We now have two functions `writeOpLoad` and `writeOpStore` which are
in charge of writing SpvOpLoad and SpvOpStore instructions.
`writeOpStore` also keeps track of pointer stores in a "store cache."
Subsequent loads from that same pointer will be found in the cache and
will return the value stored in that pointer instead.
Such a cache definitely cannot work in the face of control flow, so we
make the following concessions:
- `pruneReachableOps` is now `pruneConditionalOps`. Any pointers that
are altered inside a potentially-unreachable block are cleared from
the cache entirely.
- The entire store cache is cleared at all OpLabels within a loop.
The cache also cannot work in the presence of swizzled stores, so we
make another significant concession:
- The entire store cache is cleared whenever we store into a non-memory
pointer (e.g., assigning into a swizzled LValue, such as `foo.xz`).
Despite these significant limitations, this manages to dramatically
shrink many real-world examples.
Change-Id: I0981a0cf7b45b064e153e9ada271494c8e00cad5
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/530054
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This is a reland of commit 60ff0facbf
Structs are now deduplicated using a [Type*, SpvId] map.
Original change's description:
> Use op cache when emitting types.
>
> We no longer need to maintain a separate `fTypeMap` for mapping types
> to SpvIds, since the op cache handles this automatically.
>
> We also now support deduplicating instructions that don't have a result,
> such as decorations. (In particular, we needed to avoid emitting the
> SpvDecorationArrayStride op every time the array type was accessed, but
> this op doesn't have a result ID.)
>
> Change-Id: I779b8c8e3de5973b8f487b28c0a8ece9a1041845
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/529732
> Reviewed-by: Brian Osman <brianosman@google.com>
> Commit-Queue: John Stiles <johnstiles@google.com>
> Auto-Submit: John Stiles <johnstiles@google.com>
Change-Id: I9f6a78d58e8af38a1fd690a8860d8b5aa3193be6
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/529748
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
This reverts commit 60ff0facbf.
Reason for revert: Broke D3D bots
Original change's description:
> Use op cache when emitting types.
>
> We no longer need to maintain a separate `fTypeMap` for mapping types
> to SpvIds, since the op cache handles this automatically.
>
> We also now support deduplicating instructions that don't have a result,
> such as decorations. (In particular, we needed to avoid emitting the
> SpvDecorationArrayStride op every time the array type was accessed, but
> this op doesn't have a result ID.)
>
> Change-Id: I779b8c8e3de5973b8f487b28c0a8ece9a1041845
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/529732
> Reviewed-by: Brian Osman <brianosman@google.com>
> Commit-Queue: John Stiles <johnstiles@google.com>
> Auto-Submit: John Stiles <johnstiles@google.com>
Change-Id: I0e2187f88f2a945fd6f88ce75ff815e03d2f7df5
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/529747
Auto-Submit: Brian Osman <brianosman@google.com>
Commit-Queue: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
We no longer need to maintain a separate `fTypeMap` for mapping types
to SpvIds, since the op cache handles this automatically.
We also now support deduplicating instructions that don't have a result,
such as decorations. (In particular, we needed to avoid emitting the
SpvDecorationArrayStride op every time the array type was accessed, but
this op doesn't have a result ID.)
Change-Id: I779b8c8e3de5973b8f487b28c0a8ece9a1041845
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/529732
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Previously, we stringized the types and put them into fTypeMap. Using
the op cache is a simpler mechanism that should work equally well.
Output diffs are almost all ID reorderings. In a few cases we
managed to deduplicate function types that stringize differently but
come out the same in SPIR-V (e.g. no float/half distinction).
Change-Id: If7de5b2dafa12d05c3c2c497a243e9e3908dfee7
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/529805
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Brian Osman <brianosman@google.com>
The output changes here are almost entirely a wash, because we already
had support for caching scalars and vectors. Almost all changes are just
inconsequential reorderings of IDs, and the removal of RelaxedPrecision
decorators on constants (which were not meaningful).
Change-Id: I45340c4a240cb504b7c4a934b3db178d2f39ec99
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/528709
Reviewed-by: Arman Uguray <armansito@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
This reverts commit 0f4304e6e7.
Reason for revert: breaks Adreno 6xx
Original change's description:
> Add RelaxedPrecision decoration to function-call temp vars.
>
> This is really same basic issue as http://review.skia.org/446640. We
> were creating a temp variable but ignoring its type's precision.
>
> Change-Id: I9a5fedd7ada864d36757fc196f42ff95bac7d706
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/446718
> Commit-Queue: John Stiles <johnstiles@google.com>
> Auto-Submit: John Stiles <johnstiles@google.com>
> Reviewed-by: Brian Osman <brianosman@google.com>
Change-Id: I6ae4e264b60f7f38a1abb5f1d0324461a33c896d
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/446742
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
Bot-Commit: Rubber Stamper <rubber-stamper@appspot.gserviceaccount.com>
This is really same basic issue as http://review.skia.org/446640. We
were creating a temp variable but ignoring its type's precision.
Change-Id: I9a5fedd7ada864d36757fc196f42ff95bac7d706
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/446718
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Making a VectorCast from a compile-time constant will perform the cast
at compile-time instead; previously, we did not apply this optimization.
This simplified a few test outputs in subtle ways. (In particular, the
SPIR-V codegen used to occasionally decorate OpConstantComposite of
constant numbers with RelaxedPrecision, and no longer appears to do
this. This should have no effect on results either way AFAICS.)
Because we don't return VectorCast constructors containing compile-time
constant values, we do not need to implement compareConstant for this
constructor; they only wrap non-compile-time-constant expressions.
Change-Id: I28c1f337f64d6f20fb86bc0f58e225af4bd7b26c
Bug: skia:11032
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/392197
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This reverts commit 7bba1f55e8.
Change-Id: I707a3c215f37376086e22eaa43916afeed6da4c7
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/388456
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
SPIR-V doesn't seem to mind overlapping function names, since they're
not load-bearing in any way, but this keeps us consistent with the other
code generators.
Change-Id: Ifdb4cb17795da88eabc0db841af746fb76caf423
Bug: skia:10851
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/387757
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Arguments without side-effects that aren't read from more than once can
be moved directly into the inlined function, and don't need a scratch
variable. This can allow functions like `guarded_divide` to inline
completely in more cases.
Change-Id: I0bfce35635cf9779f4af1bc0790da966ccfe4230
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/386678
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
This reverts commit a9c187e5cc.
Change-Id: Icbfb8abdfc67fc2e6428d97a6cdede2726fb56e4
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/385596
Auto-Submit: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: John Stiles <johnstiles@google.com>
This can eliminate const variables which have been completely folded
away, unnecessary synthetic variables created during codegen/inlining,
or code that simply didn't need to exist at all.
Change-Id: I37a65e455e6527a6a6c2f4dde918f48d84dc2638
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/383496
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
This reverts commit 92748af1a5.
Reason for revert: SkSLCommaSideEffects_GPU crashing on Android
Original change's description:
> Inline functions of the form 'return (expr)' only.
>
> This drastically reduces the number of functions which we allow to be
> inlined. If this change does not hurt our performance, it will allow us
> to trivially remove hundreds of LOC. All current data leads us to
> believe that it may affect the Mali 400 but is highly unlikely to change
> results on any other device in the tree.
>
> More info: http://go/optimization-in-sksl-inliner
>
> Change-Id: Ia6b706742ce5407453e0e697b6c1f9201084c0e8
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/384858
> Auto-Submit: John Stiles <johnstiles@google.com>
> Commit-Queue: John Stiles <johnstiles@google.com>
> Reviewed-by: Brian Osman <brianosman@google.com>
> Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
TBR=brianosman@google.com,ethannicholas@google.com,johnstiles@google.com
Change-Id: I6a670dacaa58fe3386ff50375ac6d1cac4fd7f2c
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/385161
Reviewed-by: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
This reverts commit 9e476b798f.
Reason for revert: Angry Vulkan bots
Original change's description:
> Refactored SPIR-V RelaxedPrecision handling
>
> The RelaxedPrecision decoration is now handled by nextId(), to make it
> easier to see all spots where a RelaxedPrecision decoration might be
> necessary. The goal of this initial refactor is not to actually fix the
> issues with RelaxedPrecision decorations, but rather to lay the
> groundwork for doing so in followup CLs.
>
> The initial intent of this change was to not affect the SPIR-V at all,
> saving modifications for followups, but there ended up being three kinds
> of changes to the output:
>
> 1. Doing things at nextId() time rather than later means some
> decorations move to an earlier spot in the output. This results in
> diffs, but should not cause any behavioral changes.
> 2. We were incorrectly tagging bools as RelaxedPrecision in some
> situations. By funneling things through fewer code paths, the refactor
> would have caused this to happen in even more situations, and the code
> responsible for the bug was being rewritten in this CL anyway, so it
> seemed worth just fixing the issue as part of this change.
> 3. Funneling things through fewer code paths ended up adding
> (correct) RelaxedPrecision modifiers to binary operations that had
> previously been missing them. It seemed better to just let it happen
> than to try to maintain bug-for-bug compatibility with the previous
> approach.
>
> Change-Id: Ia9654d6b5754e2c797e02226660cb618c9189b36
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/384318
> Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
> Reviewed-by: Brian Osman <brianosman@google.com>
TBR=brianosman@google.com,ethannicholas@google.com,johnstiles@google.com
Change-Id: I9ada728e5fd5798bc1179640560c2e6045b7efd1
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/385158
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
The RelaxedPrecision decoration is now handled by nextId(), to make it
easier to see all spots where a RelaxedPrecision decoration might be
necessary. The goal of this initial refactor is not to actually fix the
issues with RelaxedPrecision decorations, but rather to lay the
groundwork for doing so in followup CLs.
The initial intent of this change was to not affect the SPIR-V at all,
saving modifications for followups, but there ended up being three kinds
of changes to the output:
1. Doing things at nextId() time rather than later means some
decorations move to an earlier spot in the output. This results in
diffs, but should not cause any behavioral changes.
2. We were incorrectly tagging bools as RelaxedPrecision in some
situations. By funneling things through fewer code paths, the refactor
would have caused this to happen in even more situations, and the code
responsible for the bug was being rewritten in this CL anyway, so it
seemed worth just fixing the issue as part of this change.
3. Funneling things through fewer code paths ended up adding
(correct) RelaxedPrecision modifiers to binary operations that had
previously been missing them. It seemed better to just let it happen
than to try to maintain bug-for-bug compatibility with the previous
approach.
Change-Id: Ia9654d6b5754e2c797e02226660cb618c9189b36
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/384318
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This drastically reduces the number of functions which we allow to be
inlined. If this change does not hurt our performance, it will allow us
to trivially remove hundreds of LOC. All current data leads us to
believe that it may affect the Mali 400 but is highly unlikely to change
results on any other device in the tree.
More info: http://go/optimization-in-sksl-inliner
Change-Id: Ia6b706742ce5407453e0e697b6c1f9201084c0e8
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/384858
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
This reverts commit a04692f69e.
Reason for revert: Angry Vulkan bots.
Original change's description:
> Fixed a number of spots where we should have been using RelaxedPrecision
>
> Our SPIR-V output was missing many RelaxedPrecision decorations, which
> was presumably impacting performance.
>
> Change-Id: Iee32d4a42f37af167fe0e45f3db94c2142129695
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/384178
> Reviewed-by: Brian Osman <brianosman@google.com>
> Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
TBR=egdaniel@google.com,brianosman@google.com,ethannicholas@google.com,johnstiles@google.com
Change-Id: If4fe945cb363c9b61b5a4abfde649a437689d2eb
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/384217
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
Our SPIR-V output was missing many RelaxedPrecision decorations, which
was presumably impacting performance.
Change-Id: Iee32d4a42f37af167fe0e45f3db94c2142129695
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/384178
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: Ethan Nicholas <ethannicholas@google.com>
This CL will be used to test for potential performance regressions (or
improvements) that we might cause by disabling this optimization pass.
It will be reverted in ~1 day.
Change-Id: I26b7687c341eb6d81231406381c39869cfccf6d6
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/381259
Auto-Submit: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
This reverts commit 50b1b2b90d.
Reason for revert: ending experiment
Original change's description:
> Disable control-flow analysis in SkSL. (Performance experiment)
>
> This CL will be used to test for potential performance regressions (or
> improvements?) that we might incur by disabling this optimization pass.
>
> It will be reverted in ~1 day.
>
> Change-Id: I775cdb0c95df81fa25ebbd66e4ff01f64c660f68
> Bug: skia:11319
> Reviewed-on: https://skia-review.googlesource.com/c/skia/+/378456
> Commit-Queue: John Stiles <johnstiles@google.com>
> Reviewed-by: Brian Osman <brianosman@google.com>
> Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
TBR=brianosman@google.com,ethannicholas@google.com,johnstiles@google.com
Change-Id: Ie385a82db237ff5651348d82b9651f8ba09375b9
No-Presubmit: true
No-Tree-Checks: true
No-Try: true
Bug: skia:11319
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/379581
Reviewed-by: John Stiles <johnstiles@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
This CL will be used to test for potential performance regressions (or
improvements?) that we might incur by disabling this optimization pass.
It will be reverted in ~1 day.
Change-Id: I775cdb0c95df81fa25ebbd66e4ff01f64c660f68
Bug: skia:11319
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/378456
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
These variables were later being eliminated by the dead-code-elimination
pass, so you can't see them directly in the final output, but removing
them affects the name mangling off all future symbols, so it causes an
enormous ripple effect in the diff. And of course, it's a waste of time
and memory to synthesize IRNodes just to destroy them later.
If we disable control-flow analysis, we lose the dead-code-elimination
pass entirely; this change is also beneficial for emitting better code
when optimizations are turned off.
Change-Id: I882b3be4f3fd99b77d99b6abe128f26bb9252c89
Bug: skia:11319
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/375776
Commit-Queue: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Reviewed-by: Ethan Nicholas <ethannicholas@google.com>
This allows uniforms to be specified without an explicit `layout(set=N)`
modifier. They will assume a default set value instead.
This turns out to fix a handful of tests in Metal/SPIR-V which were
written with GLSL in mind, or adapted from real generated GLSL code, and
didn't have layout information specified on their uniforms. It will also
make it easier to write SkSL tests using uniforms that can compile
either as a runtime effect or as plain Metal/SPIR-V code.
Change-Id: Id79ec06f278b913a45c09c2e6211195dc98b42c0
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/359838
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
Commit-Queue: John Stiles <johnstiles@google.com>
This will allow us to load these inputs for unit testing in `dm`.
Change-Id: Id256ba7c30d3ec94b98048e47af44cf9efe580d5
Bug: skia:11009
Reviewed-on: https://skia-review.googlesource.com/c/skia/+/357282
Commit-Queue: John Stiles <johnstiles@google.com>
Auto-Submit: John Stiles <johnstiles@google.com>
Reviewed-by: Brian Osman <brianosman@google.com>
