Build libclang on Windows: Difference between revisions

From Qt Wiki
Jump to navigation Jump to search
(Performance)
Line 55: Line 55:
We have measurements comparison between different llvm builds. The most important there are parse and re-parse times.
We have measurements comparison between different llvm builds. The most important there are parse and re-parse times.


This is the comparison for sample file that was done for our modified llvm 3.9 builds:
This is the comparison for sample file that was done for our llvm builds (all compilers are 64-bit):
{| class="wikitable"
{| class="wikitable"
|action
|action
|LLVM_msvc2015_64
|LLVM 5.0 mingw PGO
|LLVM_mingw64
|LLVM 5.0 intel compiler PGO
|LLVM_msvc2017_64
|LLVM 3.9 mingw PGO
|LLVM_mingw64_profile
|LLVM 3.9 msvc2015
|LLVM_msvc17_profile
|LLVM 3.9 msvc2017 PGO
|LLVM_clang
|LLVM 3.9 clang PGO
|LLVM_clang_profile
|LLVM 3.9 mingw
|LLVM 3.9 clang
|LLVM 3.9 msvc2017
|-
|-
|Parsing
|Parsing
|1.306
|1.5426
|1.2874
|2.7482
|2.7482
|1.473
|1.5015
|1.5797
|1.5797
|1.7486
|1.8487
|1.8487
|1.2874
|1.473
|1.7486
|1.5015
|-
|-
|Reparsing
|Reparsing
|1.9258
|2.2318
|2.3362
|2.9665
|2.9665
|2.5521
|2.6727
|3.183
|3.183
|2.8431
|3.0255
|3.0255
|2.3362
|2.5521
|2.8431
|2.6727
|}
|}
That means that currently the fastest build is mingw with applied profile-guided optimization.
That means that currently the fastest build is mingw with applied profile-guided optimization.

Revision as of 13:22, 12 March 2018

Configuring builds

MSVC (2015 and 2017) and MinGW

Can be configured with Qt Creator. It's enough to manually set CMAKE_INSTALL_PREFIX, CMAKE_BUILD_TYPE and LLVM_ENABLE_RTTI

For MinGW there are extra flags though to link libraries into libclang.dll statically (CMAKE_C_FLAGS and CMAKE_CXX_FLAGS): -static-libgcc -static-libstdc++ -static

Clang

Configurable with cmake: You need msvc command line and append PATH with bin directory of clang (set PATH=%PATH%;path\to\clang\bin)

cmake -G <Generator ("NMake Makefiles JOM" or "NMake Makefiles")> -DCMAKE_BUILD_TYPE=<Release|Debug> -DLLVM_ENABLE_RTTI=1 -DCMAKE_C_COMPILER=clang-cl -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_INSTALL_PREFIX=path\to\install -DCMAKE_C_FLAGS="-fms-compatibility-version=19.0" -DCMAKE_CXX_FLAGS="-fms-compatibility-version=19.0" -DMSVC=1 -DCMAKE_CL_64=<1|0> path\to\llvm

Intel Compiler

Currently llvm requires some tweaks to compile with icl (mostly inline namespaces from windows headers and binary ops with enums). After you apply tweaks for llvm and windows headers it compiles but performance is similar to msvc version so there's not much profit in icl builds currently.

Build

In creator and command line you can just use "jom install" or use default step from Qt Creator

Profile-guided builds extra configuration

MSVC

* First build round:
  CMAKE_C_FLAGS and CMAKE_CXX_FLAGS: /GL
  CMAKE_EXE_LINKER_FLAGS and CMAKE_SHARED_LINKER_FLAGS: /LTCG and /GENPROFILE:PGD=pgd\files\path
* Training
  Just run parsing/completion
* Second build round:
  CMAKE_C_FLAGS and CMAKE_CXX_FLAGS: /GL
  CMAKE_EXE_LINKER_FLAGS and CMAKE_SHARED_LINKER_FLAGS: /LTCG and /USEPROFILE:PGD=same\pgd\files\path

MinGW

* First build round:
  CMAKE_C_FLAGS and CMAKE_CXX_FLAGS: -fprofile-generate=path\to\profile\files
* Training
  Just run parsing/completion
* Second build round:
  CMAKE_C_FLAGS and CMAKE_CXX_FLAGS: -fprofile-use=path\to\profile\files

Clang

Clang requires compiler-rt to build llvm with -fprofile-instr-generate flag Follow the https://compiler-rt.llvm.org/ to build it. You need additional configuration though: -DCMAKE_INSTALL_PREFIX=<path\to\llvm\install>\lib\clang\5.0.0 and a proper -DCMAKE_BUILD_TYPE

* First build round:
  cmake -G "NMake Makefiles JOM" -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_RTTI=1 -DCMAKE_C_COMPILER=clang-cl -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_INSTALL_PREFIX=path\to\install -DCMAKE_C_FLAGS="-fms-compatibility-version=19.0 -fprofile-instr-generate=path\llvm-%p.profraw" -DCMAKE_CXX_FLAGS="-fms-compatibility-version=19.0 -fprofile-instr-generate=path\llvm-%p.profraw" -DMSVC=1 -DCMAKE_CL_64=<1|0> -DCMAKE_EXE_LINKER_FLAGS="<path\to\llvm\install>\lib\clang\5.0.0\lib\windows\clang_rt.profile-x86_64.lib /FORCE:MULTIPLE" path\to\llvm
* Training
* Merging train data
  llvm-profdata merge -output= llvm.profdata llvm-*.profraw
* Second build round:
  cmake -G "NMake Makefiles JOM" -DCMAKE_BUILD_TYPE=Release -DLLVM_ENABLE_RTTI=1 -DCMAKE_C_COMPILER=clang-cl -DCMAKE_CXX_COMPILER=clang-cl -DCMAKE_INSTALL_PREFIX=path\to\install -DCMAKE_C_FLAGS="-fms-compatibility-version=19.0 -fprofile-instr-use=path\llvm.prodata" -DCMAKE_CXX_FLAGS="-fms-compatibility-version=19.0 -fprofile-instr-use=path\llvm.prodata" -DMSVC=1 -DCMAKE_CL_64=<1|0> path\to\llvm

Performance

We have measurements comparison between different llvm builds. The most important there are parse and re-parse times.

This is the comparison for sample file that was done for our llvm builds (all compilers are 64-bit):

action LLVM 5.0 mingw PGO LLVM 5.0 intel compiler PGO LLVM 3.9 mingw PGO LLVM 3.9 msvc2015 LLVM 3.9 msvc2017 PGO LLVM 3.9 clang PGO LLVM 3.9 mingw LLVM 3.9 clang LLVM 3.9 msvc2017
Parsing 1.306 1.5426 1.2874 2.7482 1.473 1.5015 1.5797 1.7486 1.8487
Reparsing 1.9258 2.2318 2.3362 2.9665 2.5521 2.6727 3.183 2.8431 3.0255

That means that currently the fastest build is mingw with applied profile-guided optimization.