Integrate internal changes (#589)

This change integrates the following internal changes from oldest
to newest:

```
--
1e0da2f0c6ee3fee2b3a63f224bbc8a85850b75e by Chris Bate <[email protected]>:

[tensorrt] Fix build for TensorRT 8.6

Fixes issues in plugin utilities that caused build breakage for TRT 8.6
for going on 4 weeks. Partially reverts 44ccb0370f3435b3f3366dfc534e7d17a0f2dc4e
(just the testing portions, we don't lose much) to restore everything to
working order.

--
4791ad391a765ce55a3e1a7fc0ef704a9c767028 by Chris Bate <[email protected]>:

[tensorrt] Add change missing from 1e0da2f0

Adds test changes missing from TRT 8 build fix change.

--
f63b24fcfd3f47ee1c5081ebf524b6b51df004fe by Chris Bate <[email protected]>:

[tensorrt] Fix TRT 10.9 plugin-related errors and warnings

Fix misc issues related to TRT 10.9-conditional code for the new
plugin types introduced.

--
39005ce88949b3aaa08e7497c1f325de78fd5ffa by Chris Bate <[email protected]>:

[compiler] Remove deprecated APIs

Removes APIs from Python/C for creation of explicit objects that model
the "StablehloToExecutableOptions" and the corresponding pipeline (this
was the initial API for "StableHLOToExecutable" before we realized it
obviously doesn't scale to more tasks/pipelines). Removal of these APIs
also lets us delete some ancillary C++ code.

--
ff52dce59cd2d026423ce196501df660505af73c by Chris Bate <[email protected]>:

[compiler] refactor implementation of top-level pipeline options

Previously, we used a custom `llvm::cl::SubCommand` subclass called
"OptionsContext" to implement classes/structs that encapsulated pipeline
options and could be printed/parsed using the llvm CL infrastructure.

The OptionsContext had the feature of allowing creation of an "adaptor"
that could be bound to classes that didn't depend on the CL classes at
all (essentially, using all options with external storage), which is
some that MLIR's builtin "mlir::detail::PassOptions|mlir::PassPipelineOptions"
does not support.

However, the use of this specific feature/mechanism dwindled, and use
of a class other than that derived from `mlir::detail::PassOptions` has
some big downsides -- if you want to use your class with the MLIR pass
infrastructure to manage options, then we had to create a convoluted
adaptor struct that would produce a PassPipelineOptions on the fly.
It was very convoluted and had some major limitations (no support
for enum options in the conversion adaptor, for example).

I bit the bullet and just ported the existing infrastructure to be
based on `mlir::detail::PassOptions` (which is a subclass of `llvm::cl::SubCommand`)
instead of having a custom llvm::cl::SubCommand class.

To handle "debug options" (e.g. all those flags corresponding to MLIR's
global flags for IR printing, crash reproducers, timing, etc), I added
an option to optionally attach these flags to the pipeline's options,
allowing one to avoid use of any global CL options if desired. Previously,
we always had these options but "hid" them so that their descriptions
wouldn't appear in CLI tools.

We could in the future also do something like XLA's
`LLVMCommandLineOptionsLock`, and let the compilation task entrypoint functions
(e.g. `client.get_compilation_task`) provide a different set of options to be
parsed using the global LLVM CL subcommand.

--
9a305370418e284a27edd0401a8ed822fda0d132 by Christopher Bate <[email protected]>:

[runtime] Lua TRT module: allow multi-dim shape tensors

TensorRT has support for multi-dimensional shape tensors. Remove the
assertion that shape tensors have rank-1. There's no harm in removing
the assertion because TensorRT will issue an error at build time if
it a shape tensor violates some type constraint. Predicating the assertion
based on loaded TRT version is also unnecessary since running with
a mismatched TRT runtime/compile-time version is not supported outside
of whatever guarantees TensorRT makes.

Also ensures that support for shape tensors of various data types is
future-proofed.

--
e77c68c7be74311c542b0e05d449634e0a668dca by Christopher Bate <[email protected]>:

[compiler] Dialect/Plan: improve 'plan-eliminate-shape-ops' pass

This change fixes the logic in 'plan-eliminate-shape-ops' which was
not correctly checking the callers of a function before removing
unused arguments. It also adds support for 'func.call' caller in
addition to the tensorrt call operations.

We only recently observed bugs associated with this because previous
test cases all happened to avoid dead arguments in shape tensor
functions.

--
41b25b50d20e7447f797ee6a9d68350f0bf12abf by Christopher Bate <[email protected]>:

[compiler] Improve 'stablehlo-scalar-to-arith' support for 'stablehlo.reduce'

The converter should have allowed converting any 'stablehlo.reduce' operation
that reduces the entire tensor. This is fixed and additional tests are added.

--
b4024a211652fb0c24f570a05441fd83e45b0993 by Christopher Bate <[email protected]>:

[compiler] Improve conversion and folding of 'stablehlo.scatter'

This change improves our ability to convert 'stablehlo.scatter' to
TensorRT. Previously, we were too restrictive in recognizing scatter ops
that have a semantic corresponding to ONNX/TensorRT ScatterND.

Furthermore, Stablehlo upstream contains no routines that simplify
stablehlo.scatter or recognize scatters that are trivial no-ops. This
change adds a pattern that recognize when a scatter is a slice insertion
that completely overwrites the source tensor. This op's result can just
be replaced by the (possibly reshaped) update tensor.

--
1f7187668ef5af3d633927ad0cf648dee25bea60 by Christopher Bate <[email protected]>:

[tensorrt] If shape tensor estimation was incorrect, don't set value bounds

Sometimes we may incorrectly predict a tensor is a shape tenosr, but
TensorRT thinks it is an execution tensor. When this occurs, we emit
a warning. Previously we also tried to set value bounds on the execution
tensor. This causes TensorRT to also emit a warning. However, when the
tensor is larger than the max shape tensor size, it can also cause TensorRT
to crash. This change ensures that we don't set the bounds, which was
the original intended behavior.

--
4a12de1a0b851601467e27571223bd92ababed57 by Christopher Bate <[email protected]>:

[compiler] NFC: Fix duplicate instrumentation added

Fixes duplicate pass pipeline instrumentation setup occuring in
the StablehloToExecutable and CompilationTaskBase constructors; it
should only occur in the base class.

--
2de0f76d609b463e1beaab72a72ee39c9ca2e41c by Chris Bate <[email protected]>:

[tensorrt] Adds an option to workaround TRT limitations by forcing slice parameters in-bounds

Adds option 'force-default-slice-in-bounds' to the
'tensorrt-apply-bug-wars' pass. This updates the dynamic offset and
dynamic size operands  all slice with "default"
mode (no OOB access allowed) to clamp offsets to zero when negative and
clamp the size to the shape of the input tensor. This works around
a limitations of TRT where, under certain conditions, it will fail if we
don't have very specific bounds for input shape tensors.

--
a3c52b09859e42c53d784c1c7bd74f018ca77c4c by Christopher Bate <[email protected]>:

[compiler] Don't analyze nested modules in alloc-tensors

There are two locations where we perform bufferization analysis in
`plan-alloc-tensors`. In the last update which introduced improved
Plan dialect bufferization routines, I forgot to update one call.

This doesn't have a functional effect other than preventing the analysis
from uselessly analyzing potentially nested modules.

--
ff915bd515ca532985eed6e1469d10708b03049c by Chris Bate <[email protected]>:

[compiler] Fix incorrect assumption in modification of 'scf.while'

In 'plan-alloc-tensors', we perform empty tensor elimination in the
bodies of loops and attempt to establish DPS connectivity to the loop-carried
arguments. In this process, we had an incorrect assumption that the
yielded values of the 'scf.while' "before" region were of the same size
as the region arguments. This is not true, since they can differ in number.

This change adds a small procedure to perform best effort mapping of
yielded values to the region arguments that should be used for empty
tensor elimination.

--
f62e7df1ca41684c10b8901bf280d33d95ab6de2 by Chris Bate <[email protected]>:

[tensorrt] Update TRT "layer name" translations

In the TensorRT NvInfer API, each created layer can be given a name. Previously,
we used a combination of the MLIR op name and the printed form of the MLIR
location. However, because MLIR locations can recursively refer to other locations,
this caused us to generate large layer names for code coming from JAX,
and TensorRT would emit a warning about exceeding the character limit.

This change translates the locations while avoiding printing recursive
locations and truncates layer names as if they beyond the character limit.

Note that this feature of setting the layer name based on the MLIR op name
and location information is purely for debugging and hasn't really proven
to be very useful.

--
3185b4b4e4e0692e1772361394e9669df8085dde by Zixin Huang <[email protected]>:

[compiler] Fix SerializationBitWidth computation

This MR fixes the SerializationBitWidth computation.

Test case is added in
`executor/test/Translation/translate-constant-resource.mlir`

--
90b9ad5784fe6306b6a625d2c6b101a40b11a837 by Chris Bate <[email protected]>:

[executor] Rename 'executor.constant_resource' to 'executor.data_segment'

In preparation for the refactoring of how we manage global variables
and global initialization in the Executor dialect/runtime, this change
renames 'executor.constant_resource' to 'executor.data_segment'. This
is done in order to better reflect the purpose.
The 'executor.data_segment' operation will gain the ability to represent
mutable, runtime allocated memory regions in a future commit, while
currently it only represents read-only binary data encoded into the
program.

--
7df71589bb68738370fbd949c24e7517d1ddb3fb by Chris Bate <[email protected]>:

[executor] NFC: Rename Constant -> DataSegment in Executable schema

--
1712da1da26a3b1f3a3cd1f51a2827f30f476383 by Chris Bate <[email protected]>:

[executor] Add new fields to DataSegmentOp

Adds additional fields to the DataSegmentOp to indicate access type, whether the
segment is uninitialized, and the required alignment. Previously the op was only
used as a way to store constant data, uses are updated appropriately.

--
15f62a72dd27ca3d41a9557f1c521e3b3f0d023c by Christopher Bate <[email protected]>:

[compiler] Dialect/Plan: Don't modify nested modules in plan-alloc-tensors

This fixes an issue where the final pattern set applied in the
`plan-alloc-tensors` pass was not restricted to functions in the
top-level module scope.

--
17d04aa2dccf89310fe929f31f86a1e92ed0e952 by Chris Bate <[email protected]>:

Integrate LLVM-Project @ 729416e586fba71b4f63d71b1b5c765aefbf200b

This change upgrades our llvm-project and llvm-solid dependencies by
bumping ~7 days forward from our last rebase point.

--
e6689fd28afac42239aa277501738ba8a21c2866 by Christopher Bate <[email protected]>:

[executor] Fix dyn_cast -> dyn_cast_if_present if type conversion

In target materializations, the "originalType" may be null. Discovered
in non-standard use of the ExecutorTypeConverter in a downstream
project.

--
2459e7c9a948578adee9ab131e0318d397c5325c by Chris Bate <[email protected]>:

[compiler] Improve joint bufferization of nested modules

This change updates 'plan-module-bufferize' so that it bufferizes inner
modules prior to outer modules. In addition, it saves the final
`FuncAnalysisState` of each bufferized module into a cache. This enables
outer modules that call into inner modules to access the appropriate
function analysis state required for giving more refined answers to
bufferization interface queries.

--
7f7a7338819f0276a7a8574b1091bfe8ff5e59a1 by Chris Bate <[email protected]>:

[compiler] Move cluster kind definitions to `compiler/lib/Backends`

This moves the cluster kind definitions to `compiler/lib/Backends` to
make the Plan dialect more self-contained. Previously, it was difficult
for developers to discover where the definitions of each "ClusterKind"
were located, since we had this notion of "builtin" cluster kinds which
were located in `compiler/lib/Dialect/Plan/IR/BuiltinClusterKinds.cpp`.
We now no longer have "builtins" and the cluster/backend kinds
TensorRT and Host are now located under the "Backends" directory. They
are extensions to the Plan dialect instead of being "builtin".

This also makes some Plan dialect transformations easier to grok. The
CreateClosedRegions and OutlineClusters passes no longer have special
logic for handling TensorRT or Host cluster kinds. Instead,
that logic is now moved to be colocated with the backend definitions.
Some minor improvements were made to the ClusterKindAttrInterface to
support moving the logic.

--
415792e81123f784948af29b896da23e660e48fd by Chris Bate <[email protected]>:

[compiler] Further improve Plan cluster backend organization

Finally, this change adds a `plan::InputKind` attribute to the Plan dialect
to indicate the category of input IR that is entered into the Plan
segmentation pipeline. This information can be passed along to the backends
via the `plan::ClusterKindAttr` interface methods, and different backends
can indicate whether they support a given input kind and change their
behavior based on the input kind. This further allows us to begin removing
the tight coupling of Stablehlo to the Plan dialect transformations.
```

GitOrigin-RevId: ac4be0cd2e515cf3762f579bbcd27a7a023a4932

Co-authored-by: Copybara Bot <[email protected]>

Author: christopherbate

mlir-tensorrt/DependencyProvider.cmake

@@ -17,7 +17,7 @@ if("${MLIR_TRT_USE_LLVM}" STREQUAL "prebuilt")
   set(MTRT_BUILD_LLVM_FROM_SOURCE OFF)
 endif()
 
-set(MLIR_TRT_LLVM_COMMIT "ec66c4af09263e68d800971906e60afc27d54a06")
+set(MLIR_TRT_LLVM_COMMIT "729416e586fba71b4f63d71b1b5c765aefbf200b")
 
 set(mlir_patch_dir "${CMAKE_CURRENT_LIST_DIR}/build_tools/patches/mlir")
 
@@ -42,7 +42,6 @@ else()
     PATCHES
       "${mlir_patch_dir}/0005-mlir-memref-Fix-memref.global-overly-constrained-ver.patch"
       "${mlir_patch_dir}/0006-mlir-emitc-Fix-two-EmitC-bugs.patch"
-      "${mlir_patch_dir}/0008-MLIR-Remove-unnecessary-include-from-MathToEmitC.h-t.patch"
       "${mlir_patch_dir}/0009-mlir-Support-FileLineColRange-in-LLVM-debug-translat.patch"
       "${mlir_patch_dir}/0010-MLIR-Fix-LLVMIRTransforms-build-failure-125485.patch"
     # Set the CPM cache key to the Git hash for easy navigation.
@@ -102,14 +101,17 @@ set(stablehlo_patch_dir "${CMAKE_SOURCE_DIR}/build_tools/patches/stablehlo")
 
 nv_register_package(
   NAME Stablehlo
-  VERSION 1.8.0
-  GIT_TAG 48a1e14edc8219577fcad53de1924876f855f431
+  VERSION 1.9.3
+  GIT_TAG 459897561d365ef97caba46984847f9184d472ec
   GIT_REPOSITORY "https://github.com/openxla/stablehlo.git"
   PATCHES
     "${stablehlo_patch_dir}/0001-Fix-a-couple-missing-checks-for-static-shapes-in-sta.patch"
+    "${stablehlo_patch_dir}/0002-cmake-Update-usage-of-HandleLLVMOptions-and-LLVM_DEF.patch"
     "${stablehlo_patch_dir}/0003-Don-t-insert-unnecessary-arith.index_cast-ops.patch"
     "${stablehlo_patch_dir}/0004-Fix-ZeroExtent-condition-in-simplification-pattern.patch"
-    "${stablehlo_patch_dir}/0002-cmake-Update-usage-of-HandleLLVMOptions-and-LLVM_DEF.patch"
+    "${stablehlo_patch_dir}/0005-Fix-crash-on-ComplexType-in-PointwiseToLinalgMapConv.patch"
+    "${stablehlo_patch_dir}/0006-Remove-explicit-use-of-LLVMSupport.patch"
+    "${stablehlo_patch_dir}/0007-Fix-circular-dependence-between-StablehloPasses-and-.patch"
   OPTIONS
     "STABLEHLO_ENABLE_BINDINGS_PYTHON ON"
     "STABLEHLO_BUILD_EMBEDDED ON"
@@ -161,8 +163,11 @@ set(torch_mlir_patch_dir "${CMAKE_SOURCE_DIR}/build_tools/patches/torch_mlir")
 nv_register_package(
   NAME torch_mlir
   GIT_REPOSITORY https://github.com/llvm/torch-mlir.git
-  GIT_TAG 169032010793ee7fe3e305ab920e4119fdfc3b11
-  PATCHES "${torch_mlir_patch_dir}/0001-cmake-Allow-finding-Stablehlo-via-find_package.patch"
+  GIT_TAG 0bb263e99415d43255350d29263097b4980303bf
+  PATCHES 
+    "build_tools/patches/torch_mlir/0001-cmake-Allow-finding-Stablehlo-via-find_package.patch"
+    "build_tools/patches/torch_mlir/0002-Make-compatible-with-more-recent-Stablehlo-version.patch"
+    "build_tools/patches/torch_mlir/0003-Fix-some-configuration-paths-in-LIT-cfg.patch"
   EXCLUDE_FROM_ALL TRUE
   # We need to specify an existing directory that is not actually a submodule
   # since GIT_SUBMODULES does not except the empty string due to
@@ -172,8 +177,7 @@ nv_register_package(
     "TORCH_MLIR_OUT_OF_TREE_BUILD ON"
     "TORCH_MLIR_ENABLE_STABLEHLO ON"
     "TORCH_MLIR_EXTERNAL_STABLEHLO_DIR find_package"
-    "MLIR_DIR ${CMAKE_BINARY_DIR}/lib/cmake/mlir"
-    "LLVM_DIR ${LLVM_BINARY_DIR}/lib/cmake/llvm"
+    "TORCH_MLIR_ENABLE_TOSA OFF"
 )
 
 #-------------------------------------------------------------------------------------

mlir-tensorrt/build_tools/cmake/Targets.cmake

@@ -82,3 +82,33 @@ function(add_mlir_tensorrt_doc name)
   add_custom_target("MLIRTensorRT${name}DocGen" DEPENDS "${CMAKE_BINARY_DIR}/${ARG_OUTPUT_FILE}")
   add_dependencies("mlir-tensorrt-doc" "MLIRTensorRT${name}DocGen")
 endfunction()
+
+# ------------------------------------------------------------------------------
+# Declare a Plan dialect extension backend library.
+# ------------------------------------------------------------------------------
+function(add_mlir_tensorrt_backend_library target)
+  cmake_parse_arguments(ARG "" "TD" "" ${ARGN})
+  cmake_path(SET SRC_TD
+    NORMALIZE
+    "${CMAKE_CURRENT_SOURCE_DIR}/../../../include/${ARG_TD}")
+  cmake_path(SET BIN_TD
+    NORMALIZE
+    "${CMAKE_CURRENT_BINARY_DIR}/../../../include/${ARG_TD}")
+  # Tablegen output paths have to be relative to CMAKE_CURRENT_BINARY_DIR.
+  cmake_path(RELATIVE_PATH BIN_TD
+    BASE_DIRECTORY ${CMAKE_CURRENT_BINARY_DIR})
+
+  set(LLVM_TARGET_DEFINITIONS "${SRC_TD}")
+  
+  string(REPLACE ".td" "Attrs.h.inc" h_inc_file ${BIN_TD})
+  string(REPLACE ".td" "Attrs.cpp.inc" cpp_inc_file ${BIN_TD})
+  mlir_tablegen("${h_inc_file}" -gen-attrdef-decls)
+  mlir_tablegen("${cpp_inc_file}" -gen-attrdef-defs)  
+
+  add_public_tablegen_target(${target}IncGen)
+
+  add_mlir_tensorrt_library(${target}
+    PARTIAL_SOURCES_INTENDED
+    ${ARG_UNPARSED_ARGUMENTS}
+    DEPENDS ${target}IncGen)
+endfunction()

mlir-tensorrt/build_tools/patches/stablehlo/0001-Fix-a-couple-missing-checks-for-static-shapes-in-sta.patch

@@ -1,12 +1,12 @@
-From cc45fc70ad111e768404524b29e9f168a31baec4 Mon Sep 17 00:00:00 2001
+From b387be5903482200f4c36f64f8ed102c288c0c29 Mon Sep 17 00:00:00 2001
 From: Christopher Bate <[email protected]>
 Date: Wed, 27 Nov 2024 00:10:11 +0000
-Subject: [PATCH 1/4] Fix a couple missing checks for static shapes in
+Subject: [PATCH 1/7] Fix a couple missing checks for static shapes in
  `stablehlo-aggressive-folder`
 
 ---
  .../stablehlo_aggressive_folder.mlir          | 27 +++++++++++++------
- .../transforms/StablehloAggressiveFolder.cpp  |  9 +++++++
+ .../StablehloAggressiveFolder.cpp             |  9 +++++++
  2 files changed, 28 insertions(+), 8 deletions(-)
 
 diff --git a/stablehlo/tests/transforms/stablehlo_aggressive_folder.mlir b/stablehlo/tests/transforms/stablehlo_aggressive_folder.mlir
@@ -79,11 +79,11 @@ index 5b21a10d..c90c89c6 100644
  }
 
  // -----
-diff --git a/stablehlo/transforms/StablehloAggressiveFolder.cpp b/stablehlo/transforms/StablehloAggressiveFolder.cpp
-index 49942c42..5fd20fec 100644
---- a/stablehlo/transforms/StablehloAggressiveFolder.cpp
-+++ b/stablehlo/transforms/StablehloAggressiveFolder.cpp
-@@ -258,6 +258,9 @@ struct FoldAddOpPattern final : OpRewritePattern<mlir::stablehlo::AddOp> {
+diff --git a/stablehlo/transforms/optimization/StablehloAggressiveFolder.cpp b/stablehlo/transforms/optimization/StablehloAggressiveFolder.cpp
+index 2b5198b4..52a28e97 100644
+--- a/stablehlo/transforms/optimization/StablehloAggressiveFolder.cpp
++++ b/stablehlo/transforms/optimization/StablehloAggressiveFolder.cpp
+@@ -257,6 +257,9 @@ struct FoldAddOpPattern final : OpRewritePattern<mlir::stablehlo::AddOp> {
 
    LogicalResult matchAndRewrite(mlir::stablehlo::AddOp op,
                                  PatternRewriter& rewriter) const override {

mlir-tensorrt/build_tools/patches/stablehlo/0002-cmake-Update-usage-of-HandleLLVMOptions-and-LLVM_DEF.patch

@@ -1,7 +1,7 @@
-From f45c6b58fb35aeb40c71bfc5e3f18c20c6294238 Mon Sep 17 00:00:00 2001
+From 4949f0f91bf256e01f13dae1ddcd2139d2c41d85 Mon Sep 17 00:00:00 2001
 From: Christopher Bate <[email protected]>
 Date: Sat, 15 Feb 2025 22:02:17 +0000
-Subject: [PATCH 2/4] [cmake] Update usage of `HandleLLVMOptions` and
+Subject: [PATCH 2/7] [cmake] Update usage of `HandleLLVMOptions` and
  `LLVM_DEFINITIONS`
 
 This change attempts to resolve issues with use of `HandleLLVMOptions`

mlir-tensorrt/build_tools/patches/stablehlo/0003-Don-t-insert-unnecessary-arith.index_cast-ops.patch

@@ -1,16 +1,16 @@
-From d31c6ef5be9d1fca897dbeceee740085840b5f08 Mon Sep 17 00:00:00 2001
+From 1e5096183747c9d41eec0d624726d25454f10f9c Mon Sep 17 00:00:00 2001
 From: Christopher Bate <[email protected]>
 Date: Mon, 10 Mar 2025 22:04:05 +0000
-Subject: [PATCH 3/4] Don't insert unnecessary `arith.index_cast` ops
+Subject: [PATCH 3/7] Don't insert unnecessary `arith.index_cast` ops
 
 ---
  .../StablehloAggressiveSimplification.cpp     | 20 ++-----------------
  1 file changed, 2 insertions(+), 18 deletions(-)
 
-diff --git a/stablehlo/transforms/StablehloAggressiveSimplification.cpp b/stablehlo/transforms/StablehloAggressiveSimplification.cpp
-index ebb6c027..8527739a 100644
---- a/stablehlo/transforms/StablehloAggressiveSimplification.cpp
-+++ b/stablehlo/transforms/StablehloAggressiveSimplification.cpp
+diff --git a/stablehlo/transforms/optimization/StablehloAggressiveSimplification.cpp b/stablehlo/transforms/optimization/StablehloAggressiveSimplification.cpp
+index f32f8d66..8028f714 100644
+--- a/stablehlo/transforms/optimization/StablehloAggressiveSimplification.cpp
++++ b/stablehlo/transforms/optimization/StablehloAggressiveSimplification.cpp
 @@ -394,34 +394,18 @@ struct DynamicIotaOpToBroadcast : public OpRewritePattern<DynamicIotaOp> {
 
      auto iotaDimension = static_cast<int64_t>(iota.getIotaDimension());

mlir-tensorrt/build_tools/patches/stablehlo/0004-Fix-ZeroExtent-condition-in-simplification-pattern.patch

@@ -1,17 +1,17 @@
-From 97c312a13288e95bdb2be5c197ebf0ed4019f434 Mon Sep 17 00:00:00 2001
+From 94b386bd28b610a3218508c391acd926412e57f1 Mon Sep 17 00:00:00 2001
 From: Christopher Bate <[email protected]>
 Date: Mon, 10 Mar 2025 22:51:38 +0000
-Subject: [PATCH 4/4] Fix ZeroExtent condition in simplification pattern
+Subject: [PATCH 4/7] Fix ZeroExtent condition in simplification pattern
 
 Attribute doesn't have to be a DenseElementsAttr.
 ---
- .../transforms/StablehloAggressiveSimplificationPatterns.td    | 3 ++-
+ .../optimization/StablehloAggressiveSimplificationPatterns.td  | 3 ++-
  1 file changed, 2 insertions(+), 1 deletion(-)
 
-diff --git a/stablehlo/transforms/StablehloAggressiveSimplificationPatterns.td b/stablehlo/transforms/StablehloAggressiveSimplificationPatterns.td
+diff --git a/stablehlo/transforms/optimization/StablehloAggressiveSimplificationPatterns.td b/stablehlo/transforms/optimization/StablehloAggressiveSimplificationPatterns.td
 index 9cbcc07c..60396cc4 100644
---- a/stablehlo/transforms/StablehloAggressiveSimplificationPatterns.td
-+++ b/stablehlo/transforms/StablehloAggressiveSimplificationPatterns.td
+--- a/stablehlo/transforms/optimization/StablehloAggressiveSimplificationPatterns.td
++++ b/stablehlo/transforms/optimization/StablehloAggressiveSimplificationPatterns.td
 @@ -94,7 +94,8 @@ def SortedDims : AttrConstraint<
      "is sorted dimensions">;
 

mlir-tensorrt/build_tools/patches/stablehlo/0005-Fix-crash-on-ComplexType-in-PointwiseToLinalgMapConv.patch

@@ -0,0 +1,95 @@
+From 139c779d447d6163c51dbe9d8735b2062025f032 Mon Sep 17 00:00:00 2001
+From: Christopher Bate <[email protected]>
+Date: Fri, 21 Mar 2025 03:28:26 +0000
+Subject: [PATCH 5/7] Fix crash on ComplexType in PointwiseToLinalgMapConverter
+
+---
+ .../conversions/linalg/tests/pointwise.mlir   | 23 ++++++++++++++
+ .../transforms/StablehloToLinalgPointwise.cpp | 30 +++++++++++++++----
+ 2 files changed, 48 insertions(+), 5 deletions(-)
+
+diff --git a/stablehlo/conversions/linalg/tests/pointwise.mlir b/stablehlo/conversions/linalg/tests/pointwise.mlir
+index 6dc76f24..7a9f71aa 100644
+--- a/stablehlo/conversions/linalg/tests/pointwise.mlir
++++ b/stablehlo/conversions/linalg/tests/pointwise.mlir
+@@ -23,6 +23,29 @@ func.func @float_add(%lhs: tensor<2x2xf32>,
+ 
+ // -----
+ 
++// CHECK: #map = affine_map<(d0, d1) -> (d0, d1)>
++// CHECK-LABEL: func @complex_add_const
++// CHECK-PRIMITIVE-LABEL: func @complex_add_const
++func.func @complex_add_const(%lhs: tensor<2x2xcomplex<f32>>,
++                             %rhs: tensor<2x2xcomplex<f32>>)
++                             -> tensor<2x2xcomplex<f32>> {
++
++  // CHECK: %[[CST:.+]] = complex.constant [1.000000e-01 : f32, 2.000000e-01 : f32] : complex<f32>
++  // CHECK: linalg.generic
++  // CHECK: ^bb0(%[[IN:.+]]: complex<f32>, %[[OUT:.+]]: complex<f32>)
++  // CHECK: %[[RESULT:[a-zA-Z0-9_]*]] = complex.add %[[IN]], %[[CST]]
++  // CHECK: linalg.yield %[[RESULT]]
++
++  // CHECK-PRIMITIVE: linalg.map
++  // CHECK-PRIMITIVE: complex.add
++  %cst = stablehlo.constant dense<(0.1, 0.2)> : tensor<2x2xcomplex<f32>>
++  %0 = "stablehlo.add"(%lhs, %cst) {someattr}
++      : (tensor<2x2xcomplex<f32>>, tensor<2x2xcomplex<f32>>) -> tensor<2x2xcomplex<f32>>
++  func.return %0 : tensor<2x2xcomplex<f32>>
++}
++
++// -----
++
+ // CHECK-LABEL: func @float_add_dynamic_encoding
+ // CHECK-PRIMITIVE-LABEL: func @float_add_dynamic_encoding
+ func.func @float_add_dynamic_encoding(
+diff --git a/stablehlo/conversions/linalg/transforms/StablehloToLinalgPointwise.cpp b/stablehlo/conversions/linalg/transforms/StablehloToLinalgPointwise.cpp
+index 707db6a7..301dfdc2 100644
+--- a/stablehlo/conversions/linalg/transforms/StablehloToLinalgPointwise.cpp
++++ b/stablehlo/conversions/linalg/transforms/StablehloToLinalgPointwise.cpp
+@@ -114,6 +114,28 @@ FailureOr<PointwiseConversionInfo> checkOperandsAndResults(
+   return PointwiseConversionInfo{maxRank, resultTy};
+ }
+ 
++/// If `input` is a splat constant value, materialize the scalar splat
++/// value. Otherwise, return nullopt.
++std::optional<Value> materializeSplatScalarConstant(RewriterBase &rewriter,
++                                                    Location loc, Value input) {
++  SplatElementsAttr attr;
++  Type elementType = mlir::getElementTypeOrSelf(input.getType());
++  if (!matchPattern(input, m_Constant(&attr))) return {};
++  if (isa<IntegerType, FloatType, IndexType>(elementType)) {
++    return rewriter
++        .create<arith::ConstantOp>(loc, elementType,
++                                   attr.getSplatValue<TypedAttr>())
++        .getResult();
++  }
++  if (isa<ComplexType>(elementType)) {
++    return rewriter
++        .create<complex::ConstantOp>(loc, elementType,
++                                     attr.getSplatValue<ArrayAttr>())
++        .getResult();
++  }
++  return {};
++}
++
+ /// Converts a HLO operation to a linalg.map op that contains the corresponding
+ /// scalar operations.
+ template <typename OpTy>
+@@ -160,11 +182,9 @@ struct PointwiseToLinalgMapConverter : OpConversionPattern<OpTy> {
+     SmallVector<Value> mappedInputs;
+     SmallVector<Value> scalarInputs;
+     for (Value input : adaptor.getOperands()) {
+-      DenseElementsAttr attr;
+-      if (matchPattern(input, m_Constant(&attr)) && attr.isSplat()) {
+-        scalarInputs.push_back(rewriter.create<arith::ConstantOp>(
+-            loc, cast<ShapedType>(input.getType()).getElementType(),
+-            attr.getSplatValue<TypedAttr>()));
++      if (std::optional<Value> splatVal =
++              materializeSplatScalarConstant(rewriter, loc, input)) {
++        scalarInputs.push_back(*splatVal);
+       } else if (getRank(input) == maxRank) {
+         mappedInputs.push_back(coerceTensorShape(
+             rewriter, loc, cast<TypedValue<ShapedType>>(input),
+-- 
+2.46.0
+

mlir-tensorrt/build_tools/patches/stablehlo/0006-Remove-explicit-use-of-LLVMSupport.patch

@@ -0,0 +1,24 @@
+From fafd462cdae170e3b615cb559e907b30840d7cf7 Mon Sep 17 00:00:00 2001
+From: Christopher Bate <[email protected]>
+Date: Fri, 21 Mar 2025 04:00:05 +0000
+Subject: [PATCH 6/7] Remove explicit use of LLVMSupport
+
+---
+ stablehlo/transforms/conversions/CMakeLists.txt | 1 -
+ 1 file changed, 1 deletion(-)
+
+diff --git a/stablehlo/transforms/conversions/CMakeLists.txt b/stablehlo/transforms/conversions/CMakeLists.txt
+index e1da2c8b..7aac80f1 100644
+--- a/stablehlo/transforms/conversions/CMakeLists.txt
++++ b/stablehlo/transforms/conversions/CMakeLists.txt
+@@ -18,7 +18,6 @@ add_mlir_library(StablehloTypeConversion
+   TypeConversion.cpp
+ 
+   LINK_LIBS PUBLIC
+-  LLVMSupport
+   MLIRIR
+   MLIRSupport
+   MLIRTransformUtils
+-- 
+2.46.0
+

mlir-tensorrt/build_tools/patches/stablehlo/0007-Fix-circular-dependence-between-StablehloPasses-and-.patch

@@ -0,0 +1,63 @@
+From e0b197588de8367b729b726009be028da3ed74a7 Mon Sep 17 00:00:00 2001
+From: Christopher Bate <[email protected]>
+Date: Sun, 23 Mar 2025 01:54:33 +0000
+Subject: [PATCH 7/7] Fix circular dependence between StablehloPasses and
+ StablehloOptimizationPasses
+
+Fixes build when BUILD_SHARED_LIBS=ON.
+---
+ stablehlo/transforms/CMakeLists.txt              | 12 +++++++++++-
+ stablehlo/transforms/optimization/CMakeLists.txt |  1 +
+ 2 files changed, 12 insertions(+), 1 deletion(-)
+
+diff --git a/stablehlo/transforms/CMakeLists.txt b/stablehlo/transforms/CMakeLists.txt
+index 4787369d..50c87304 100644
+--- a/stablehlo/transforms/CMakeLists.txt
++++ b/stablehlo/transforms/CMakeLists.txt
+@@ -39,6 +39,16 @@ set(LLVM_TARGET_DEFINITIONS VhloToVersionPatterns.td)
+ mlir_tablegen(VhloToVersionPatterns.h.inc --gen-rewriters)
+ add_public_tablegen_target(VhloToVersionPatterns)
+ 
++add_mlir_dialect_library(StablehloPassUtils
++  PARTIAL_SOURCES_INTENDED
++  PassUtils.cpp
++
++  LINK_LIBS PUBLIC
++  ChloOps
++  MLIRIR
++  StablehloOps
++  )
++
+ 
+ add_mlir_dialect_library(StablehloPasses
+   PARTIAL_SOURCES_INTENDED
+@@ -59,7 +69,6 @@ add_mlir_dialect_library(StablehloPasses
+   StablehloRefineShapes.cpp
+   VhloLegalizeToStablehlo.cpp
+   VhloToVersion.cpp
+-  PassUtils.cpp
+ 
+   DEPENDS
+   ChloDecompositionPatternsIncGen
+@@ -90,6 +99,7 @@ add_mlir_dialect_library(StablehloPasses
+   StablehloLinalgTransforms
+   StablehloOps
+   StablehloOptimizationPasses
++  StablehloPassUtils
+   StablehloTypeInference
+   VhloOps
+ )
+diff --git a/stablehlo/transforms/optimization/CMakeLists.txt b/stablehlo/transforms/optimization/CMakeLists.txt
+index d43d77be..d063a49d 100644
+--- a/stablehlo/transforms/optimization/CMakeLists.txt
++++ b/stablehlo/transforms/optimization/CMakeLists.txt
+@@ -41,5 +41,6 @@ add_mlir_dialect_library(StablehloOptimizationPasses
+   MLIRTransformUtils
+   StablehloBase
+   StablehloOps
++  StablehloPassUtils
+   StablehloTypeInference
+ )
+-- 
+2.46.0
+