Add ldc-profgen tool (ldc-developers#4536)

Also fixes cross build action: asserts yes/no must be same as host package (never asserts for android cross build)

.github/actions/3-build-cross/action.yml

@@ -33,8 +33,20 @@ runs:
         else
           tag=CI
         fi
+
+        if [[ '${{ inputs.os }}' == android ]]; then
+          assertsSuffix=""
+        else
+          # Use assertions for untagged builds. Must be the same as for the host LLVM package, because
+          # llvm-config of host package will be used for the cross build configuration.
+          assertsSuffix="-withAsserts"
+          if [[ '${{ github.ref }}' = refs/tags/* ]]; then
+            assertsSuffix=""
+          fi
+        fi
+
         curl -fL --retry 3 --max-time 300 -o llvm-cross.tar.xz \
-          https://github.com/ldc-developers/llvm-project/releases/download/$tag/llvm-$version-${{ inputs.os }}-${{ inputs.arch }}.tar.xz
+          https://github.com/ldc-developers/llvm-project/releases/download/$tag/llvm-$version-${{ inputs.os }}-${{ inputs.arch }}$assertsSuffix.tar.xz
         mkdir llvm-cross
         tar -xf llvm-cross.tar.xz --strip 1 -C llvm-cross
         rm llvm-cross.tar.xz
@@ -145,4 +157,4 @@ runs:
           ${{ inputs.cmake_flags }}
           ${{ inputs.with_pgo == 'true' && '-DDFLAGS_LDC=-fprofile-use=../pgo-ldc/merged.profdata' || '' }}
           ${{ env.CROSS_CMAKE_FLAGS }}
-        build_targets: ldc2 ldmd2 ldc-build-runtime ldc-build-plugin ldc-profdata ldc-prune-cache timetrace2txt
+        build_targets: ldc2 ldmd2 ldc-build-runtime ldc-build-plugin ldc-profdata ldc-profgen ldc-prune-cache timetrace2txt

.github/actions/5-install/action.yml

@@ -21,7 +21,7 @@ runs:
           cd ..
         else
           mkdir -p install/bin
-          cp build-cross/bin/{ldc2,ldmd2,ldc-build-runtime,ldc-profdata,ldc-prune-cache,timetrace2txt} install/bin/
+          cp build-cross/bin/{ldc2,ldmd2,ldc-build-runtime,ldc-profdata,ldc-profgen,ldc-prune-cache,timetrace2txt} install/bin/
           cp build-cross/bin/ldc-build-plugin install/bin/ || true
           cp -R build-cross-libs/lib install/
           cp build-cross/lib/{libldc_rt.*,libLTO-ldc.dylib,LLVMgold-ldc.so} install/lib/ || true

CHANGELOG.md

@@ -3,7 +3,8 @@
 #### Big news
 - Frontend, druntime and Phobos are at version [2.106.0](https://dlang.org/changelog/2.106.0.html). (#4522)
 - New command-line options `-fno-{exceptions,moduleinfo,rtti}` to selectively enable some `-betterC` effects. (#4522)
-- New command-line option `-fprofile-sample-use` for using sample-based profile data for optimization. Functionality and usage is identical to Clang's option with same name. (#4531).
+- New command-line option `-fprofile-sample-use` for using sample-based profile data for optimization. Functionality and usage is identical to Clang's option with same name. (#4531)
+- New `ldc-profgen` tool for sample-based PGO, a copy of LLVM's [llvm-profgen](https://llvm.org/docs/CommandGuide/llvm-profgen.html). (#4536)
 
 #### Platform support
 

CMakeLists.txt

@@ -40,7 +40,7 @@ find_package(LLVM 11.0 REQUIRED
     instcombine ipo instrumentation irreader libdriver linker lto mc
     mcdisassembler mcparser objcarcopts object option profiledata scalaropts
     selectiondag support tablegen target transformutils vectorize
-    windowsdriver windowsmanifest ${EXTRA_LLVM_MODULES})
+    windowsdriver windowsmanifest symbolize ${EXTRA_LLVM_MODULES})
 math(EXPR LDC_LLVM_VER ${LLVM_VERSION_MAJOR}*100+${LLVM_VERSION_MINOR})
 message(STATUS "Using LLVM Version ${LLVM_VERSION_MAJOR}.${LLVM_VERSION_MINOR}")
 # Remove LLVMTableGen library from list of libraries

tests/CMakeLists.txt

@@ -1,5 +1,6 @@
 set( LDC2_BIN          ${PROJECT_BINARY_DIR}/bin/${LDC_EXE} )
 set( LDCPROFDATA_BIN   ${PROJECT_BINARY_DIR}/bin/ldc-profdata )
+set( LDCPROFGEN_BIN    ${PROJECT_BINARY_DIR}/bin/ldc-profgen )
 set( LDCPRUNECACHE_BIN ${PROJECT_BINARY_DIR}/bin/${LDCPRUNECACHE_EXE} )
 set( LDCBUILDPLUGIN_BIN ${PROJECT_BINARY_DIR}/bin/${LDC_BUILD_PLUGIN_EXE} )
 set( TIMETRACE2TXT_BIN ${PROJECT_BINARY_DIR}/bin/${TIMETRACE2TXT_EXE} )

tests/lit.site.cfg.in

@@ -16,6 +16,7 @@ OFF = False
 ## Auto-initialized variables by cmake:
 config.ldc2_bin            = "@LDC2_BIN@"
 config.ldcprofdata_bin     = "@LDCPROFDATA_BIN@"
+config.ldcprofgen_bin      = "@LDCPROFGEN_BIN@"
 config.ldcprunecache_bin   = "@LDCPRUNECACHE_BIN@"
 config.ldcbuildplugin_bin  = "@LDCBUILDPLUGIN_BIN@"
 config.timetrace2txt_bin   = "@TIMETRACE2TXT_BIN@"
@@ -156,6 +157,7 @@ config.environment['PATH'] = path
 config.substitutions.append( ('%ldc', config.ldc2_bin) )
 config.substitutions.append( ('%gnu_make', config.gnu_make_bin) )
 config.substitutions.append( ('%profdata', config.ldcprofdata_bin) )
+config.substitutions.append( ('%profgen', config.ldcprofgen_bin) )
 config.substitutions.append( ('%prunecache', config.ldcprunecache_bin) )
 config.substitutions.append( ('%buildplugin', config.ldcbuildplugin_bin + " --ldcSrcDir=" + config.ldc2_source_dir ) )
 config.substitutions.append( ('%timetrace2txt', config.timetrace2txt_bin) )

tools/CMakeLists.txt

@@ -40,6 +40,32 @@ else()
     message(WARNING "ldc-profdata source (${LDCPROFDATA_SRC}) not found")
 endif()
 
+#############################################################################
+# Build ldc-profgen utility that generates a profile data file from given perf script
+# data files for sample-based profile guided optimization (-fprofile-sample-use).
+# https://llvm.org/docs/CommandGuide/llvm-profgen.html
+# The source in ldc-profgen/ldc-profgen-xx.x is an unmodified copy of llvm's llvm-profgen source dir.
+if(LDC_LLVM_VER GREATER_EQUAL 1400)
+    macro(add_llvm_tool llvm_name)
+        string(REPLACE "llvm-" "ldc-" ldc_name ${llvm_name})
+        message(STATUS "Configuring ${ldc_name} build target")
+        add_executable(${ldc_name} ${ARGN})
+        set_target_properties(
+            ${ldc_name} PROPERTIES
+            RUNTIME_OUTPUT_DIRECTORY ${PROJECT_BINARY_DIR}/bin
+            COMPILE_FLAGS "${LLVM_CXXFLAGS} ${LDC_CXXFLAGS}"
+            LINK_FLAGS "${SANITIZE_LDFLAGS}"
+        )
+        target_link_libraries(${ldc_name} ${LLVM_LIBRARIES} ${CMAKE_DL_LIBS} ${LLVM_LDFLAGS})
+        install(TARGETS ${ldc_name} DESTINATION ${CMAKE_INSTALL_PREFIX}/bin)
+    endmacro()
+    if (IS_DIRECTORY ${CMAKE_CURRENT_SOURCE_DIR}/ldc-profgen/ldc-profgen-${LLVM_VERSION_MAJOR}.${LLVM_VERSION_MINOR})
+        add_subdirectory(ldc-profgen/ldc-profgen-${LLVM_VERSION_MAJOR}.${LLVM_VERSION_MINOR})
+    else()
+        message(WARNING "ldc-profgen source not found (${CMAKE_CURRENT_SOURCE_DIR}/ldc-profgen/ldc-profgen-${LLVM_VERSION_MAJOR}.${LLVM_VERSION_MINOR})")
+    endif()
+endif()
+
 #############################################################################
 # Build timetrace2txt
 set(TIMETRACE2TXT_EXE timetrace2txt)

tools/README.md

@@ -7,4 +7,6 @@ The `/tools` directory contains user tools that accompany LDC and that should be
 
 `ldc-profdata` converts raw profiling data to a profile data format that can be used by LDC. The source is copied from LLVM (`llvm-profdata`), and is versioned for each LLVM version that we support because the version has to match exactly with LDC's LLVM version.
 
+`ldc-profgen` converts perf sample profiling data to a profile data format that can be used by LDC. The source is copied from LLVM (`llvm-profgen`), and is versioned for each LLVM version that we support because the version has to match exactly with LDC's LLVM version.
+
 `timetrace2txt` converts the .timetrace output of `--ftime-trace` (which is in [Chromium's trace event JSON format](https://www.chromium.org/developers/how-tos/trace-event-profiling-tool/)) to a text file that is easier for humans to read.

tools/ldc-profgen/ldc-profgen-14.0/CMakeLists.txt

@@ -0,0 +1,23 @@
+
+set(LLVM_LINK_COMPONENTS
+  AllTargetsDescs
+  AllTargetsDisassemblers
+  AllTargetsInfos
+  DebugInfoDWARF
+  Core
+  MC
+  IPO
+  MCDisassembler
+  Object
+  ProfileData
+  Support
+  Symbolize
+  )
+
+add_llvm_tool(llvm-profgen
+  llvm-profgen.cpp
+  PerfReader.cpp
+  CSPreInliner.cpp
+  ProfiledBinary.cpp
+  ProfileGenerator.cpp
+  )