[NVVM] Drop Dealloc of shared memory buffers in conversion to NVVM (i…

…ree-org#9743)

Shared memory does not have traditional alloc/dealloc semantics but rather a semantics tht is closer to alloca with block-wide visibility and addressability (i.e. not thread-local stack visibility).
Dealloc ops of shared memory may serve the purpose of annotations for computing live ranges to better reuse limited shared memory space but cannot be lowered to NVVM as there is no abstraction to represent this.

This revision simply drops such deallocs when converting to NVVM.

Integrate llvm-project and bump dependencies. (iree-org#9700)

MHLO: df41f42da81a59c23391f994d4bced6afc95a505
LLVM: aa58b7b1e30f
TF: 8a3ae2cc003b7d005d1a3d901d03cc7373fbaf3c

* Cherry-picking RVV revert commit:
  * https://reviews.llvm.org/D127871
  * Issue reported on patch. If not reverted before next integrate, will need to be carried forward.

Removing todo for a flow.tensor.cast equivalent. (iree-org#9691)

Since we don't handle unranked tensors there's not much tensor.cast
seems to do that tensor.reshape doesn't as it doesn't allow element
type conversion.

Closes iree-org#6418.

Add utility functions for VM to EmitC conversion (iree-org#9708)

Add support for lowering memref.alloca to VM dialect. (iree-org#9694)

Co-authored-by: Hanhan Wang <[email protected]>

Migrate TFLite's stateful lowerings to MLProgram (iree-org#9671)

We have stateful behavior supported via MLProgram. Migrating the TFLite
lowerings from directly lowering to IREE means we can better seperate
TFLite legalization from being IREE specific.

Migrate TFLite's stateful lowerings to MLProgram (iree-org#9671)

We have stateful behavior supported via MLProgram. Migrating the TFLite
lowerings from directly lowering to IREE means we can better seperate
TFLite legalization from being IREE specific.

[Transform] Add a new transform op that applies patterns (iree-org#9676)

This revision introduces a `transform.iree.apply_patterns` operation that operates on
an isolated from above op and applies a set of patterns while listening and updating
transform dialect handles.

The list of patterns is specified via attributes that are additive and roughly play the role of
populate functions.

iree-opt is extended so that it can run transform dialect directly without files, which
results in general simplifications.

Report an error when an export function does not map to a vm.func. (i…

…ree-org#9665)

Fixes iree-org#4245

[vulkan] Clean up various Vulkan specific e2e tests (iree-org#9672)

These tests are accumulated over time but we haven't re-check
whether they are useful anymore.

* `compare.mlir` is identical to the one under `xla_ops/`.
* `dot_general.mlir` is a subset of the one under `xla_ops/`.
* `vectorized_conv.mlir` is merged into `conv.mlir`.
* `log_plus_one.mlir` is identical to the one under `xla_ops/`.
* `pw_add_multiwg.mlir` was added during the early time and
  just checks a pointwise addition; not really useful now.
* `reduce.mlir` is merged into the one under `xla_ops/`.