https://github.com/w2yehia updated
https://github.com/llvm/llvm-project/pull/124353
>From abef90fe8f46431a5fb8b7fe717c9fb65eb30266 Mon Sep 17 00:00:00 2001
From: Wael Yehia <wye...@ca.ibm.com>
Date: Thu, 23 Jan 2025 00:03:15 +0000
Subject: [PATCH 1/2] [PGO] Add a clang option -fprofile-continuous that
enables PGO continuous mode
---
clang/docs/UsersManual.rst | 8 ++++
clang/include/clang/Basic/CodeGenOptions.def | 1 +
clang/include/clang/Driver/Options.td | 5 +++
clang/lib/CodeGen/BackendUtil.cpp | 42 +++++++++++---------
clang/lib/Driver/ToolChains/Clang.cpp | 29 ++++++++++++++
clang/test/CodeGen/profile-continuous.c | 16 ++++++++
6 files changed, 83 insertions(+), 18 deletions(-)
create mode 100644 clang/test/CodeGen/profile-continuous.c
diff --git a/clang/docs/UsersManual.rst b/clang/docs/UsersManual.rst
index 260e84910c6f78..1e509906733997 100644
--- a/clang/docs/UsersManual.rst
+++ b/clang/docs/UsersManual.rst
@@ -3033,6 +3033,14 @@ indexed format, regardeless whether it is produced by
frontend or the IR pass.
overhead. ``prefer-atomic`` will be transformed to ``atomic`` when supported
by the target, or ``single`` otherwise.
+.. option:: -fprofile-continuous
+
+ Enables continuous PGO mode where profile counter updates are continuously
+ synced to a file. This option sets any neccessary modifiers (currently
``%c``)
+ in the default profile filename and passes any necessary flags to the
+ middle-end to support this mode. Value profiling is not supported in
+ continuous mode.
+
.. option:: -ftemporal-profile
Enables the temporal profiling extension for IRPGO to improve startup time by
diff --git a/clang/include/clang/Basic/CodeGenOptions.def
b/clang/include/clang/Basic/CodeGenOptions.def
index 0f4ed13d5f3d8c..bbaf8b183222e9 100644
--- a/clang/include/clang/Basic/CodeGenOptions.def
+++ b/clang/include/clang/Basic/CodeGenOptions.def
@@ -221,6 +221,7 @@ AFFECTING_VALUE_CODEGENOPT(OptimizationLevel, 2, 0) ///<
The -O[0-3] option spec
AFFECTING_VALUE_CODEGENOPT(OptimizeSize, 2, 0) ///< If -Os (==1) or -Oz (==2)
is specified.
CODEGENOPT(AtomicProfileUpdate , 1, 0) ///< Set -fprofile-update=atomic
+CODEGENOPT(ContinuousProfileSync, 1, 0) ///< Enable continuous PGO mode
/// Choose profile instrumenation kind or no instrumentation.
ENUM_CODEGENOPT(ProfileInstr, ProfileInstrKind, 2, ProfileNone)
/// Choose profile kind for PGO use compilation.
diff --git a/clang/include/clang/Driver/Options.td
b/clang/include/clang/Driver/Options.td
index 2721c1b5d8dc55..5a7e64d5b5a96f 100644
--- a/clang/include/clang/Driver/Options.td
+++ b/clang/include/clang/Driver/Options.td
@@ -1886,6 +1886,11 @@ def fprofile_update_EQ : Joined<["-"],
"fprofile-update=">,
Values<"atomic,prefer-atomic,single">,
MetaVarName<"<method>">, HelpText<"Set update method of profile counters">,
MarshallingInfoFlag<CodeGenOpts<"AtomicProfileUpdate">>;
+def fprofile_continuous : Flag<["-"], "fprofile-continuous">,
+ Group<f_Group>, Visibility<[ClangOption, CC1Option]>,
+ HelpText<"Enable Continuous PGO mode">,
+ MarshallingInfoFlag<CodeGenOpts<"ContinuousProfileSync">>;
+
defm pseudo_probe_for_profiling : BoolFOption<"pseudo-probe-for-profiling",
CodeGenOpts<"PseudoProbeForProfiling">, DefaultFalse,
PosFlag<SetTrue, [], [ClangOption], "Emit">,
diff --git a/clang/lib/CodeGen/BackendUtil.cpp
b/clang/lib/CodeGen/BackendUtil.cpp
index 3951ad01497cca..afafa8af585c71 100644
--- a/clang/lib/CodeGen/BackendUtil.cpp
+++ b/clang/lib/CodeGen/BackendUtil.cpp
@@ -133,6 +133,16 @@ std::string getDefaultProfileGenName() {
: "default_%m.profraw";
}
+// Path and name of file used for profile generation
+std::string getProfileGenName(const CodeGenOptions &CodeGenOpts) {
+ std::string FileName = CodeGenOpts.InstrProfileOutput.empty()
+ ? getDefaultProfileGenName()
+ : CodeGenOpts.InstrProfileOutput;
+ if (CodeGenOpts.ContinuousProfileSync)
+ FileName = "%c" + FileName;
+ return FileName;
+}
+
class EmitAssemblyHelper {
CompilerInstance &CI;
DiagnosticsEngine &Diags;
@@ -550,7 +560,9 @@ getInstrProfOptions(const CodeGenOptions &CodeGenOpts,
return std::nullopt;
InstrProfOptions Options;
Options.NoRedZone = CodeGenOpts.DisableRedZone;
- Options.InstrProfileOutput = CodeGenOpts.InstrProfileOutput;
+ Options.InstrProfileOutput = CodeGenOpts.ContinuousProfileSync
+ ? ("%c" + CodeGenOpts.InstrProfileOutput)
+ : CodeGenOpts.InstrProfileOutput;
Options.Atomic = CodeGenOpts.AtomicProfileUpdate;
return Options;
}
@@ -811,13 +823,12 @@ void EmitAssemblyHelper::RunOptimizationPipeline(
if (CodeGenOpts.hasProfileIRInstr())
// -fprofile-generate.
- PGOOpt = PGOOptions(
- CodeGenOpts.InstrProfileOutput.empty() ? getDefaultProfileGenName()
- :
CodeGenOpts.InstrProfileOutput,
- "", "", CodeGenOpts.MemoryProfileUsePath, nullptr, PGOOptions::IRInstr,
- PGOOptions::NoCSAction, ClPGOColdFuncAttr,
- CodeGenOpts.DebugInfoForProfiling,
- /*PseudoProbeForProfiling=*/false, CodeGenOpts.AtomicProfileUpdate);
+ PGOOpt = PGOOptions(getProfileGenName(CodeGenOpts), "", "",
+ CodeGenOpts.MemoryProfileUsePath, nullptr,
+ PGOOptions::IRInstr, PGOOptions::NoCSAction,
+ ClPGOColdFuncAttr, CodeGenOpts.DebugInfoForProfiling,
+ /*PseudoProbeForProfiling=*/false,
+ CodeGenOpts.AtomicProfileUpdate);
else if (CodeGenOpts.hasProfileIRUse()) {
// -fprofile-use.
auto CSAction = CodeGenOpts.hasProfileCSIRUse() ? PGOOptions::CSIRUse
@@ -861,18 +872,13 @@ void EmitAssemblyHelper::RunOptimizationPipeline(
PGOOpt->Action != PGOOptions::SampleUse &&
"Cannot run CSProfileGen pass with ProfileGen or SampleUse "
" pass");
- PGOOpt->CSProfileGenFile = CodeGenOpts.InstrProfileOutput.empty()
- ? getDefaultProfileGenName()
- : CodeGenOpts.InstrProfileOutput;
+ PGOOpt->CSProfileGenFile = getProfileGenName(CodeGenOpts);
PGOOpt->CSAction = PGOOptions::CSIRInstr;
} else
- PGOOpt = PGOOptions("",
- CodeGenOpts.InstrProfileOutput.empty()
- ? getDefaultProfileGenName()
- : CodeGenOpts.InstrProfileOutput,
- "", /*MemoryProfile=*/"", nullptr,
- PGOOptions::NoAction, PGOOptions::CSIRInstr,
- ClPGOColdFuncAttr,
CodeGenOpts.DebugInfoForProfiling);
+ PGOOpt = PGOOptions("", getProfileGenName(CodeGenOpts), "",
+ /*MemoryProfile=*/"", nullptr, PGOOptions::NoAction,
+ PGOOptions::CSIRInstr, ClPGOColdFuncAttr,
+ CodeGenOpts.DebugInfoForProfiling);
}
if (TM)
TM->setPGOOption(PGOOpt);
diff --git a/clang/lib/Driver/ToolChains/Clang.cpp
b/clang/lib/Driver/ToolChains/Clang.cpp
index 33f08cf28feca1..8aefa82951edbc 100644
--- a/clang/lib/Driver/ToolChains/Clang.cpp
+++ b/clang/lib/Driver/ToolChains/Clang.cpp
@@ -580,6 +580,7 @@ static void addPGOAndCoverageFlags(const ToolChain &TC,
Compilation &C,
const ArgList &Args, SanitizerArgs &SanArgs,
ArgStringList &CmdArgs) {
const Driver &D = TC.getDriver();
+ const llvm::Triple &T = TC.getTriple();
auto *PGOGenerateArg = Args.getLastArg(options::OPT_fprofile_generate,
options::OPT_fprofile_generate_EQ,
options::OPT_fno_profile_generate);
@@ -785,6 +786,34 @@ static void addPGOAndCoverageFlags(const ToolChain &TC,
Compilation &C,
D.Diag(diag::err_drv_unsupported_option_argument)
<< A->getSpelling() << Val;
}
+ if (const auto *A = Args.getLastArg(options::OPT_fprofile_continuous)) {
+ if (!PGOGenerateArg && !CSPGOGenerateArg && !ProfileGenerateArg)
+ D.Diag(clang::diag::err_drv_argument_only_allowed_with)
+ << A->getSpelling()
+ << "-fprofile-generate, -fprofile-instr-generate, or "
+ "-fcs-profile-generate";
+ else {
+ CmdArgs.push_back("-fprofile-continuous");
+ // Platforms that require a bias variable:
+ if (T.isOSFuchsia() || T.isOSBinFormatELF() || T.isOSAIX()) {
+ CmdArgs.push_back("-mllvm");
+ CmdArgs.push_back("-runtime-counter-relocation");
+ }
+ // -fprofile-instr-generate does not decide the profile file name in the
+ // FE, and so it does not define the filename symbol
+ // (__llvm_profile_filename). Instead, the runtime uses the name
+ // "default.profraw" for the profile file. When continuous mode is ON, we
+ // will create the filename symbol so that we can insert the "%c"
+ // modifier.
+ if (ProfileGenerateArg &&
+ (ProfileGenerateArg->getOption().matches(
+ options::OPT_fprofile_instr_generate) ||
+ (ProfileGenerateArg->getOption().matches(
+ options::OPT_fprofile_instr_generate_EQ) &&
+ strlen(ProfileGenerateArg->getValue()) == 0)))
+ CmdArgs.push_back("-fprofile-instrument-path=default.profraw");
+ }
+ }
int FunctionGroups = 1;
int SelectedFunctionGroup = 0;
diff --git a/clang/test/CodeGen/profile-continuous.c
b/clang/test/CodeGen/profile-continuous.c
new file mode 100644
index 00000000000000..1db31c2f162b27
--- /dev/null
+++ b/clang/test/CodeGen/profile-continuous.c
@@ -0,0 +1,16 @@
+// RUN: %clang %s -S -emit-llvm -fprofile-generate -fprofile-continuous -o - |
FileCheck %s --check-prefix=IRPGO
+// RUN: %clang %s -S -emit-llvm -fprofile-generate=mydir -fprofile-continuous
-o - | FileCheck %s --check-prefix=IRPGO_EQ
+// RUN: %clang %s -S -emit-llvm -fcs-profile-generate -fprofile-continuous -O
-o - | FileCheck %s --check-prefix=CSIRPGO
+// RUN: %clang %s -S -emit-llvm -fprofile-instr-generate -fprofile-continuous
-o - | FileCheck %s --check-prefix=CLANG_PGO
+// RUN: %clang %s -S -emit-llvm -fprofile-instr-generate= -fprofile-continuous
-o - | FileCheck %s --check-prefix=CLANG_PGO
+// RUN: %clang %s -S -emit-llvm -fprofile-instr-generate=foo.profraw
-fprofile-continuous -o - | FileCheck %s --check-prefix=CLANG_PGO_EQ
+
+// RUN: not %clang -### %s -fprofile-continuous -c 2>&1 | FileCheck %s
--check-prefix=ERROR
+
+// IRPGO: @__llvm_profile_filename = {{.*}} c"%cdefault_%m.profraw\00"
+// IRPGO_EQ: @__llvm_profile_filename = {{.*}} c"%cmydir/default_%m.profraw\00"
+// CSIRPGO: @__llvm_profile_filename = {{.*}} c"%cdefault_%m.profraw\00"
+// CLANG_PGO: @__llvm_profile_filename = {{.*}} c"%cdefault.profraw\00"
+// CLANG_PGO_EQ: @__llvm_profile_filename = {{.*}} c"%cfoo.profraw\00"
+// ERROR: clang: error: invalid argument '-fprofile-continuous' only allowed
with '-fprofile-generate, -fprofile-instr-generate, or -fcs-profile-generate'
+void foo(){}
>From 5d9db9a4c3dd33d26a5e8578807d81242ae476ce Mon Sep 17 00:00:00 2001
From: Wael Yehia <wye...@ca.ibm.com>
Date: Mon, 27 Jan 2025 17:18:37 +0000
Subject: [PATCH 2/2] Add example to the user documenation
---
clang/docs/UsersManual.rst | 10 ++++++++++
1 file changed, 10 insertions(+)
diff --git a/clang/docs/UsersManual.rst b/clang/docs/UsersManual.rst
index 1e509906733997..d9b29433468f93 100644
--- a/clang/docs/UsersManual.rst
+++ b/clang/docs/UsersManual.rst
@@ -3041,6 +3041,16 @@ indexed format, regardeless whether it is produced by
frontend or the IR pass.
middle-end to support this mode. Value profiling is not supported in
continuous mode.
+ .. code-block:: console
+
+ $ clang++ -O2 -fprofile-generate -fprofile-continuous code.cc -o code
+
+ Running `./code` will collect the profile and write it to the
+ `default_xxxx.profraw` file. However, if `./code` abruptly terminates or does
+ not call `exit()`, in continuous mode the profile collected up to the point
of
+ termination will be available in `default_xxxx.profraw` while in the
+ non-continuous mode, no profile file is generated.
+
.. option:: -ftemporal-profile
Enables the temporal profiling extension for IRPGO to improve startup time by
_______________________________________________
cfe-commits mailing list
cfe-commits@lists.llvm.org
https://lists.llvm.org/cgi-bin/mailman/listinfo/cfe-commits
