BRL-CAD/OpenShadingLanguage
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

deps: Raise llvm minimum to 14.0 (#2019)

Browse Source 
Drop support for llvm 11, 12, 13.

Simplify lots of conditional compilation.

With 14 as the minimum, the new pass manager and opaque pointers are
both supported, so we are able to get rid of all code related to
support of the legacy pass manager, as well as typed pointers. This
is a lot of removed code!

We also simplify the testing matrix as a result of having a less
expansive range of LLVM versions to consider.

When updating INSTALL.md, I noticed that we still had **bold** for
"new minimum dependencies" left over from when we first released 1.14.
So I'm unbolding them now, except for the ones we actually have
changed for 1.15.

I had originally intended to pull the minimum to 15, but actually that
makes it more difficult to continue CI testing vs VFX Platform 2022,
since the ASWF-docker container for 2022 only support up to LLVM 14.
And simultaneously, once I removed the code for legacy PM and typed
pointers, there was, in comparison, only a small amount of additional
simplification we could squeeze out of moving the minimum to llvm 15.
So for OSL 1.15, we will have a minimum llvm 14, and then for OSL 1.16
when we would feel more comfortable droping all VFX Platform 2022
support and move to the 2023 containers as our oldest testing
environment, we can bump to llvm 15.

Signed-off-by: Larry Gritz <lg@larrygritz.com>
This commit is contained in:
Larry Gritz
2025-08-11 07:27:37 -07:00
committed by GitHub
parent 89fc3a24d3
commit 04307c7488
11 changed files with 107 additions and 664 deletions
Download Patch File Download Diff File Expand all files Collapse all files
.github/workflows/ci.yml
+16 -30
View File
@@ -51,24 +51,10 @@ jobs:
fail-fast: false
matrix:
include:
# - desc: gcc9/C++17 llvm11 py3.7 oiio2.5 sse4 batch-b8avx2
# nametag: linux-vfx2021
# runner: ubuntu-latest
# container: aswftesting/ci-osl:2021-clang11
# vfxyear: 2021
# old_node: 1
# cxx_std: 17
# openexr_ver: v3.1.3
# openimageio_ver: v2.5.4.0
# python_ver: 3.7
# pybind11_ver: v2.7.0
# simd: sse4.2
# batched: b8_AVX2_noFMA
# setenvs: export USE_OPENVDB=0
- desc: gcc9/C++17 llvm13 py3.9 exr3.1 oiio-rel avx2
- desc: gcc9/C++17 llvm14 py3.9 exr3.1 oiio-rel avx2
nametag: linux-vfx2022
runner: ubuntu-latest
container: aswftesting/ci-osl:2022-clang13
container: aswftesting/ci-osl:2022-clang14
vfxyear: 2022
old_node: 1
cxx_std: 17
@@ -77,10 +63,10 @@ jobs:
pybind11_ver: v2.9.0
simd: avx2,f16c
batched: b8_AVX2
- desc: clang12/C++17 llvm12 oiio-main py3.9 avx2 batch-avx512
nametag: linux-clang12-llvm12-batch
- desc: clang14/C++17 llvm14 oiio-main py3.9 avx2 batch-avx512
nametag: linux-clang14-llvm14-batch
runner: ubuntu-latest
container: aswftesting/ci-osl:2022-clang12
container: aswftesting/ci-osl:2022-clang14
vfxyear: 2022
old_node: 1
cxx_std: 17
@@ -91,10 +77,10 @@ jobs:
simd: avx2,f16c
batched: b8_AVX2,b8_AVX512,b16_AVX512
setenvs: USE_OPENVDB=0
- desc: gcc9/C++17 llvm11 py3.9 exr3.1 oiio3.0 sse2 batch-b4sse2
- desc: gcc9/C++17 llvm14 py3.9 exr3.1 oiio3.0 sse2 batch-b4sse2
nametag: linux-vfx2022-clang
runner: ubuntu-latest
container: aswftesting/ci-osl:2022-clang13
container: aswftesting/ci-osl:2022-clang14
vfxyear: 2022
old_node: 1
cxx_std: 17
@@ -104,10 +90,10 @@ jobs:
pybind11_ver: v2.9.0
simd: sse2
batched: b4_SSE2
- desc: oldest everything gcc9/C++17 llvm11 py3.9 oiio2.5 no-simd
- desc: oldest everything gcc9/C++17 llvm14 py3.9 oiio2.5 no-simd
nametag: linux-oldest
runner: ubuntu-latest
container: aswftesting/ci-osl:2022-clang11
container: aswftesting/ci-osl:2022-clang14
vfxyear: 2022
old_node: 1
cxx_std: 17
@@ -280,7 +266,7 @@ jobs:
# ^^ exclude python-oslquery test until the ASWF container properly
# includes OIIO's python bindings, then we can remove that.
# Address and leak sanitizers
# Address and leak sanitizers (debug build)
- desc: sanitizers
nametag: sanitizer
runner: ubuntu-latest
@@ -386,8 +372,8 @@ jobs:
USE_OPENVDB=0
OPENCOLORIO_CMAKE_FLAGS="-DCMAKE_CXX_COMPILER=g++"
- desc: Debug gcc9/C++17 llvm11 py3.10 oiio2.5 exr3.1 sse4
nametag: linux-debug-gcc7-llvm11
- desc: Debug gcc9/C++17 llvm14 py3.10 oiio2.5 exr3.1 sse4
nametag: linux-debug-gcc9-llvm14
runner: ubuntu-22.04
cxx_compiler: g++-9
cxx_std: 17
@@ -397,11 +383,11 @@ jobs:
python_ver: "3.10"
simd: sse4.2
setenvs: export CMAKE_BUILD_TYPE=Debug
LLVM_VERSION=11.0.0 LLVM_DISTRO_NAME=ubuntu-20.04
LLVM_VERSION=14.0.0 LLVM_DISTRO_NAME=ubuntu-18.04
PUGIXML_VERSION=v1.9
CTEST_TEST_TIMEOUT=240
- desc: gcc10/C++17 llvm11 oiio-2.5 avx2
nametag: linux-2021ish-gcc10-llvm11
- desc: gcc10/C++17 llvm14 oiio-2.5 avx2
nametag: linux-2021ish-gcc10-llvm14
runner: ubuntu-22.04
cxx_compiler: g++-10
cxx_std: 17
@@ -411,7 +397,7 @@ jobs:
pybind11_ver: v2.8.1
python_ver: "3.10"
simd: avx2,f16c
setenvs: export LLVM_VERSION=11.0.0 LLVM_DISTRO_NAME=ubuntu-20.04
setenvs: export LLVM_VERSION=14.0.0 LLVM_DISTRO_NAME=ubuntu-18.04
OPENIMAGEIO_CMAKE_FLAGS="-DBUILD_FMT_VERSION=7.0.1"
PUGIXML_VERSION=v1.10
- desc: latest releases gcc11/C++17 llvm17 oiio-3.0 exr3.2 py3.12 avx2 batch-b16avx512
INSTALL.md
+9 -9
View File
@@ -19,17 +19,17 @@ Dependencies
OSL requires the following dependencies or tools.
NEW or CHANGED minimum dependencies since the last major release are **bold**.
* Build system: [CMake](https://cmake.org/) **3.19 or newer** (tested
* Build system: [CMake](https://cmake.org/) 3.19 or newer (tested
through 4.0)
* A suitable C++17 compiler to build OSL itself, which may be any of:
- **GCC 9.3** or newer (tested through gcc 13.1)
- **Clang 5** or newer (tested through clang 19)
- GCC 9.3 or newer (tested through gcc 13.1)
- Clang 5 or newer (tested through clang 20)
- Microsoft Visual Studio 2017 or newer
- Intel C++ compiler **icc version 19** or newer or LLVM-based icx compiler
- Intel C++ compiler icc version 19 or newer or LLVM-based icx compiler
version 2022 or newer.
* **[OpenImageIO](http://openimageio.org) 2.5 or newer** (tested through 3.0
* [OpenImageIO](http://openimageio.org) 2.5 or newer (tested through 3.0
and main)
OSL uses OIIO both for its texture mapping functionality as well as
@@ -49,15 +49,15 @@ NEW or CHANGED minimum dependencies since the last major release are **bold**.
$OpenImageIO_ROOT/lib to be in your LD_LIBRARY_PATH (or
DYLD_LIBRARY_PATH on OS X).
* [LLVM](http://www.llvm.org) 11, 12, 13, 14, 15, 16, 17, 18, 19, 20, including
clang libraries.
* [LLVM](http://www.llvm.org) **14.0 or newer**, 15, 16, 17, 18, 19, 20,
including clang libraries.
* (optional) For GPU rendering on NVIDIA GPUs:
* [OptiX](https://developer.nvidia.com/rtx/ray-tracing/optix) 7.0 or higher.
* [Cuda](https://developer.nvidia.com/cuda-downloads) 9.0 or higher. It is
recommended that you use 11.0 or higher.
* [Imath](https://github.com/AcademySoftwareFoundation/Imath) **3.1 or newer**.
* [Imath](https://github.com/AcademySoftwareFoundation/Imath) 3.1 or newer.
* [Flex](https://github.com/westes/flex) 2.5.35 or newer and
[GNU Bison](https://www.gnu.org/software/bison/) 2.7 or newer.
Note that on some MacOS/xcode releases, the system-installed Bison is too
@@ -70,7 +70,7 @@ NEW or CHANGED minimum dependencies since the last major release are **bold**.
* (optional) Python: If you are building the Python bindings or running the
testsuite:
* **Python >= 3.9** (tested through 3.13)
* **pybind11 >= 2.7** (tested through 3.0)
* pybind11 >= 2.7 (tested through 3.0)
* NumPy (tested through 2.2.4)
* (optional) Qt5 >= 5.6 or Qt6 (tested Qt5 through 5.15 and Qt6 through 6.8).
If not found at build time, the `osltoy` application will be disabled.
Makefile
+1 -1
View File
@@ -359,7 +359,7 @@ help:
@echo " Finding and Using Dependencies:"
@echo " USE_QT=0 Skip anything that needs Qt"
@echo " LLVM-related options:"
@echo " LLVM_VERSION=12.0 Specify which LLVM version to use"
@echo " LLVM_VERSION=19.0 Specify which LLVM version to use"
@echo " LLVM_DIRECTORY=xx Specify where LLVM lives"
@echo " LLVM_NAMESPACE=xx Specify custom LLVM namespace"
@echo " LLVM_STATIC=1 Use static LLVM libraries"
src/cmake/cuda_macros.cmake
+4 -14
View File
@@ -134,12 +134,6 @@ function ( MAKE_CUDA_BITCODE src suffix generated_bc extra_clang_args )
set (CLANG_MSVC_FIX -Wno-ignored-attributes -Wno-unknown-attributes)
endif ()
if (NOT LLVM_OPAQUE_POINTERS AND ${LLVM_VERSION} VERSION_GREATER_EQUAL 15.0)
# Until we fully support opaque pointers, we need to disable
# them when using LLVM 15.
list (APPEND LLVM_COMPILE_FLAGS -Xclang -no-opaque-pointers)
endif ()
if (NOT CUDA_NO_FTZ)
set (CLANG_FTZ_FLAG "-fcuda-flush-denormals-to-zero")
endif ()
@@ -221,14 +215,10 @@ function ( CUDA_SHADEOPS_COMPILE prefix output_bc output_ptx input_srcs headers
list ( APPEND shadeops_bc_list ${shadeops_bc} )
endforeach ()
if (LLVM_NEW_PASS_MANAGER)
# There is no --nvptx-assign-valid-global-names flag for the new
# pass manager, but it appears to run this pass by default.
string(REPLACE "-O" "O" opt_tool_flags ${CUDA_OPT_FLAG_CLANG})
set (opt_tool_flags -passes="default<${opt_tool_flags}>")
else()
set (opt_tool_flags ${CUDA_OPT_FLAG_CLANG} --nvptx-assign-valid-global-names)
endif ()
# There is no --nvptx-assign-valid-global-names flag for the new
# pass manager, but it appears to run this pass by default.
string(REPLACE "-O" "O" opt_tool_flags ${CUDA_OPT_FLAG_CLANG})
set (opt_tool_flags -passes="default<${opt_tool_flags}>")
# Link all of the individual LLVM bitcode files, and emit PTX for the linked bitcode
add_custom_command ( OUTPUT ${linked_bc} ${linked_ptx}
src/cmake/externalpackages.cmake
+1 -30
View File
@@ -57,7 +57,7 @@ checked_find_package (pugixml REQUIRED
# LLVM library setup
checked_find_package (LLVM REQUIRED
VERSION_MIN 11.0
VERSION_MIN 14.0
VERSION_MAX 20.9
PRINT LLVM_SYSTEM_LIBRARIES CLANG_LIBRARIES
LLVM_SHARED_MODE)
@@ -79,36 +79,7 @@ if (LLVM_VERSION VERSION_GREATER_EQUAL 15.0 AND CMAKE_COMPILER_IS_CLANG
"If you are using LLVM 15 or higher, you should also use clang version "
"15 or higher, or you may get build errors.${ColorReset}\n")
endif ()
if (LLVM_VERSION VERSION_GREATER_EQUAL 16.0)
if (CMAKE_CXX_STANDARD VERSION_LESS 17)
message (WARNING "${ColorYellow}LLVM 16+ requires C++17 or higher. "
"Please set CMAKE_CXX_STANDARD to 17 or higher.${ColorReset}\n")
endif ()
if (CMAKE_COMPILER_IS_GNUCC AND (GCC_VERSION VERSION_LESS 7.0))
message (WARNING "${ColorYellow}LLVM 16+ requires gcc 7.0 or higher.${ColorReset}\n")
endif ()
if (CMAKE_COMPILER_IS_CLANG
AND NOT (CLANG_VERSION_STRING VERSION_GREATER_EQUAL 5.0
OR APPLECLANG_VERSION_STRING VERSION_GREATER_EQUAL 5.0))
message (WARNING "${ColorYellow}LLVM 16+ requires clang 5.0 or higher.${ColorReset}\n")
endif ()
endif ()
# Use opaque pointers starting with LLVM 16
if (${LLVM_VERSION} VERSION_GREATER_EQUAL 16.0)
set(LLVM_OPAQUE_POINTERS ON)
add_compile_definitions (OSL_LLVM_OPAQUE_POINTERS)
else()
set(LLVM_OPAQUE_POINTERS OFF)
endif()
# Enable new pass manager for LLVM 16+
if (${LLVM_VERSION} VERSION_GREATER_EQUAL 16.0)
set(LLVM_NEW_PASS_MANAGER ON)
add_compile_definitions (OSL_LLVM_NEW_PASS_MANAGER)
else()
set(LLVM_NEW_PASS_MANAGER OFF)
endif()
checked_find_package (partio)
src/cmake/llvm_macros.cmake
-6
View File
@@ -74,12 +74,6 @@ function ( EMBED_LLVM_BITCODE_IN_CPP src_list suffix output_name list_to_append_
list (TRANSFORM include_dirs PREPEND -I
OUTPUT_VARIABLE ALL_INCLUDE_DIRS)
if (NOT LLVM_OPAQUE_POINTERS AND ${LLVM_VERSION} VERSION_GREATER_EQUAL 15.0)
# Until we fully support opaque pointers, we need to disable
# them when using LLVM 15.
list (APPEND LLVM_COMPILE_FLAGS -Xclang -no-opaque-pointers)
endif ()
# Command to turn the .cpp file into LLVM assembly language .s, into
# LLVM bitcode .bc, then back into a C++ file with the bc embedded!
add_custom_command ( OUTPUT ${src_bc}
src/include/OSL/llvm_util.h
-1
View File
@@ -1240,7 +1240,6 @@ private:
llvm::Value* op_combine_4x_vectors(llvm::Value* half_vec_1,
llvm::Value* half_vec_2);
void setup_legacy_optimization_passes(int optlevel, bool target_host);
void setup_new_optimization_passes(int optlevel, bool target_host);
};
src/liboslexec/batched_backendllvm.cpp
-19
View File
@@ -1304,25 +1304,6 @@ BatchedBackendLLVM::llvm_store_value(llvm::Value* new_val, llvm::Value* dst_ptr,
if (!type.is_closure_based() && t.aggregate > 1)
dst_ptr = ll.GEP(dst_type, dst_ptr, 0, component);
#ifndef OSL_LLVM_OPAQUE_POINTERS
if ((const llvm::Type*)ll.type_ptr(ll.llvm_typeof(new_val))
!= ll.llvm_typeof(dst_ptr)) {
std::cerr << " new_val type=";
{
llvm::raw_os_ostream os_cerr(std::cerr);
ll.llvm_typeof(new_val)->print(os_cerr);
}
std::cerr << " dest_ptr type=";
{
llvm::raw_os_ostream os_cerr(std::cerr);
ll.llvm_typeof(dst_ptr)->print(os_cerr);
}
std::cerr << std::endl;
}
OSL_ASSERT((const llvm::Type*)ll.type_ptr(ll.llvm_typeof(new_val))
== ll.llvm_typeof(dst_ptr));
#endif
// Finally, store the value.
ll.op_store(new_val, dst_ptr);
return true;
src/liboslexec/batched_llvm_gen.cpp
-30
View File
@@ -1056,18 +1056,6 @@ LLVMGEN(llvm_gen_andor)
// sometimes we could not force the data type to be an bool and it remains
// an int, for those cases we will need to convert the boolean to int
if (!result.forced_llvm_bool()) {
#ifndef OSL_LLVM_OPAQUE_POINTERS
llvm::Type* resultType = rop.ll.llvm_typeof(
rop.llvm_get_pointer(result));
OSL_ASSERT(
(resultType
== reinterpret_cast<llvm::Type*>(rop.ll.type_wide_int_ptr()))
|| (resultType
== reinterpret_cast<llvm::Type*>(rop.ll.type_int_ptr())));
llvm::Type* typeOfR = rop.ll.llvm_typeof(i1_res);
OSL_ASSERT(typeOfR == rop.ll.type_wide_bool()
|| typeOfR == rop.ll.type_bool());
#endif
i1_res = rop.ll.op_bool_to_int(i1_res);
}
@@ -2170,15 +2158,6 @@ LLVMGEN(llvm_gen_bitwise_binary_op)
// We allowed boolean values to pass through or,and,xor
// so we might need to promote to integer before storage
if (!Result.forced_llvm_bool()) {
#ifndef OSL_LLVM_OPAQUE_POINTERS
llvm::Type* resultType = rop.ll.llvm_typeof(
rop.llvm_get_pointer(Result));
OSL_ASSERT(
(resultType
== reinterpret_cast<llvm::Type*>(rop.ll.type_wide_int_ptr()))
|| (resultType
== reinterpret_cast<llvm::Type*>(rop.ll.type_int_ptr())));
#endif
llvm::Type* typeOfR = rop.ll.llvm_typeof(r);
if (typeOfR == rop.ll.type_wide_bool()
|| typeOfR == rop.ll.type_bool()) {
@@ -3075,15 +3054,6 @@ LLVMGEN(llvm_gen_compare_op)
final_result = rop.ll.llvm_mask_to_native(final_result);
}
} else {
#ifndef OSL_LLVM_OPAQUE_POINTERS
llvm::Type* resultType = rop.ll.llvm_typeof(
rop.llvm_get_pointer(Result));
OSL_ASSERT(
(resultType
== reinterpret_cast<llvm::Type*>(rop.ll.type_wide_int_ptr()))
|| (resultType
== reinterpret_cast<llvm::Type*>(rop.ll.type_int_ptr())));
#endif
final_result = rop.ll.op_bool_to_int(final_result);
}
src/liboslexec/llvm_passes.h
+2 -39
View File
@@ -68,16 +68,10 @@ public:
// what the native mask type should be. And if necessary maintain a
// 16 bit and 32 bit native mask representation to be passed as a
// livein.
m_native_mask_type = llvm::FixedVectorType::get(llvm_type_int32,
WidthT);
#if OSL_LLVM_VERSION >= 112
m_native_mask_type = llvm::FixedVectorType::get(llvm_type_int32,
WidthT);
m_wide_zero_initializer = llvm::ConstantDataVector::getSplat(
WidthT, llvm::ConstantInt::get(context, llvm::APInt(32, 0)));
#else
m_wide_zero_initializer = llvm::ConstantVector::getSplat(
llvm::ElementCount(WidthT, false),
llvm::ConstantInt::get(context, llvm::APInt(32, 0)));
#endif
}
bool run(llvm::Function& F) const
@@ -404,37 +398,6 @@ public:
}
};
// Legacy pass manager adapter
template<int WidthT>
class LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks final
: public llvm::FunctionPass {
PreventBitMasksFromBeingLiveinsToBasicBlocks<WidthT> m_pass;
public:
static char ID;
LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks() : FunctionPass(ID) {}
bool doInitialization(llvm::Module& M) override
{
m_pass.initialize(M.getContext());
return false; // Module was not changed
}
bool runOnFunction(llvm::Function& F) override { return m_pass.run(F); }
};
// No need to worry about static variable collisions if included multiple
// places because of the anonymous namespace, each translation unit
// including this file will need its own static members defined. LLVM will
// assign IDs when they get registered, so this initialization value is not
// important.
template<> char LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks<4>::ID = 0;
template<> char LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks<8>::ID = 0;
template<> char LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks<16>::ID = 0;
} // end of anonymous namespace
} // namespace pvt
src/liboslexec/llvm_util.cpp
+74 -485
View File
@@ -13,8 +13,8 @@
#include <OSL/oslconfig.h>
#include <OSL/wide.h>
#if OSL_LLVM_VERSION < 110
# error "LLVM minimum version required for OSL is 11.0"
#if OSL_LLVM_VERSION < 140
# error "LLVM minimum version required for OSL is 14.0"
#endif
OSL_PRAGMA_WARNING_PUSH
@@ -51,12 +51,8 @@ OSL_GCC_PRAGMA(GCC diagnostic ignored "-Wmaybe-uninitialized")
#else
# include <llvm/TargetParser/Host.h>
#endif
#include <llvm/MC/TargetRegistry.h>
#include <llvm/Support/raw_os_ostream.h>
#if OSL_LLVM_VERSION < 140
# include <llvm/Support/TargetRegistry.h>
#else
# include <llvm/MC/TargetRegistry.h>
#endif
#include <llvm/Analysis/BasicAliasAnalysis.h>
#include <llvm/Analysis/TargetTransformInfo.h>
@@ -86,46 +82,39 @@ OSL_GCC_PRAGMA(GCC diagnostic ignored "-Wmaybe-uninitialized")
#include <llvm/Support/DynamicLibrary.h>
#if OSL_LLVM_VERSION >= 120
# include <llvm/CodeGen/Passes.h>
#endif
#include <llvm/CodeGen/Passes.h>
#ifdef OSL_LLVM_NEW_PASS_MANAGER
// New pass manager
# include <llvm/Analysis/LoopAnalysisManager.h>
# include <llvm/Passes/PassBuilder.h>
# include <llvm/Transforms/IPO/ArgumentPromotion.h>
# include <llvm/Transforms/IPO/ConstantMerge.h>
# include <llvm/Transforms/IPO/DeadArgumentElimination.h>
# include <llvm/Transforms/IPO/GlobalDCE.h>
# include <llvm/Transforms/IPO/GlobalOpt.h>
# include <llvm/Transforms/IPO/SCCP.h>
# include <llvm/Transforms/IPO/StripDeadPrototypes.h>
# include <llvm/Transforms/Scalar/ADCE.h>
# include <llvm/Transforms/Scalar/CorrelatedValuePropagation.h>
# include <llvm/Transforms/Scalar/DCE.h>
# include <llvm/Transforms/Scalar/DeadStoreElimination.h>
# include <llvm/Transforms/Scalar/EarlyCSE.h>
# include <llvm/Transforms/Scalar/IndVarSimplify.h>
# include <llvm/Transforms/Scalar/JumpThreading.h>
# include <llvm/Transforms/Scalar/LICM.h>
# include <llvm/Transforms/Scalar/LoopDeletion.h>
# include <llvm/Transforms/Scalar/LoopIdiomRecognize.h>
# include <llvm/Transforms/Scalar/LoopRotation.h>
# include <llvm/Transforms/Scalar/LoopUnrollPass.h>
# include <llvm/Transforms/Scalar/LowerExpectIntrinsic.h>
# include <llvm/Transforms/Scalar/MemCpyOptimizer.h>
# include <llvm/Transforms/Scalar/Reassociate.h>
# include <llvm/Transforms/Scalar/SCCP.h>
# include <llvm/Transforms/Scalar/SROA.h>
# include <llvm/Transforms/Scalar/SimpleLoopUnswitch.h>
# include <llvm/Transforms/Scalar/SimplifyCFG.h>
# include <llvm/Transforms/Scalar/TailRecursionElimination.h>
# include <llvm/Transforms/Utils/Mem2Reg.h>
#else
// Legacy pass manager
# include <llvm/Transforms/IPO/PassManagerBuilder.h>
#endif
#include <llvm/Analysis/LoopAnalysisManager.h>
#include <llvm/Passes/PassBuilder.h>
#include <llvm/Transforms/IPO/ArgumentPromotion.h>
#include <llvm/Transforms/IPO/ConstantMerge.h>
#include <llvm/Transforms/IPO/DeadArgumentElimination.h>
#include <llvm/Transforms/IPO/GlobalDCE.h>
#include <llvm/Transforms/IPO/GlobalOpt.h>
#include <llvm/Transforms/IPO/SCCP.h>
#include <llvm/Transforms/IPO/StripDeadPrototypes.h>
#include <llvm/Transforms/Scalar/ADCE.h>
#include <llvm/Transforms/Scalar/CorrelatedValuePropagation.h>
#include <llvm/Transforms/Scalar/DCE.h>
#include <llvm/Transforms/Scalar/DeadStoreElimination.h>
#include <llvm/Transforms/Scalar/EarlyCSE.h>
#include <llvm/Transforms/Scalar/IndVarSimplify.h>
#include <llvm/Transforms/Scalar/JumpThreading.h>
#include <llvm/Transforms/Scalar/LICM.h>
#include <llvm/Transforms/Scalar/LoopDeletion.h>
#include <llvm/Transforms/Scalar/LoopIdiomRecognize.h>
#include <llvm/Transforms/Scalar/LoopRotation.h>
#include <llvm/Transforms/Scalar/LoopUnrollPass.h>
#include <llvm/Transforms/Scalar/LowerExpectIntrinsic.h>
#include <llvm/Transforms/Scalar/MemCpyOptimizer.h>
#include <llvm/Transforms/Scalar/Reassociate.h>
#include <llvm/Transforms/Scalar/SCCP.h>
#include <llvm/Transforms/Scalar/SROA.h>
#include <llvm/Transforms/Scalar/SimpleLoopUnswitch.h>
#include <llvm/Transforms/Scalar/SimplifyCFG.h>
#include <llvm/Transforms/Scalar/TailRecursionElimination.h>
#include <llvm/Transforms/Utils/Mem2Reg.h>
// additional includes for PTX generation
#include <llvm/Analysis/TargetLibraryInfo.h>
@@ -198,9 +187,7 @@ static std::unique_ptr<std::vector<std::shared_ptr<LLVMMemoryManager>>>
static int jit_mem_hold_users = 0;
#if OSL_LLVM_VERSION >= 120
llvm::raw_os_ostream raw_cout(std::cout);
#endif
}; // namespace
@@ -396,14 +383,11 @@ public:
// New pass manager state, mainly here because these are template classes
// for which forward declarations in a public header are tricky.
struct LLVM_Util::NewPassManager {
#ifdef OSL_LLVM_NEW_PASS_MANAGER
llvm::LoopAnalysisManager loop_analysis_manager;
llvm::FunctionAnalysisManager function_analysis_manager;
llvm::CGSCCAnalysisManager cgscc_analysis_manager;
llvm::ModuleAnalysisManager module_analysis_manager;
llvm::ModulePassManager module_pass_manager;
#endif
};
@@ -457,12 +441,6 @@ LLVM_Util::LLVM_Util(const PerThreadInfo& per_thread_info, int debuglevel,
OIIO::spin_lock lock(llvm_global_mutex);
if (!m_thread->llvm_context) {
m_thread->llvm_context = new llvm::LLVMContext();
#if OSL_LLVM_VERSION >= 150 && !defined(OSL_LLVM_OPAQUE_POINTERS)
m_thread->llvm_context->setOpaquePointers(false);
// FIXME: For now, keep using typed pointers. We're going to have
// to fix this and switch to opaque pointers by llvm 16.
#endif
//static SetCommandLineOptionsForLLVM sSetCommandLineOptionsForLLVM;
}
if (!m_thread->llvm_jitmm) {
@@ -622,28 +600,6 @@ LLVM_Util::SetupLLVM()
llvm::initializeTarget(registry);
llvm::initializeCodeGen(registry);
#ifndef OSL_LLVM_NEW_PASS_MANAGER
// LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks
static llvm::RegisterPass<
LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks<4>>
sRegCustomPass2(
"PreventBitMasksFromBeingLiveinsToBasicBlocks<4>",
"Prevent Bit Masks <4xi1> From Being Liveins To Basic Blocks Pass",
false /* Only looks at CFG */, false /* Analysis Pass */);
static llvm::RegisterPass<
LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks<8>>
sRegCustomPass0(
"PreventBitMasksFromBeingLiveinsToBasicBlocks<8>",
"Prevent Bit Masks <8xi1> From Being Liveins To Basic Blocks Pass",
false /* Only looks at CFG */, false /* Analysis Pass */);
static llvm::RegisterPass<
LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks<16>>
sRegCustomPass1(
"PreventBitMasksFromBeingLiveinsToBasicBlocks<16>",
"Prevent Bit Masks <16xi1> From Being Liveins To Basic Blocks Pass",
false /* Only looks at CFG */, false /* Analysis Pass */);
#endif
if (debug()) {
for (auto t : llvm::TargetRegistry::targets())
std::cout << "Target: '" << t.getName() << "' "
@@ -1559,12 +1515,9 @@ LLVM_Util::make_jit_execengine(std::string* err, TargetISA requestedISA,
options.NoZerosInBSS = false;
options.GuaranteedTailCallOpt = false;
#if OSL_LLVM_VERSION < 120
options.StackAlignmentOverride = 0;
#endif
options.FunctionSections = true;
options.UseInitArray = false;
options.FloatABIType = llvm::FloatABI::Default;
options.FunctionSections = true;
options.UseInitArray = false;
options.FloatABIType = llvm::FloatABI::Default;
#if OSL_LLVM_VERSION < 190
options.RelaxELFRelocations = false;
#endif
@@ -1818,10 +1771,7 @@ LLVM_Util::nvptx_target_machine()
options.AllowFPOpFusion = llvm::FPOpFusion::Fast;
options.NoZerosInBSS = 0;
options.GuaranteedTailCallOpt = 0;
#if OSL_LLVM_VERSION < 120
options.StackAlignmentOverride = 0;
#endif
options.UseInitArray = 0;
options.UseInitArray = 0;
// Verify that the NVPTX target has been initialized
std::string error;
@@ -1892,21 +1842,12 @@ LLVM_Util::InstallLazyFunctionCreator(void* (*P)(const std::string&))
void
LLVM_Util::setup_optimization_passes(int optlevel, bool target_host)
{
#ifdef OSL_LLVM_NEW_PASS_MANAGER
setup_new_optimization_passes(optlevel, target_host);
#else
setup_legacy_optimization_passes(optlevel, target_host);
#endif
}
void
LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
{
#ifdef OSL_LLVM_NEW_PASS_MANAGER
# if OSL_LLVM_VERSION <= 110
# error "New pass manager not supported in LLVM 11 and earlier"
# endif
OSL_DEV_ONLY(std::cout << "setup_new_optimization_passes " << optlevel);
OSL_ASSERT(m_new_pass_manager == nullptr);
@@ -2004,7 +1945,7 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
case 12: {
llvm::ModulePassManager& mpm = m_new_pass_manager->module_pass_manager;
# if 0 // PRETTY_GOOD_KEEP_AS_REF
#if 0 // PRETTY_GOOD_KEEP_AS_REF
mpm.addPass(llvm::ModuleInlinerWrapperPass());
mpm.addPass(
llvm::createModuleToFunctionPassAdaptor(llvm::SimplifyCFGPass()));
@@ -2013,11 +1954,11 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
{
llvm::FunctionPassManager fpm;
fpm.addPass(llvm::SimplifyCFGPass());
# if OSL_LLVM_VERSION < 160
# if OSL_LLVM_VERSION < 160
fpm.addPass(llvm::SROAPass());
# else
# else
fpm.addPass(llvm::SROAPass(llvm::SROAOptions::PreserveCFG));
# endif
# endif
fpm.addPass(llvm::EarlyCSEPass());
fpm.addPass(llvm::ReassociatePass());
@@ -2031,11 +1972,11 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
fpm.addPass(llvm::DCEPass());
fpm.addPass(llvm::JumpThreadingPass());
# if OSL_LLVM_VERSION < 160
# if OSL_LLVM_VERSION < 160
fpm.addPass(llvm::SROAPass());
# else
# else
fpm.addPass(llvm::SROAPass(llvm::SROAOptions::PreserveCFG));
# endif
# endif
fpm.addPass(llvm::InstCombinePass());
// Added
@@ -2045,7 +1986,7 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
mpm.addPass(llvm::GlobalDCEPass());
mpm.addPass(llvm::ConstantMergePass());
# else
#else
mpm.addPass(llvm::ModuleInlinerWrapperPass());
mpm.addPass(
llvm::createModuleToFunctionPassAdaptor(llvm::SimplifyCFGPass()));
@@ -2054,14 +1995,14 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
{
llvm::FunctionPassManager fpm;
fpm.addPass(llvm::SimplifyCFGPass());
# if OSL_LLVM_VERSION < 160
# if OSL_LLVM_VERSION < 160
fpm.addPass(llvm::SROAPass());
# else
# else
// PreserveCFG is the same behavior as earlier versions, but changing
// to ModifyCFG here and other places may improve performance.
// https://reviews.llvm.org/D138238
fpm.addPass(llvm::SROAPass(llvm::SROAOptions::PreserveCFG));
# endif
# endif
fpm.addPass(llvm::EarlyCSEPass());
// Eliminate and remove as much as possible up front
@@ -2082,7 +2023,11 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
llvm::LoopPassManager lpm;
const bool use_memory_ssa = true; // Needed by LICM
lpm.addPass(llvm::LoopRotatePass());
# if OSL_LLVM_VERSION >= 150
lpm.addPass(llvm::LICMPass(llvm::LICMOptions()));
# else
lpm.addPass(llvm::LICMPass());
# endif
lpm.addPass(llvm::SimpleLoopUnswitchPass(false));
fpm.addPass(createFunctionToLoopPassAdaptor(std::move(lpm),
use_memory_ssa));
@@ -2121,7 +2066,7 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
mpm.addPass(llvm::GlobalDCEPass());
mpm.addPass(llvm::ConstantMergePass());
# endif
#endif
break;
}
case 13: {
@@ -2132,11 +2077,11 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
{
llvm::FunctionPassManager fpm;
fpm.addPass(llvm::SimplifyCFGPass());
# if OSL_LLVM_VERSION < 160
#if OSL_LLVM_VERSION < 160
fpm.addPass(llvm::SROAPass());
# else
#else
fpm.addPass(llvm::SROAPass(llvm::SROAOptions::PreserveCFG));
# endif
#endif
fpm.addPass(llvm::EarlyCSEPass());
fpm.addPass(llvm::LowerExpectIntrinsicPass());
@@ -2153,11 +2098,11 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
fpm.addPass(llvm::InstCombinePass());
fpm.addPass(llvm::DCEPass());
# if OSL_LLVM_VERSION < 160
#if OSL_LLVM_VERSION < 160
fpm.addPass(llvm::SROAPass());
# else
#else
fpm.addPass(llvm::SROAPass(llvm::SROAOptions::PreserveCFG));
# endif
#endif
fpm.addPass(llvm::InstCombinePass());
fpm.addPass(llvm::SimplifyCFGPass());
fpm.addPass(llvm::PromotePass());
@@ -2201,10 +2146,10 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
llvm::createModuleToFunctionPassAdaptor(std::move(fpm)));
}
# if OSL_LLVM_VERSION < 150
#if OSL_LLVM_VERSION < 150
mpm.addPass(llvm::createModuleToPostOrderCGSCCPassAdaptor(
llvm::ArgumentPromotionPass()));
# endif
#endif
{
llvm::FunctionPassManager fpm;
@@ -2212,11 +2157,11 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
fpm.addPass(llvm::InstCombinePass());
fpm.addPass(llvm::JumpThreadingPass());
fpm.addPass(llvm::SimplifyCFGPass());
# if OSL_LLVM_VERSION < 160
#if OSL_LLVM_VERSION < 160
fpm.addPass(llvm::SROAPass());
# else
#else
fpm.addPass(llvm::SROAPass(llvm::SROAOptions::PreserveCFG));
# endif
#endif
fpm.addPass(llvm::InstCombinePass());
fpm.addPass(llvm::TailCallElimPass());
mpm.addPass(
@@ -2238,18 +2183,18 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
mpm.addPass(llvm::ModuleInlinerWrapperPass());
# if OSL_LLVM_VERSION < 150
#if OSL_LLVM_VERSION < 150
mpm.addPass(llvm::createModuleToPostOrderCGSCCPassAdaptor(
llvm::ArgumentPromotionPass()));
# endif
#endif
{
llvm::FunctionPassManager fpm;
# if OSL_LLVM_VERSION < 160
#if OSL_LLVM_VERSION < 160
fpm.addPass(llvm::SROAPass());
# else
#else
fpm.addPass(llvm::SROAPass(llvm::SROAOptions::PreserveCFG));
# endif
#endif
fpm.addPass(llvm::InstCombinePass());
fpm.addPass(llvm::SimplifyCFGPass());
fpm.addPass(llvm::ReassociatePass());
@@ -2258,7 +2203,11 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
llvm::LoopPassManager lpm;
const bool use_memory_ssa = true; // Needed by LICM
lpm.addPass(llvm::LoopRotatePass());
#if OSL_LLVM_VERSION >= 150
lpm.addPass(llvm::LICMPass(llvm::LICMOptions()));
#else
lpm.addPass(llvm::LICMPass());
#endif
lpm.addPass(llvm::SimpleLoopUnswitchPass(false));
fpm.addPass(createFunctionToLoopPassAdaptor(std::move(lpm),
use_memory_ssa));
@@ -2332,323 +2281,8 @@ LLVM_Util::setup_new_optimization_passes(int optlevel, bool target_host)
};
}
}
#endif
}
void
LLVM_Util::setup_legacy_optimization_passes(int optlevel, bool target_host)
{
#ifndef OSL_LLVM_NEW_PASS_MANAGER
# if OSL_LLVM_VERSION >= 160
# error "Legacy pass manager not supported in LLVM 16 and newer"
# endif
OSL_DEV_ONLY(std::cout << "setup_legacy_optimization_passes " << optlevel);
OSL_DASSERT(m_llvm_module_passes == NULL && m_llvm_func_passes == NULL);
// Construct the per-function passes and module-wide (interprocedural
// optimization) passes.
m_llvm_func_passes = new llvm::legacy::FunctionPassManager(module());
llvm::legacy::FunctionPassManager& fpm = (*m_llvm_func_passes);
m_llvm_module_passes = new llvm::legacy::PassManager;
llvm::legacy::PassManager& mpm = (*m_llvm_module_passes);
llvm::TargetMachine* target_machine = nullptr;
if (target_host) {
target_machine = execengine()->getTargetMachine();
llvm::Triple ModuleTriple(module()->getTargetTriple());
// Add an appropriate TargetLibraryInfo pass for the module's triple.
llvm::TargetLibraryInfoImpl TLII(ModuleTriple);
mpm.add(new llvm::TargetLibraryInfoWrapperPass(TLII));
mpm.add(createTargetTransformInfoWrapperPass(
target_machine ? target_machine->getTargetIRAnalysis()
: llvm::TargetIRAnalysis()));
fpm.add(createTargetTransformInfoWrapperPass(
target_machine ? target_machine->getTargetIRAnalysis()
: llvm::TargetIRAnalysis()));
}
// llvm_optimize 0-3 corresponds to the same set of optimizations
// as clang: -O0, -O1, -O2, -O3
// Tests on production shaders suggest the sweet spot between JIT time
// and runtime performance is O1.
//
// Optlevels 10, 11, 12, 13 explicitly create optimization passes. They
// are stripped down versions of clang's -O0, -O1, -O2, -O3. They try to
// provide similar results with improved optimization time by removing
// some expensive passes that were repeated many times and omitting
// other passes that are not applicable or not profitable. Useful for
// debugging, optlevel 10 adds next to no additional passes.
switch (optlevel) {
default: {
// For LLVM 3.0 and higher, llvm_optimize 1-3 means to use the
// same set of optimizations as clang -O1, -O2, -O3
llvm::PassManagerBuilder builder;
builder.OptLevel = optlevel;
// Time spent in JIT is considerably higher if there is no inliner specified
builder.Inliner = llvm::createFunctionInliningPass();
builder.DisableUnrollLoops = false;
builder.SLPVectorize = false;
builder.LoopVectorize = false;
if (target_machine)
target_machine->adjustPassManager(builder);
builder.populateFunctionPassManager(fpm);
builder.populateModulePassManager(mpm);
break;
}
case 10:
// truly add no optimizations
break;
case 11: {
// The least we would want to do
mpm.add(llvm::createFunctionInliningPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createGlobalDCEPass());
break;
}
case 12: {
# if 0 // PRETTY_GOOD_KEEP_AS_REF
mpm.add(llvm::createFunctionInliningPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createGlobalDCEPass());
mpm.add(llvm::createTypeBasedAAWrapperPass());
mpm.add(llvm::createBasicAAWrapperPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createSROAPass());
mpm.add(llvm::createEarlyCSEPass());
mpm.add(llvm::createReassociatePass());
mpm.add(llvm::createConstantPropagationPass());
mpm.add(llvm::createDeadCodeEliminationPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createPromoteMemoryToRegisterPass());
mpm.add(llvm::createAggressiveDCEPass());
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createDeadCodeEliminationPass());
mpm.add(llvm::createJumpThreadingPass());
mpm.add(llvm::createSROAPass());
mpm.add(llvm::createInstructionCombiningPass());
// Added
mpm.add(llvm::createDeadStoreEliminationPass());
mpm.add(llvm::createGlobalDCEPass());
mpm.add(llvm::createConstantMergePass());
# else
mpm.add(llvm::createFunctionInliningPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createGlobalDCEPass());
mpm.add(llvm::createTypeBasedAAWrapperPass());
mpm.add(llvm::createBasicAAWrapperPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createSROAPass());
mpm.add(llvm::createEarlyCSEPass());
// Eliminate and remove as much as possible up front
mpm.add(llvm::createReassociatePass());
# if OSL_LLVM_VERSION < 120
mpm.add(llvm::createConstantPropagationPass());
# endif
mpm.add(llvm::createDeadCodeEliminationPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createPromoteMemoryToRegisterPass());
mpm.add(llvm::createAggressiveDCEPass());
// mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createReassociatePass());
// TODO: investigate if the loop optimization passes rely on metadata from clang
// we might need to recreate that meta data in OSL's loop code to enable these passes
mpm.add(llvm::createLoopRotatePass());
mpm.add(llvm::createLICMPass());
# if OSL_LLVM_VERSION < 150
mpm.add(llvm::createLoopUnswitchPass(false));
# endif
// mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createIndVarSimplifyPass());
// Don't think we emitted any idioms that should be converted to a loop
// mpm.add(llvm::createLoopIdiomPass());
mpm.add(llvm::createLoopDeletionPass());
mpm.add(llvm::createLoopUnrollPass());
// GVN is expensive but should pay for itself in reducing JIT time
mpm.add(llvm::createGVNPass());
mpm.add(llvm::createSCCPPass());
// mpm.add(llvm::createInstructionCombiningPass());
// JumpThreading combo had a good improvement on JIT time
mpm.add(llvm::createJumpThreadingPass());
// optional, didn't seem to help more than it cost
// mpm.add(llvm::createCorrelatedValuePropagationPass());
mpm.add(llvm::createDeadStoreEliminationPass());
mpm.add(llvm::createAggressiveDCEPass());
mpm.add(llvm::createCFGSimplificationPass());
// Place late as possible to minimize #instrs it has to process
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createPromoteMemoryToRegisterPass());
mpm.add(llvm::createDeadCodeEliminationPass());
mpm.add(llvm::createGlobalDCEPass());
mpm.add(llvm::createConstantMergePass());
# endif
break;
}
case 13: {
mpm.add(llvm::createGlobalDCEPass());
mpm.add(llvm::createTypeBasedAAWrapperPass());
mpm.add(llvm::createBasicAAWrapperPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createSROAPass());
mpm.add(llvm::createEarlyCSEPass());
mpm.add(llvm::createLowerExpectIntrinsicPass());
mpm.add(llvm::createReassociatePass());
# if OSL_LLVM_VERSION < 120
mpm.add(llvm::createConstantPropagationPass());
# endif
mpm.add(llvm::createDeadCodeEliminationPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createPromoteMemoryToRegisterPass());
mpm.add(llvm::createAggressiveDCEPass());
// The InstructionCombining is much more expensive that all the other
// optimizations, should attempt to reduce the number of times it is
// executed, if at all
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createDeadCodeEliminationPass());
mpm.add(llvm::createSROAPass());
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createPromoteMemoryToRegisterPass());
mpm.add(llvm::createGlobalOptimizerPass());
mpm.add(llvm::createReassociatePass());
# if OSL_LLVM_VERSION < 120
mpm.add(llvm::createIPConstantPropagationPass());
# else
// createIPConstantPropagationPass disappeared with LLVM 12.
// Comments in their PR indicate that IPSCCP is better, but I don't
// know if that means such a pass should be *right here*. I leave it
// to others who use opt==13 to continue to curate this particular
// list of passes.
mpm.add(llvm::createIPSCCPPass());
# endif
mpm.add(llvm::createDeadArgEliminationPass());
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createPruneEHPass());
mpm.add(llvm::createPostOrderFunctionAttrsLegacyPass());
mpm.add(llvm::createReversePostOrderFunctionAttrsPass());
mpm.add(llvm::createFunctionInliningPass());
# if OSL_LLVM_VERSION < 120
mpm.add(llvm::createConstantPropagationPass());
# endif
mpm.add(llvm::createDeadCodeEliminationPass());
mpm.add(llvm::createCFGSimplificationPass());
# if OSL_LLVM_VERSION < 150
mpm.add(llvm::createArgumentPromotionPass());
# endif
mpm.add(llvm::createAggressiveDCEPass());
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createJumpThreadingPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createSROAPass());
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createTailCallEliminationPass());
mpm.add(llvm::createFunctionInliningPass());
# if OSL_LLVM_VERSION < 120
mpm.add(llvm::createConstantPropagationPass());
# endif
mpm.add(llvm::createIPSCCPPass());
mpm.add(llvm::createDeadArgEliminationPass());
mpm.add(llvm::createAggressiveDCEPass());
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createFunctionInliningPass());
# if OSL_LLVM_VERSION < 150
mpm.add(llvm::createArgumentPromotionPass());
# endif
mpm.add(llvm::createSROAPass());
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createReassociatePass());
mpm.add(llvm::createLoopRotatePass());
mpm.add(llvm::createLICMPass());
# if OSL_LLVM_VERSION < 150
mpm.add(llvm::createLoopUnswitchPass(false));
# endif
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createIndVarSimplifyPass());
mpm.add(llvm::createLoopIdiomPass());
mpm.add(llvm::createLoopDeletionPass());
mpm.add(llvm::createLoopUnrollPass());
mpm.add(llvm::createGVNPass());
mpm.add(llvm::createMemCpyOptPass());
mpm.add(llvm::createSCCPPass());
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createJumpThreadingPass());
mpm.add(llvm::createCorrelatedValuePropagationPass());
mpm.add(llvm::createDeadStoreEliminationPass());
mpm.add(llvm::createAggressiveDCEPass());
mpm.add(llvm::createCFGSimplificationPass());
mpm.add(llvm::createInstructionCombiningPass());
mpm.add(llvm::createFunctionInliningPass());
mpm.add(llvm::createAggressiveDCEPass());
mpm.add(llvm::createStripDeadPrototypesPass());
mpm.add(llvm::createGlobalDCEPass());
mpm.add(llvm::createConstantMergePass());
mpm.add(llvm::createVerifierPass());
break;
}
}; // switch(optlevel)
// Add some extra passes if they are needed
if (target_host) {
if (!m_supports_llvm_bit_masks_natively) {
switch (m_vector_width) {
case 16:
// MUST BE THE FINAL PASS!
mpm.add(
new LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks<16>());
break;
case 8:
// MUST BE THE FINAL PASS!
mpm.add(
new LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks<8>());
break;
case 4:
// MUST BE THE FINAL PASS!
mpm.add(
new LegacyPreventBitMasksFromBeingLiveinsToBasicBlocks<4>());
break;
default:
std::cout << "m_vector_width = " << m_vector_width << "\n";
OSL_ASSERT(0 && "unsupported bit mask width");
};
}
}
#endif
}
void
@@ -2662,19 +2296,9 @@ LLVM_Util::do_optimize(std::string* out_err)
return;
#endif
#ifdef OSL_LLVM_NEW_PASS_MANAGER
// New pass manager
m_new_pass_manager->module_pass_manager.run(
*m_llvm_module, m_new_pass_manager->module_analysis_manager);
#else
// Legacy pass manager
m_llvm_func_passes->doInitialization();
for (auto&& I : m_llvm_module->functions())
if (!I.isDeclaration())
m_llvm_func_passes->run(I);
m_llvm_func_passes->doFinalization();
m_llvm_module_passes->run(*m_llvm_module);
#endif
}
@@ -4492,13 +4116,6 @@ llvm::Value*
LLVM_Util::op_load(llvm::Type* type, llvm::Value* ptr,
const std::string& llname)
{
#ifndef OSL_LLVM_OPAQUE_POINTERS
OSL_PRAGMA_WARNING_PUSH
OSL_GCC_PRAGMA(GCC diagnostic ignored "-Wdeprecated-declarations")
OSL_ASSERT(type
== ptr->getType()->getScalarType()->getPointerElementType());
OSL_PRAGMA_WARNING_POP
#endif
return builder().CreateLoad(type, ptr, llname);
}
@@ -6168,13 +5785,6 @@ llvm::Value*
LLVM_Util::GEP(llvm::Type* type, llvm::Value* ptr, llvm::Value* elem,
const std::string& llname)
{
#ifndef OSL_LLVM_OPAQUE_POINTERS
OSL_PRAGMA_WARNING_PUSH
OSL_GCC_PRAGMA(GCC diagnostic ignored "-Wdeprecated-declarations")
OSL_ASSERT(type
== ptr->getType()->getScalarType()->getPointerElementType());
OSL_PRAGMA_WARNING_POP
#endif
return builder().CreateGEP(type, ptr, elem, llname);
}
@@ -6184,13 +5794,6 @@ llvm::Value*
LLVM_Util::GEP(llvm::Type* type, llvm::Value* ptr, int elem,
const std::string& llname)
{
#ifndef OSL_LLVM_OPAQUE_POINTERS
OSL_PRAGMA_WARNING_PUSH
OSL_GCC_PRAGMA(GCC diagnostic ignored "-Wdeprecated-declarations")
OSL_ASSERT(type
== ptr->getType()->getScalarType()->getPointerElementType());
OSL_PRAGMA_WARNING_POP
#endif
return builder().CreateConstGEP1_32(type, ptr, elem, llname);
}
@@ -6200,13 +5803,6 @@ llvm::Value*
LLVM_Util::GEP(llvm::Type* type, llvm::Value* ptr, int elem1, int elem2,
const std::string& llname)
{
#ifndef OSL_LLVM_OPAQUE_POINTERS
OSL_PRAGMA_WARNING_PUSH
OSL_GCC_PRAGMA(GCC diagnostic ignored "-Wdeprecated-declarations")
OSL_ASSERT(type
== ptr->getType()->getScalarType()->getPointerElementType());
OSL_PRAGMA_WARNING_POP
#endif
return builder().CreateConstGEP2_32(type, ptr, elem1, elem2, llname);
}
@@ -6215,13 +5811,6 @@ llvm::Value*
LLVM_Util::GEP(llvm::Type* type, llvm::Value* ptr, int elem1, int elem2,
int elem3, const std::string& llname)
{
#ifndef OSL_LLVM_OPAQUE_POINTERS
OSL_PRAGMA_WARNING_PUSH
OSL_GCC_PRAGMA(GCC diagnostic ignored "-Wdeprecated-declarations")
OSL_ASSERT(type
== ptr->getType()->getScalarType()->getPointerElementType());
OSL_PRAGMA_WARNING_POP
#endif
llvm::Value* elements[3] = { constant(elem1), constant(elem2),
constant(elem3) };
return builder().CreateGEP(type, ptr, toArrayRef(elements), llname);