will be used in later commits.

Depends on:

• [naga spv-out] Expand LocalType to permit pointers to matrices. #6387
• [naga spv-out] Simplify Writer::get_pointer_id. #6388
• [naga spv-out] Delete BlockContext::is_intermediate; use types. #6389

[jimblandy]

cc @sagudev This is mostly right, I think, but the types on the OpAccessChain instructions are wrong. I think that should be easy to fix.

[jimblandy]

@sagudev I think this is ready to test.

In the end, I think this wasn't simpler than yours, but I don't think it's too hard to understand, and I got a bunch of nice cleanups in there in the process.

[jimblandy]

Tweaked comments, and moved out another logically independent change. I think the final commit should be a pretty clear explanation of the idea now.

[teoxoy]

This looks great!

[teoxoy]

Should we move the test added by the initial PR into the new snapshot?

https://github.com/gfx-rs/wgpu/pull/6188/files#diff-f63b9d8b161661aa5df1539e321be5f259f9c2e508f3ac970a62f8a050b0fd36

[teoxoy]

For linking purposes so that we remember to update this in the future: This is not spec compliant but we already have #4390 tracking this.