Follow the actual SPIR-V type when computing an access chain expression,
instead of the canonicalized view in the optimizer's type manager.
Do this so we can generate the correct member name for a struct,
rather than using the member name for the other representative
struct type. The optimizer's type canonicalizer is insensitive to
struct member names.
Fixes tint:213
Bug: tint:3, tint:213
Change-Id: I88ec42a4cb049b011a59d5522e4cb39bc181a4fb
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/27602
Reviewed-by: dan sinclair <dsinclair@chromium.org>
Commit-Queue: dan sinclair <dsinclair@chromium.org>
This CL updates the SPIR-V Reader to create BlockStatements instead of
StatementLists.
Bug: tint:136
Change-Id: I957019446ca00306187de701f86ae3e0dd5c5eb8
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/25740
Reviewed-by: Sarah Mashayekhi <sarahmashay@google.com>
This CL removes the KillStatement from the AST and replaces all test
usage with DiscardStatement.
Bug: tint:169
Change-Id: Ie68dd3cdd54056f144d10506f05cc1f6903d1cda
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/25605
Reviewed-by: David Neto <dneto@google.com>
This CL changes the SPIRV-Reader to generate DiscardStatements instead
of KillStatements for an OpKill.
Bug: tint:168
Change-Id: Ic4223b614c822c6ad449f61aedaddc9605c42535
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/25642
Reviewed-by: David Neto <dneto@google.com>
Still TODO: OpSelect over arrays and structures, as permitted in SPIR-V 1.4
Bug: tint:3, tint:99
Change-Id: I70f6c8a43ea3339cd715813c6eb0128d66ff0df8
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/25301
Reviewed-by: dan sinclair <dsinclair@chromium.org>
This CL adds support for handling uniform data. Currently the uniform is
added to a buffer where the number is the binding value. This will need
to be updated to accept the correct mapping from the embedder.
Bug: tint:8
Change-Id: Icccccbe599a9555defa6136e384745f4093df020
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/25104
Reviewed-by: David Neto <dneto@google.com>
Old way:
- struct decorated with BufferBlock
- Uniform storage class
New way
- struct decorated with Block
- StorageBuffer storage class
Also fixes the result type for an access chain.
Bug: tint:99
Change-Id: I2324ba94bb19b369d206313de798bdfec6099fe0
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/24605
Reviewed-by: dan sinclair <dsinclair@chromium.org>
Emits it as a module-level variable. Deconstruct and throw away
the gl_PerVertex struct.
Not handled: unusual patterns that are technically valid but
which don't occur in practice:
- loading, storing, or producing intermediate values of the whole structure.
- multiple definitions of the per-vertex structure (e.g. if someone had
put both a vertex shader and a tessellation shader in the same
module.)
Bug: tint:3, tint:99
Change-Id: I3ad9ff6ab780a002367f01f385bfa7d6ddba6db9
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/24880
Reviewed-by: dan sinclair <dsinclair@chromium.org>
Take the type of a const-definition (or a write to a hoisted variable)
from the generated expression rather than the directly-converted type
from the SPIR-V ID.
This prepares the way for remapping storage buffers from
Uniform/BufferBlock to StorageBuffer/Block representation.
Bug: tint:99
Change-Id: I6104d8ad96312053562468a3cdb247932c1cb6cb
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/24604
Reviewed-by: dan sinclair <dsinclair@chromium.org>
Use the fact that in WGSL the scope corresponding to the loop construct
encloses the scope for its associated continue construct.
In our construct data structure, the two are adjacent but not
overlapping.
This improvement means that when a definition is in a loop construct,
but used only in the loop or associated continue construct, then no
hoisting is required.
Bug: tint:3
Change-Id: I8d33b8f76303ab2868306847e846b4c26899e746
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/24420
Reviewed-by: dan sinclair <dsinclair@google.com>
When we hoist a variable out of a continue construct, put it
in associated loop construct, if it exists. This reduces its
lifetime in WGSL, and easier to understand as a code reader.
Change-Id: I8f0cc37640bfe67874cbc27b55029e79e9a8992c
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/24321
Reviewed-by: dan sinclair <dsinclair@google.com>
The special case is not "single-block-loop" but rather the
case where a continue target is also its own loop header.
This can occur for single-block loops and multi-block loops.
Bug: tint:3
Change-Id: I4af1410793caf8b26a1f781e221fc0b395f07aa3
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/24320
Reviewed-by: dan sinclair <dsinclair@google.com>
For each OpPhi, make a variable to carry values from predecessor blocks
to the OpPhi. Declare the variable at the smallest scope enclosing all
the predecessor blocks (where we write to it), and the OpPhi (where we
read from it).
Bug: tint:3
Change-Id: I7898b4b903d9ee1a25a7466e3c5aaf6840550e2d
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/24181
Reviewed-by: dan sinclair <dsinclair@google.com>
Compensate for the fact that dominance does not correspond
exactly to scoping. A definition can dominate a use, but when mapped
in a naive way to constant definitiion and its use, the definition
name goes out of scope by the time you reach the use.
This is correct for storable types.
Bug: tint:3
Change-Id: I03e6c5ba68393151485ed4cdbe6b2b3d7773d1ad
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/24141
Reviewed-by: dan sinclair <dsinclair@google.com>
Avoid sinking expensive operations into control flow such as loops.
The heuristic way to achieve that is to avoid moving combinatorial
values across *any* structured construct boundaries.
Bug: tint:3
Change-Id: I91502b01166a0db64c0e652331591850df75f9d4
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/24140
Reviewed-by: dan sinclair <dsinclair@google.com>
TODOs in the test file are or were in the function_arithmetic_test.cc
Change-Id: I679dbf0019e12e26488ccc310dfb03a535628e0b
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/23462
Reviewed-by: dan sinclair <dsinclair@google.com>
This CL cleans up various formatting and lint errors.
Change-Id: Ieee14db90e36acc8b469d58abb84fcf3595321bb
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/23224
Reviewed-by: David Neto <dneto@google.com>
Add a guard variable for flow control within that if-selection.
Also, the premerge blocks are always surrounded by an if-selection,
to ensure we cause reconvergence at the end of the original if-selection
construct, just like in the original SPIR-V.
Bug: tint:3
Change-Id: I614c6840e539bf9a338058beb5b6f70484e3320a
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/23182
Reviewed-by: dan sinclair <dsinclair@google.com>
This supports the extended instructions used by the compute_boids
example.
Bug: tint:3
Change-Id: I364c343217139e489377dd2a9330058114023caa
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/23126
Reviewed-by: dan sinclair <dsinclair@google.com>
- Avoid redundant switch-break.
WGSL does an implicit break at the end of a switch case, because
it has fallthrough.
TODO: Emit fallthrough
Bug: tint:3
Change-Id: Ida44b13181a01a2c1459c0447dac496ba5b97ffc
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/22961
Reviewed-by: dan sinclair <dsinclair@google.com>
In BlockInfo, remove the backpointers from true-head, false-head, and
premerge-head to the if-selection header block.
Convert the forward references from if-selection to its internal heads
from pointers to IDs.
Bug: tint:3
Change-Id: Ic931df519795e14374bff4f60ad37a4b32f79c91
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/23140
Reviewed-by: dan sinclair <dsinclair@google.com>
Produce less noise in ASTs for a common case.
Also test that an empty continuing construct doesn't show up in the
AST. That's currently handled by the AST code. We want to keep this
behaviour even if the AST implementation changes. Right now
code change is needed when emitting the start of a continuing
construct.
Bug: tint:3
Change-Id: I96a12087e305c64647561f65d87acda907ae9c42
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/22844
Reviewed-by: dan sinclair <dsinclair@google.com>
This emits the equivalent of break-if, break-unless, continue-if,
continue-unless. But we do it via a regular if-then-else.
Adds a test matrix.
Adds all required tests except for those needing OpSwitch.
Bug: tint:3
Change-Id: I960a40aa00f95f394a92a099c8b12104010ad49f
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/22603
Reviewed-by: dan sinclair <dsinclair@google.com>
A merge block must be dominated by its own header.
This CL checks the cases where that fails because the
merge block is also the:
- the default or case header for a switch (a different header)
- the true-head, false-head, or premerge-head for an if-selection
with a different header
Bug: tint:3
Change-Id: I6dd1fae162e9d33bd9a0b43d3ca9558cebed058b
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/22680
Reviewed-by: dan sinclair <dsinclair@google.com>
This fixes the pathological cases nobody wants, and arguably
should be added to the SPIR-V spec.
If we really really want to support these cases, we can revisit.
Bug: tint:3
Change-Id: I0a75490d451676caa0933e3761098ba1fe3f8b60
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/22664
Reviewed-by: dan sinclair <dsinclair@google.com>
The `>>>` symbol was folded into the `>>` symbol in WGSL. This CL
removes `>>>` from Tint.
Change-Id: I9d900de9a6026a8099796b94aad44483f0c6813f
Reviewed-on: https://dawn-review.googlesource.com/c/tint/+/22582
Reviewed-by: Ryan Harrison <rharrison@chromium.org>
