PGO: switch to using compiler-rt profile library. (#2527)

Also makes sure that the PGO runtime is available when doing the PGO tests.

CMakeLists.txt

@@ -765,12 +765,14 @@ if (LDC_INSTALL_LLVM_RUNTIME_LIBS)
 
     if(APPLE)
         copy_compilerrt_lib("darwin/libclang_rt.asan_osx_dynamic.dylib" "libldc_rt.asan_osx_dynamic.dylib" TRUE)
+        copy_compilerrt_lib("darwin/libclang_rt.profile_osx.a" "libldc_rt.profile_osx.a" FALSE)
         if(NOT (LDC_LLVM_VER LESS 600))
             copy_compilerrt_lib("darwin/libclang_rt.fuzzer_osx.a" "libldc_rt.fuzzer_osx.a" FALSE)
             copy_compilerrt_lib("darwin/libclang_rt.xray_osx.a"   "libldc_rt.xray_osx.a"   FALSE)
         endif()
     elseif(UNIX)
         copy_compilerrt_lib("linux/libclang_rt.asan-x86_64.a" "libldc_rt.asan-x86_64.a" FALSE)
+        copy_compilerrt_lib("linux/libclang_rt.profile-x86_64.a" "libldc_rt.profile-x86_64.a" FALSE)
         if(NOT (LDC_LLVM_VER LESS 500))
             copy_compilerrt_lib("linux/libclang_rt.xray-x86_64.a"   "libldc_rt.xray-x86_64.a"   FALSE)
         endif()

driver/linker-gcc.cpp

@@ -53,6 +53,7 @@ private:
   virtual void addASanLinkFlags(const llvm::Triple &triple);
   virtual void addFuzzLinkFlags(const llvm::Triple &triple);
   virtual void addCppStdlibLinkFlags(const llvm::Triple &triple);
+  virtual void addProfileRuntimeLinkFlags(const llvm::Triple &triple);
   virtual void addXRayLinkFlags(const llvm::Triple &triple);
 
   virtual void addLinker();
@@ -368,6 +369,33 @@ void ArgsBuilder::addCppStdlibLinkFlags(const llvm::Triple &triple) {
   }
 }
 
+// Adds all required link flags for PGO.
+void ArgsBuilder::addProfileRuntimeLinkFlags(const llvm::Triple &triple) {
+  std::string searchPaths[] = {
+    getFullCompilerRTLibPath(triple, "libldc_rt.profile"),
+    getFullCompilerRTLibPath(triple, "libclang_rt.profile"),
+    getFullClangCompilerRTLibPath(triple, "libclang_rt.profile"),
+  };
+
+#if LDC_LLVM_VER >= 308
+  if (global.params.targetTriple->isOSLinux()) {
+    // For Linux, explicitly define __llvm_profile_runtime as undefined
+    // symbol, so that the initialization part of profile-rt is linked in.
+    addLdFlag("-u", llvm::getInstrProfRuntimeHookVarName());
+  }
+#endif
+
+  for (const auto &filepath : searchPaths) {
+    IF_LOG Logger::println("Searching profile runtime: %s", filepath.c_str());
+
+    if (llvm::sys::fs::exists(filepath)) {
+      IF_LOG Logger::println("Found, linking with %s", filepath.c_str());
+      args.push_back(filepath);
+      return;
+    }
+  }
+}
+
 void ArgsBuilder::addSanitizers(const llvm::Triple &triple) {
   if (opts::isSanitizerEnabled(opts::AddressSanitizer)) {
     addASanLinkFlags(triple);
@@ -401,14 +429,7 @@ void ArgsBuilder::build(llvm::StringRef outputPath,
 
   // Link with profile-rt library when generating an instrumented binary.
   if (opts::isInstrumentingForPGO()) {
-#if LDC_LLVM_VER >= 308
+    addProfileRuntimeLinkFlags(*global.params.targetTriple);
-    if (global.params.targetTriple->isOSLinux()) {
-      // For Linux, explicitly define __llvm_profile_runtime as undefined
-      // symbol, so that the initialization part of profile-rt is linked in.
-      addLdFlag("-u", llvm::getInstrProfRuntimeHookVarName());
-    }
-#endif
-    args.push_back("-lldc-profile-rt");
   }
 
   if (opts::enableDynamicCompile) {

runtime/CMakeLists.txt

@@ -629,9 +629,10 @@ endmacro()
 macro(build_all_runtime_variants d_flags c_flags ld_flags path_suffix outlist_targets)
     build_runtime_variants("${d_flags}" "${c_flags}" "${ld_flags}" "${path_suffix}" ${outlist_targets})
 
-    # static profile-rt
+    # static profile-rt (currently only needed on Windows)
-    build_profile_runtime("${c_flags}" "${ld_flags}" "" "${path_suffix}" ${outlist_targets})
+    if("${TARGET_SYSTEM}" MATCHES "Windows")
-    get_target_suffix("" "${path_suffix}" target_suffix)
+        build_profile_runtime("${c_flags}" "${ld_flags}" "" "${path_suffix}" ${outlist_targets})
+    endif()
 endmacro()
 
 
@@ -666,7 +667,7 @@ if(MULTILIB AND "${TARGET_SYSTEM}" MATCHES "APPLE")
 
     # KLUDGE: Cannot use `$<TARGET_LINKER_FILE:target>` in custom command.
     # Set up the list of generated libs (without 'lib' prefix) to be merged manually.
-    set(libs_to_install ldc-profile-rt.a)
+    set(libs_to_install)
     if(NOT ${BUILD_SHARED_LIBS} STREQUAL "ON")
         list(APPEND libs_to_install druntime-ldc.a druntime-ldc-debug.a
                                     phobos2-ldc.a  phobos2-ldc-debug.a

tests/PGO/allstatementtypes.d

@@ -1,5 +1,7 @@
 // Test instrumentation of all 'simple' AST statement types.
 
+// REQUIRES: PGO_RT
+
 // Disable autogenerated boundschecking to declutter the generated code.
 // See boundscheck.d for boundschecking instrumenation tests.
 

tests/PGO/branching_switch.d

@@ -1,5 +1,7 @@
 // Test instrumentation of switch with non-constant case expression.
 
+// REQUIRES: PGO_RT
+
 // RUN: %ldc -c -output-ll -fprofile-instr-generate -of=%t.ll %s && FileCheck %s --check-prefix=PROFGEN < %t.ll
 
 // RUN: %ldc -fprofile-instr-generate=%t.profraw -run %s  \

tests/PGO/break.d

@@ -1,5 +1,7 @@
 // Test calculation of execution counts with loops with break-to-label and continue-to-label.
 
+// REQUIRES: PGO_RT
+
 // RUN: %ldc -c -output-ll -fprofile-instr-generate -of=%t.ll %s && FileCheck %s --check-prefix=PROFGEN < %t.ll
 
 // RUN: %ldc -fprofile-instr-generate=%t.profraw -run %s  \

tests/PGO/exceptions.d

@@ -1,5 +1,7 @@
 // Test instrumentation of statement types involving exceptions.
 
+// REQUIRES: PGO_RT
+
 // XFAIL: Windows
 
 // RUN: %ldc -c -output-ll -fprofile-instr-generate -of=%t.ll %s && FileCheck %s --check-prefix=PROFGEN < %t.ll

tests/PGO/functions.d

@@ -3,6 +3,8 @@
 // The tests should not test function name mangling, therefore the functionname
 // matching strings contain regexp wildcards.
 
+// REQUIRES: PGO_RT
+
 // RUN: %ldc -c -output-ll -fprofile-instr-generate -of=%t.ll %s && FileCheck %s --check-prefix=PROFGEN < %t.ll
 
 // RUN: %ldc -fprofile-instr-generate=%t.profraw -run %s  \

tests/PGO/hash.d

@@ -2,6 +2,8 @@
 // applied when code has changed.
 // The code changes are simulated by version(.) blocks.
 
+// REQUIRES: PGO_RT
+
 // RUN: %ldc -d-version=ProfData -fprofile-instr-generate=%t.profraw -run %s  \
 // RUN:   &&  %profdata merge %t.profraw -o %t.profdata \
 // RUN:   &&  %ldc -d-version=ProfData -c -output-ll -of=%t2.ll -fprofile-instr-use=%t.profdata %s \

tests/PGO/indirect_calls.d

@@ -1,5 +1,7 @@
 // Test instrumentation of indirect calls
 
+// REQUIRES: PGO_RT
+
 // REQUIRES: atleast_llvm309
 
 // RUN: %ldc -c -output-ll -fprofile-instr-generate -of=%t.ll %s && FileCheck %s --check-prefix=PROFGEN < %t.ll

tests/PGO/irbased_indirect_calls.d

@@ -1,5 +1,7 @@
 // Test instrumentation of indirect calls
 
+// REQUIRES: PGO_RT
+
 // REQUIRES: atleast_llvm309
 
 // There is an LLVM bug, this test currently errors during LLVM codegen for Windows.

tests/PGO/lit.local.cfg

@@ -0,0 +1,10 @@
+import os
+import platform
+import re
+
+# Add "PGO_RT" feature if the runtime library is available
+for file in os.listdir(config.ldc2_lib_dir):
+    m = re.match('.*profile.*', file)
+    if m is not None:
+        config.available_features.add('PGO_RT')
+        continue

tests/PGO/max_function_count.d

@@ -1,5 +1,6 @@
 // Test that maximum function counts are set correctly (LLVM >= 3.8)
 
+// REQUIRES: PGO_RT
 // REQUIRES: atleast_llvm308
 // For LLVM > 3.8, a summary is emitted, see summary.d
 

tests/PGO/profile_rt_calls.d

@@ -1,5 +1,7 @@
 // Tests runtime profile-rt access.
 
+// REQUIRES: PGO_RT
+
 // RUN: %ldc -fprofile-instr-generate=%t.profraw -run %s
 
 import ldc.profile;

tests/PGO/reset_counters.d

@@ -1,5 +1,7 @@
 // Tests (the availability of) the runtime lib function to reset all profile counters.
 
+// REQUIRES: PGO_RT
+
 // RUN: %ldc -fprofile-instr-generate=%t.profraw -run %s  \
 // RUN:   &&  %profdata merge %t.profraw -o %t.profdata \
 // RUN:   &&  %ldc -c -output-ll -of=%t2.ll -fprofile-instr-use=%t.profdata %s \

tests/PGO/singleobj.d

@@ -1,5 +1,7 @@
 // Tests simulaneous compilation into singleobj
 
+// REQUIRES: PGO_RT
+
 // RUN: %ldc -fprofile-instr-generate=%t.profraw %S/inputs/singleobj_input.d -c -of=%t%obj \
 // RUN:   &&  %ldc -fprofile-instr-generate=%t.profraw %t%obj -run %s \
 // RUN:   &&  %profdata merge %t.profraw -o %t.profdata \

tests/PGO/summary.d

@@ -1,5 +1,6 @@
 // Test that maximum function counts are set correctly (LLVM >= 3.8)
 // REQUIRES: atleast_llvm309
+// REQUIRES: PGO_RT
 
 // RUN: %ldc -fprofile-instr-generate=%t.profraw -run %s  \
 // RUN:   &&  %profdata merge %t.profraw -o %t.profdata \

tests/PGO/thinlto_pgo.d

@@ -3,6 +3,7 @@
 
 // REQUIRES: atleast_llvm400
 // REQUIRES: LTO
+// REQUIRES: PGO_RT
 
 // RUN: %ldc -fprofile-instr-generate=%t.profraw -run %s  \
 // RUN:   &&  %profdata merge %t.profraw -o %t.profdata \