[CGData] Clang Options (#90304)

This adds new Clang flags to support codegen (CG) data:
- `-fcodegen-data-generate{=path}`: This flag passes
`-codegen-data-generate` as a boolean to the LLVM backend, causing the
raw CG data to be emitted into a custom section. Currently, for LLD
MachO only, it also passes `--codegen-data-generate-path=<path>` so that
the indexed CG data file can be automatically produced at link time. For
linkers that do not yet support this feature, `llvm-cgdata` can be used
manually to merge this CG data in object files.
- `-fcodegen-data-use{=path}`: This flag passes
`-codegen-data-use-path=<path>` to the LLVM backend, enabling the use of
specified CG data to optimistically outline functions.
 - The default `<path>` is set to `default.cgdata` when not specified.
 
This depends on #108733.
This is a patch for
https://discourse.llvm.org/t/rfc-enhanced-machine-outliner-part-2-thinlto-nolto/78753.

Author: kyulee-com

clang/docs/UsersManual.rst

@@ -2410,6 +2410,39 @@ are listed below.
   link-time optimizations like whole program inter-procedural basic block
   reordering.
 
+.. option:: -fcodegen-data-generate[=<path>]
+
+  Emit the raw codegen (CG) data into custom sections in the object file.
+  Currently, this option also combines the raw CG data from the object files
+  into an indexed CG data file specified by the <path>, for LLD MachO only.
+  When the <path> is not specified, `default.cgdata` is created.
+  The CG data file combines all the outlining instances that occurred locally
+  in each object file.
+
+  .. code-block:: console
+
+    $ clang -fuse-ld=lld -Oz -fcodegen-data-generate code.cc
+
+  For linkers that do not yet support this feature, `llvm-cgdata` can be used
+  manually to merge this CG data in object files.
+
+  .. code-block:: console
+
+    $ clang -c -fuse-ld=lld -Oz -fcodegen-data-generate code.cc
+    $ llvm-cgdata --merge -o default.cgdata code.o
+
+.. option:: -fcodegen-data-use[=<path>]
+
+  Read the codegen data from the specified path to more effectively outline
+  functions across compilation units. When the <path> is not specified,
+  `default.cgdata` is used. This option can create many identically outlined
+  functions that can be optimized by the conventional linker’s identical code
+  folding (ICF).
+
+  .. code-block:: console
+
+    $ clang -fuse-ld=lld -Oz -Wl,--icf=safe -fcodegen-data-use code.cc
+
 Profile Guided Optimization
 ---------------------------
 

clang/include/clang/Driver/Options.td

@@ -1894,6 +1894,18 @@ def fprofile_selected_function_group :
   Visibility<[ClangOption, CC1Option]>, MetaVarName<"<i>">,
   HelpText<"Partition functions into N groups using -fprofile-function-groups and select only functions in group i to be instrumented. The valid range is 0 to N-1 inclusive">,
   MarshallingInfoInt<CodeGenOpts<"ProfileSelectedFunctionGroup">>;
+def fcodegen_data_generate_EQ : Joined<["-"], "fcodegen-data-generate=">,
+    Group<f_Group>, Visibility<[ClangOption, CLOption]>, MetaVarName<"<path>">,
+    HelpText<"Emit codegen data into the object file. LLD for MachO (currently) merges them into the specified <path>.">;
+def fcodegen_data_generate : Flag<["-"], "fcodegen-data-generate">,
+    Group<f_Group>, Visibility<[ClangOption, CLOption]>, Alias<fcodegen_data_generate_EQ>, AliasArgs<["default.cgdata"]>,
+    HelpText<"Emit codegen data into the object file. LLD for MachO (currently) merges them into default.cgdata.">;
+def fcodegen_data_use_EQ : Joined<["-"], "fcodegen-data-use=">,
+    Group<f_Group>, Visibility<[ClangOption, CLOption]>, MetaVarName<"<path>">,
+    HelpText<"Use codegen data read from the specified <path>.">;
+def fcodegen_data_use : Flag<["-"], "fcodegen-data-use">,
+    Group<f_Group>, Visibility<[ClangOption, CLOption]>, Alias<fcodegen_data_use_EQ>, AliasArgs<["default.cgdata"]>,
+    HelpText<"Use codegen data read from default.cgdata to optimize the binary">;
 def fswift_async_fp_EQ : Joined<["-"], "fswift-async-fp=">,
     Group<f_Group>,
     Visibility<[ClangOption, CC1Option, CC1AsOption, CLOption]>,

clang/lib/Driver/ToolChains/CommonArgs.cpp

@@ -2753,6 +2753,25 @@ void tools::addMachineOutlinerArgs(const Driver &D,
       addArg(Twine("-enable-machine-outliner=never"));
     }
   }
+
+  auto *CodeGenDataGenArg =
+      Args.getLastArg(options::OPT_fcodegen_data_generate_EQ);
+  auto *CodeGenDataUseArg = Args.getLastArg(options::OPT_fcodegen_data_use_EQ);
+
+  // We only allow one of them to be specified.
+  if (CodeGenDataGenArg && CodeGenDataUseArg)
+    D.Diag(diag::err_drv_argument_not_allowed_with)
+        << CodeGenDataGenArg->getAsString(Args)
+        << CodeGenDataUseArg->getAsString(Args);
+
+  // For codegen data gen, the output file is passed to the linker
+  // while a boolean flag is passed to the LLVM backend.
+  if (CodeGenDataGenArg)
+    addArg(Twine("-codegen-data-generate"));
+
+  // For codegen data use, the input file is passed to the LLVM backend.
+  if (CodeGenDataUseArg)
+    addArg(Twine("-codegen-data-use-path=") + CodeGenDataUseArg->getValue());
 }
 
 void tools::addOpenMPDeviceRTL(const Driver &D,

clang/lib/Driver/ToolChains/Darwin.cpp

@@ -476,6 +476,13 @@ void darwin::Linker::AddLinkArgs(Compilation &C, const ArgList &Args,
         llvm::sys::path::append(Path, "default.profdata");
       CmdArgs.push_back(Args.MakeArgString(Twine("--cs-profile-path=") + Path));
     }
+
+    auto *CodeGenDataGenArg =
+        Args.getLastArg(options::OPT_fcodegen_data_generate_EQ);
+    if (CodeGenDataGenArg)
+      CmdArgs.push_back(
+          Args.MakeArgString(Twine("--codegen-data-generate-path=") +
+                             CodeGenDataGenArg->getValue()));
   }
 }
 
@@ -633,6 +640,32 @@ void darwin::Linker::ConstructJob(Compilation &C, const JobAction &JA,
   CmdArgs.push_back("-mllvm");
   CmdArgs.push_back("-enable-linkonceodr-outlining");
 
+  // Propagate codegen data flags to the linker for the LLVM backend.
+  auto *CodeGenDataGenArg =
+      Args.getLastArg(options::OPT_fcodegen_data_generate_EQ);
+  auto *CodeGenDataUseArg = Args.getLastArg(options::OPT_fcodegen_data_use_EQ);
+
+  // We only allow one of them to be specified.
+  const Driver &D = getToolChain().getDriver();
+  if (CodeGenDataGenArg && CodeGenDataUseArg)
+    D.Diag(diag::err_drv_argument_not_allowed_with)
+        << CodeGenDataGenArg->getAsString(Args)
+        << CodeGenDataUseArg->getAsString(Args);
+
+  // For codegen data gen, the output file is passed to the linker
+  // while a boolean flag is passed to the LLVM backend.
+  if (CodeGenDataGenArg) {
+    CmdArgs.push_back("-mllvm");
+    CmdArgs.push_back("-codegen-data-generate");
+  }
+
+  // For codegen data use, the input file is passed to the LLVM backend.
+  if (CodeGenDataUseArg) {
+    CmdArgs.push_back("-mllvm");
+    CmdArgs.push_back(Args.MakeArgString(Twine("-codegen-data-use-path=") +
+                                         CodeGenDataUseArg->getValue()));
+  }
+
   // Setup statistics file output.
   SmallString<128> StatsFile =
       getStatsFileName(Args, Output, Inputs[0], getToolChain().getDriver());

clang/test/Driver/codegen-data.c

@@ -0,0 +1,38 @@
+// Verify only one of codegen-data flag is passed.
+// RUN: not %clang -### -S --target=aarch64-linux-gnu -fcodegen-data-generate -fcodegen-data-use %s 2>&1 | FileCheck %s --check-prefix=CONFLICT
+// RUN: not %clang -### -S --target=arm64-apple-darwin  -fcodegen-data-generate -fcodegen-data-use %s 2>&1 | FileCheck %s --check-prefix=CONFLICT
+// CONFLICT: error: invalid argument '-fcodegen-data-generate' not allowed with '-fcodegen-data-use'
+
+// Verify the codegen-data-generate (boolean) flag is passed to LLVM
+// RUN: %clang -### -S --target=aarch64-linux-gnu -fcodegen-data-generate %s  2>&1| FileCheck %s --check-prefix=GENERATE
+// RUN: %clang -### -S --target=arm64-apple-darwin -fcodegen-data-generate %s 2>&1| FileCheck %s --check-prefix=GENERATE
+// GENERATE: "-mllvm" "-codegen-data-generate"
+
+// Verify the codegen-data-use-path flag (with a default value) is passed to LLVM.
+// RUN: %clang -### -S --target=aarch64-linux-gnu -fcodegen-data-use %s 2>&1| FileCheck %s --check-prefix=USE
+// RUN: %clang -### -S --target=arm64-apple-darwin -fcodegen-data-use %s 2>&1| FileCheck %s --check-prefix=USE
+// RUN: %clang -### -S --target=aarch64-linux-gnu -fcodegen-data-use=file %s 2>&1 | FileCheck %s --check-prefix=USE-FILE
+// RUN: %clang -### -S --target=arm64-apple-darwin -fcodegen-data-use=file %s 2>&1 | FileCheck %s --check-prefix=USE-FILE
+// USE: "-mllvm" "-codegen-data-use-path=default.cgdata"
+// USE-FILE: "-mllvm" "-codegen-data-use-path=file"
+
+// Verify the codegen-data-generate (boolean) flag with a LTO.
+// RUN: %clang -### -flto --target=aarch64-linux-gnu -fcodegen-data-generate %s 2>&1 | FileCheck %s --check-prefix=GENERATE-LTO
+// GENERATE-LTO: {{ld(.exe)?"}}
+// GENERATE-LTO-SAME: "-plugin-opt=-codegen-data-generate"
+// RUN: %clang -### -flto --target=arm64-apple-darwin -fcodegen-data-generate %s 2>&1 | FileCheck %s --check-prefix=GENERATE-LTO-DARWIN
+// GENERATE-LTO-DARWIN: {{ld(.exe)?"}}
+// GENERATE-LTO-DARWIN-SAME: "-mllvm" "-codegen-data-generate"
+
+// Verify the codegen-data-use-path flag with a LTO is passed to LLVM.
+// RUN: %clang -### -flto=thin --target=aarch64-linux-gnu -fcodegen-data-use %s 2>&1 | FileCheck %s --check-prefix=USE-LTO
+// USE-LTO: {{ld(.exe)?"}}
+// USE-LTO-SAME: "-plugin-opt=-codegen-data-use-path=default.cgdata"
+// RUN: %clang -### -flto=thin --target=arm64-apple-darwin -fcodegen-data-use %s 2>&1 | FileCheck %s --check-prefix=USE-LTO-DARWIN
+// USE-LTO-DARWIN: {{ld(.exe)?"}}
+// USE-LTO-DARWIN-SAME: "-mllvm" "-codegen-data-use-path=default.cgdata"
+
+// For now, LLD MachO supports for generating the codegen data at link time.
+// RUN: %clang -### -fuse-ld=lld -B%S/Inputs/lld --target=arm64-apple-darwin -fcodegen-data-generate %s 2>&1 | FileCheck %s --check-prefix=GENERATE-LLD-DARWIN
+// GENERATE-LLD-DARWIN: {{ld(.exe)?"}}
+// GENERATE-LLD-DARWIN-SAME: "--codegen-data-generate-path=default.cgdata"