[Clang][Driver][SamplePGO] Enable -fsample-profile-use-profi for SamplePGO

[nilanjana87]

• Enable -fsample-profile-use-profi flag by default for SamplePGO
• Add -fno_sample_profile_use_profi flag for opting out

Discourse conversation: https://discourse.llvm.org/t/turn-on-fsample-profile-use-profi-by-default/86508/2

[llvmbot]

@llvm/pr-subscribers-pgo
@llvm/pr-subscribers-clang-driver

@llvm/pr-subscribers-clang

Author: Nilanjana Basu (nilanjana87)

Changes

• Enable -fsample-profile-use-profi flag by default for SamplePGO
• Add -fno_sample_profile_use_profi flag for opting out

Discourse conversation: https://discourse.llvm.org/t/turn-on-fsample-profile-use-profi-by-default/86508/2

---

Full diff: https://github.com/llvm/llvm-project/pull/145957.diff

3 Files Affected:

• (modified) clang/include/clang/Driver/Options.td (+2)
• (modified) clang/lib/Driver/ToolChains/Clang.cpp (+2-1)
• (modified) clang/test/Driver/pgo-sample-use-profi.c (+17-2)

diff --git a/clang/include/clang/Driver/Options.td b/clang/include/clang/Driver/Options.td
index 0822df2640d23..ef6bcfe296a80 100644
--- a/clang/include/clang/Driver/Options.td
+++ b/clang/include/clang/Driver/Options.td
@@ -1698,6 +1698,8 @@ def fsample_profile_use_profi : Flag<["-"], "fsample-profile-use-profi">,
                basic block counts to branch probabilities to fix them by extended
                and re-engineered classic MCMF (min-cost max-flow) approach.}]>;
 def fno_profile_sample_accurate : Flag<["-"], "fno-profile-sample-accurate">, Group<f_Group>;
+def fno_sample_profile_use_profi : Flag<["-"], "fno-sample-profile-use-profi">,
+                                   Group<f_Group>;
 def fno_auto_profile : Flag<["-"], "fno-auto-profile">, Group<f_Group>,
     Alias<fno_profile_sample_use>;
 def fauto_profile_EQ : Joined<["-"], "fauto-profile=">,
diff --git a/clang/lib/Driver/ToolChains/Clang.cpp b/clang/lib/Driver/ToolChains/Clang.cpp
index 8a18865b899d2..7898c0a6a899c 100644
--- a/clang/lib/Driver/ToolChains/Clang.cpp
+++ b/clang/lib/Driver/ToolChains/Clang.cpp
@@ -6283,7 +6283,8 @@ void Clang::ConstructJob(Compilation &C, const JobAction &JA,
   Args.AddLastArg(CmdArgs, options::OPT_fclang_abi_compat_EQ);
 
   if (getLastProfileSampleUseArg(Args) &&
-      Args.hasArg(options::OPT_fsample_profile_use_profi)) {
+      Args.hasFlag(options::OPT_fsample_profile_use_profi,
+                   options::OPT_fno_sample_profile_use_profi, true)) {
     CmdArgs.push_back("-mllvm");
     CmdArgs.push_back("-sample-profile-use-profi");
   }
diff --git a/clang/test/Driver/pgo-sample-use-profi.c b/clang/test/Driver/pgo-sample-use-profi.c
index 0370d947ce590..9d7c7af983508 100644
--- a/clang/test/Driver/pgo-sample-use-profi.c
+++ b/clang/test/Driver/pgo-sample-use-profi.c
@@ -1,4 +1,19 @@
-/// Test if profi flat is enabled in frontend as user-facing feature.
-// RUN: %clang --target=x86_64 -c -fsample-profile-use-profi -fprofile-sample-use=/dev/null -### %s 2>&1 | FileCheck %s
+/// Ensure that profi flag is enabled by default in frontend for SamplePGO
+
+// Target specific checks:
+// RUN: %clang --target=x86_64 -c -fprofile-sample-use=/dev/null -### %s 2>&1 | FileCheck %s
+// RUN: %clang --target=AArch64 -c -fprofile-sample-use=/dev/null -### %s 2>&1 | FileCheck %s
+
+// Target agnostic checks:
+// RUN: %clang -c -fprofile-sample-use=/dev/null -### %s 2>&1 | FileCheck %s
+// RUN: %clang -c -fsample-profile-use-profi -fprofile-sample-use=/dev/null -### %s 2>&1 | FileCheck %s
+// RUN: %clang -c -fno-sample-profile-use-profi -fsample-profile-use-profi -fprofile-sample-use=/dev/null -### %s 2>&1 | FileCheck %s
+
+// Cases where profi flag is disabled:
+// RUN: %clang -c -### %s 2>&1 | FileCheck %s --check-prefixes=CHECK-NO-PROFI
+// RUN: %clang -c -fno-sample-profile-use-profi -fprofile-sample-use=/dev/null -### %s 2>&1 | FileCheck %s --check-prefixes=CHECK-NO-PROFI
+// RUN: %clang -c -fsample-profile-use-profi -fno-sample-profile-use-profi -fprofile-sample-use=/dev/null -### %s 2>&1 | FileCheck %s --check-prefixes=CHECK-NO-PROFI
+
 
 // CHECK: "-mllvm" "-sample-profile-use-profi"
+// CHECK-NO-PROFI-NOT: "-sample-profile-use-profi"

[snehasish]

We evaluated profi internally (at Google) last year. Our configuration uses AutoFDO with flow-sensitive discriminators (FS-AFDO). We found slight regressions with this configuration and didn't enable it for our workloads. With FS-AFDO we see performance improvements of ~1%. Can you share your configuration and performance numbers?