Clang PGO and LLVM PGO commandline interface

Clang Frontend
1

Hello all,
A while ago, I implemented frontend-based PGO in LDC (very similar to Clang’s) and now I’d like to add IR-based PGO. I want the commandline interface to be the same (or very similar) to Clang’s, but I am having trouble finding the current state of Clang’s PGO commandline flags (and their meaning related to Clang/LLVM-IR PGO).

Can you give me a pointer to where I can find this, or perhaps give me a brief overview of the current flags and how to select whether to use frontend-based or IR-based PGO ?

Thanks a lot,
Johan

2

https://clang.llvm.org/docs/UsersManual.html#profile-guided-optimization is the documentation for Clang’s PGO. It doesn’t explicitly state which flags map to frontend and IR-based instrumentation.

  • -fprofile-generate and -fprofile-use are for IR-based instrumentation and profile use.
  • -fprofile-instr-generate and -fprofile-instr-use are for frontend-based instrumentation and profile use.
  • -fprofile-sample-use is to use sparse profiles from hardware-counter-based profiles.
3

-fprofile-use and -fprofile-instr-use can actually be used interchangeably for front-end and IR based PGO. The difference is that -fprofile-use=<> also takes a directory as the argument.

David

4

Hi David and Pirama, (first sent off-list, but resending to list)
Thank you for your quick responses. I’ve now successfully added IR-based PGO to LDC; it’s almost trivial. The only thing needed (on LLVM side) was

// For generating pgo info

builder.EnablePGOInstrGen = true;
builder.PGOInstrGen = <outputfilename>;


// For using pgo profile
builder.PGOInstrUse = <inputfilename>;

where builder is the PassManagerBuilder.

This works well on Linux and macOS, however for a Windows target triple, it does not work. The “instrprof” pass fails in that case.
I have attached the failing testcase, that fails with “opt -instrprof crash.ll”:


> opt crash.ll -instrprof

comdat global value has private linkage
[1 x i64]* @__profc_luke
comdat global value has private linkage
[1 x i64]* @__profc_cold
comdat global value has private linkage
[1 x i64]* @__profc_main
comdat global value has private linkage
[2 x i64]* @__profc__Dmain
LLVM ERROR: Broken module found, compilation aborted!

I’ve obtained the IR by compiling the D code with -print-before-all, and crash.ll is the IR before the instrprof pass.

Now, before I file this as a bug, I want to make sure that I didn’t do anything wrong in setting up IR-based PGO. Are there any other steps needed? FYI: we don’t use a special pipeline setup, just the default what LLVM gives us: https://github.com/ldc-developers/ldc/blob/2645ae8645c70d4dc01988f85482d1b85f82fc14/gen/optimizer.cpp#L235-L331 .
(somehow “opt -instr-prof-gen” does work for Windows…)

As Pirama suggested, I tested with Clang and there it does work with a Windows triple. (https://godbolt.org/g/h5CMJx)

Thanks a lot,
Kind regards,
Johan

crash.ll (6.16 KB)