add -debug-info-for-profiling frontend flag

This achieves the same as clang's `-fdebug-info-for-profiling`, which
emits DWARF discriminators to aid in narrowing-down which basic block
corresponds to a particular instruction address. This is particularly
useful for sampling-based profiling.

rdar://135443278

Author: kavon

include/swift/AST/IRGenOptions.h

@@ -504,6 +504,9 @@ class IRGenOptions {
   /// or the empty string.
   std::string UseSampleProfile = "";
 
+  /// Controls whether DWARF discriminators are added to the IR.
+  unsigned DebugInfoForProfiling : 1;
+
   /// List of backend command-line options for -embed-bitcode.
   std::vector<uint8_t> CmdArgs;
 
@@ -588,7 +591,8 @@ class IRGenOptions {
         ColocateTypeDescriptors(true), UseRelativeProtocolWitnessTables(false),
         UseFragileResilientProtocolWitnesses(false), EnableHotColdSplit(false),
         EmitAsyncFramePushPopMetadata(false), EmitYieldOnce2AsYieldOnce(true),
-        AsyncFramePointerAll(false), UseProfilingMarkerThunks(false), CmdArgs(),
+        AsyncFramePointerAll(false), UseProfilingMarkerThunks(false),
+        DebugInfoForProfiling(false), CmdArgs(),
         SanitizeCoverage(llvm::SanitizerCoverageOptions()),
         TypeInfoFilter(TypeInfoDumpFilter::All),
         PlatformCCallingConvention(llvm::CallingConv::C), UseCASBackend(false),

include/swift/Option/Options.td

@@ -1472,6 +1472,10 @@ def profile_generate : Flag<["-"], "profile-generate">,
   Flags<[FrontendOption, NoInteractiveOption]>,
   HelpText<"Generate instrumented code to collect execution counts">;
 
+def debug_info_for_profiling : Flag<["-"], "debug-info-for-profiling">,
+  Flags<[FrontendOption, NoInteractiveOption]>,
+  HelpText<"Emit extra debug info to make sample profile more accurate">;
+
 def profile_use : CommaJoined<["-"], "profile-use=">,
   Flags<[FrontendOption, NoInteractiveOption, ArgumentIsPath]>,
   MetaVarName<"<profdata>">,

lib/Frontend/CompilerInvocation.cpp

@@ -3180,6 +3180,8 @@ static bool ParseIRGenArgs(IRGenOptions &Opts, ArgList &Args,
   const Arg *ProfileSampleUse = Args.getLastArg(OPT_profile_sample_use);
   Opts.UseSampleProfile = ProfileSampleUse ? ProfileSampleUse->getValue() : "";
 
+  Opts.DebugInfoForProfiling |= Args.hasArg(OPT_debug_info_for_profiling);
+
   Opts.PrintInlineTree |= Args.hasArg(OPT_print_llvm_inline_tree);
   // Always producing all outputs when caching is enabled.
   Opts.AlwaysCompile |= Args.hasArg(OPT_always_compile_output_files) ||

lib/IRGen/IRGen.cpp

@@ -211,7 +211,21 @@ static void populatePGOOptions(std::optional<PGOOptions> &Out,
       /*FS=*/ llvm::vfs::getRealFileSystem(), // TODO: is this fine?
       /*Action=*/ PGOOptions::SampleUse,
       /*CSPGOAction=*/ PGOOptions::NoCSAction,
-      /*DebugInfoForProfiling=*/ false
+      /*DebugInfoForProfiling=*/ Opts.DebugInfoForProfiling
+    );
+    return;
+  }
+
+  if (Opts.DebugInfoForProfiling) {
+    Out = PGOOptions(
+        /*ProfileFile=*/ "",
+        /*CSProfileGenFile=*/ "",
+        /*ProfileRemappingFile=*/ "",
+        /*MemoryProfile=*/ "",
+        /*FS=*/ nullptr,
+        /*Action=*/ PGOOptions::NoAction,
+        /*CSPGOAction=*/ PGOOptions::NoCSAction,
+        /*DebugInfoForProfiling=*/ true
     );
     return;
   }
@@ -239,14 +253,15 @@ void swift::performLLVMOptimizations(const IRGenOptions &Opts,
     PTO.LoopVectorization = true;
     PTO.SLPVectorization = true;
     PTO.MergeFunctions = true;
-    populatePGOOptions(PGOOpt, Opts);
     // Splitting trades code size to enhance memory locality, avoid in -Osize.
     DoHotColdSplit = Opts.EnableHotColdSplit && !Opts.optimizeForSize();
     level = llvm::OptimizationLevel::Os;
   } else {
     level = llvm::OptimizationLevel::O0;
   }
 
+  populatePGOOptions(PGOOpt, Opts);
+
   LoopAnalysisManager LAM;
   FunctionAnalysisManager FAM;
   CGSCCAnalysisManager CGAM;

lib/IRGen/IRGenDebugInfo.cpp

@@ -839,6 +839,7 @@ class IRGenDebugInfoImpl : public IRGenDebugInfo {
                               TheCU->getProducer(), true, StringRef(), 0,
                               RemappedASTFile, llvm::DICompileUnit::FullDebug,
                               Signature);
+        // NOTE: not setting DebugInfoForProfiling here
         DIB.finalize();
       }
     }
@@ -2504,7 +2505,7 @@ IRGenDebugInfoImpl::IRGenDebugInfoImpl(const IRGenOptions &Opts,
           ? llvm::DICompileUnit::FullDebug
           : llvm::DICompileUnit::LineTablesOnly,
       /* DWOId */ 0, /* SplitDebugInlining */ true,
-      /* DebugInfoForProfiling */ false,
+      /* DebugInfoForProfiling */ Opts.DebugInfoForProfiling,
       llvm::DICompileUnit::DebugNameTableKind::Default,
       /* RangesBaseAddress */ false, DebugPrefixMap.remapPath(Sysroot), SDK);
 

lib/IRGen/IRGenSIL.cpp

@@ -1939,9 +1939,7 @@ IRGenSILFunction::IRGenSILFunction(IRGenModule &IGM, SILFunction *f)
     // LLVM doesn't have an attribute for -O
   }
 
-  const bool optimizing = IGM.IRGen.Opts.shouldOptimize() &&
-                          !IGM.IRGen.Opts.DisableLLVMOptzns;
-  if (optimizing && !IGM.IRGen.Opts.UseSampleProfile.empty()) {
+  if (!IGM.IRGen.Opts.UseSampleProfile.empty()) {
     // This attribute helps in LTO situations: https://reviews.llvm.org/D79959
     CurFn->addFnAttr("use-sample-profile");
   }

test/Profiler/samplepgo.swift

@@ -19,13 +19,7 @@
 
 // ----------------------------------------
 // Now, test cases where there there should not be any profile metadata,
-// such as when no optimizations are requested, or no data is provided
-
-// RUN: %target-swift-frontend %t/program.swift -module-name test -emit-ir \
-// RUN:            -profile-sample-use=%t/profile.txt -o %t/no-optimizations.ll
-
-// RUN: %FileCheck --check-prefix CHECK-NODATA %s < %t/no-optimizations.ll
-
+// such as when no data is provided
 
 // RUN: %target-swift-frontend %t/program.swift -module-name test -emit-ir \
 // RUN:                        -O -o %t/no-data.ll