sunxfancy / IPRA-exp

IPRA Experiments

Clone the repo

This project contains submodules, please run:

git clone --recursive [e-mail]:sunxfancy/IPRA-exp.git

or after clone, run:

git submodule update --init --recursive

Install tools

Required tools: SingularityCE - installing it following the documentation: https://docs.sylabs.io/guides/3.11/user-guide/

Build Project

1. run make singularity/image to build the image, you may need the root privileges
2. run ./make to build the clang compiler, autofdo and some small tools
3. run ./make benchmarks/${bench}/${target} can build the specific target of one of the benchmarks
4. run ./make benchmarks/${bench}/${target}.bench can run the bench for this target
5. run ./make benchmarks/${bench}/${target}.regprof3 can run the register and spill code profiling for this target

E.g. Run the clang benchmark register and spill code profiling for thin-lto build with all features eanbled (Threshold for hot function = 3, Callsite Cold Ratio = 20):

./make benchmarks/clang/pgo-full-bfdoipra6.3-20.regprof3

Quickly Reproducing the Benchmarks

Run ./make benchmarks to run all benchmarks

Benchmark Output

There are two folder will be generated: build and tmp. In the build/benchmarks folder, you will see the benchmarks result:

1. <target>.bench the result of perf stat after running 5 times
2. <target>.regprof3 the result of push pop countings and spill code size

You can use the python script in benchmarks/report.ipynb to view those data quickly, by running ./make jupyter to start a jupyter session.

How to use the compiler

After build the compiler, there is a folder install/llvm contains the clang compiler. The specific compiler flags:

1. -mllvm -fdo-ipra to enable FDOIPRA
2. -mllvm -fdoipra-both-hot=false to control applying optimzaiton for function which entry is hot or having hot loop in its body. When it is false, means only apply to function which entry is hot. (Default: true)
3. -fdoipra-cc=1 to enable optimaztion for Cold Callsite and Cold Callee (Default: true)
4. -fdoipra-ch=1 to enable optimaztion for Cold Callsite and Hot Callee (Default: false)
5. -fdoipra-hc=1 to enable optimaztion for Hot Callsite and Cold Callee (Default: false)
6. -fdoipra-ccr=10 to setup the Callsite Cold Ratio (Default: 10.0)

There should be a 3-steps build:

1. Build the pgo-full version: instrumentation using -fprofile-generate, profiling and then rebuild the PGO with FullLTO
2. Using perf to collect the sampling data and convert the profile using hot-list-creator to generate a hot function list.
3. Build the final binary

You can checkout the small example here: how-to-use/build.mk

Available Benchmarks

1. clang
2. gcc
3. mysql
4. leveldb

Available Targets

1. pgo-full PGO + FullLTO build, the baseline
2. pgo-full-ipra PGO + IPRA + FullLTO build, the existing solution

the following 3 are applying optimazition for function which entry is hot:

3. pgo-full-fdoipra PGO + FullLTO with ColdCallSite-ColdCallee using no_caller_saved_registers attributes
4. pgo-full-fdoipra2 PGO + FullLTO with ColdCallSite-ColdCallee and ColdCallsite-HotCallee using no_caller_saved_registers attributes and a proxy call
5. pgo-full-fdoipra3 PGO + FullLTO with ColdCallSite-ColdCallee, ColdCallsite-HotCallee and indirect calls using no_caller_saved_registers attributes and a proxy call

the following 3 are applying optimzaiton for function which entry is hot or having hot loop in its body:

6. pgo-full-bfdoipra PGO + FullLTO with ColdCallSite-ColdCallee using no_caller_saved_registers attributes
7. pgo-full-bfdoipra2 PGO + FullLTO with ColdCallSite-ColdCallee and ColdCallsite-HotCallee using no_caller_saved_registers attributes and a proxy call
8. pgo-full-bfdoipra3 PGO + FullLTO with ColdCallSite-ColdCallee, ColdCallsite-HotCallee and indirect calls using no_caller_saved_registers attributes and a proxy call

The following 6 are applying no_callee_saved_registers to cold caller and hot callee functions, corresponding to the previous 6 items:

9. pgo-full-fdoipra4 PGO + FullLTO with ColdCallSite-ColdCallee using no_caller_saved_registers attributes
10. pgo-full-fdoipra5 PGO + FullLTO with ColdCallSite-ColdCallee and ColdCallsite-HotCallee using no_caller_saved_registers attributes and a proxy call
11. pgo-full-fdoipra6 PGO + FullLTO with ColdCallSite-ColdCallee, ColdCallsite-HotCallee and indirect calls using no_caller_saved_registers attributes and a proxy call
12. pgo-full-bfdoipra4 PGO + FullLTO with ColdCallSite-ColdCallee using no_caller_saved_registers attributes
13. pgo-full-bfdoipra5 PGO + FullLTO with ColdCallSite-ColdCallee and ColdCallsite-HotCallee using no_caller_saved_registers attributes and a proxy call
14. pgo-full-bfdoipra6 PGO + FullLTO with ColdCallSite-ColdCallee, ColdCallsite-HotCallee and indirect calls using no_caller_saved_registers attributes and a proxy call

Target variants

There are variants for each target, used to configure the threshold for each target. Format: pgo-full-fdoipra.A-B

Number A: The threshold for how many times could be considered as a hot function in sampling data. e.g 3 means if a function has been seen 3 times in the perf sampling data, the system will consider it as a hot function. Number B: Callsite Cold Ratio, in the PGO profiling data, which callsite should be considered as a cold callsite. e.g. 10 means if the hit frequence at entry of the caller function has 10 times larger than the hit frequence at callsite, this is a cold callsite.

Available A: 1 3 5 10 Availabel B: 10 20

How to optimize Google's binary

1. Apply patches to LLVM 16 and build autofdo

IPRA-exp/fdoipra.patch at main · sunxfancy/IPRA-exp (github.com)

IPRA-exp/fix.patch at main · sunxfancy/IPRA-exp (github.com)

Build LLVM and Clang

Build modified version of autofdo (dev branch) sunxfancy/autofdo at dev (github.com)

2. Build PGO LTO version of your binary

a. Build instrumented version:

-fprofile-generate=<output_path>

b. Run tests and profile merge:

… // run tests

cd <output_path> && llvm-profdata merge -output=instrumented.profdata *

c. Build optimized version:

-flto=thin -fprofile-use=instrumented.profdata -Wl,--lto-basic-block-sections=labels -Wl,--build-id

Other flags to avoid bugs:

-fno-optimize-sibling-calls -Wl,-mllvm -Wl,-fast-isel=false -Wl,-Bsymbolic-non-weak-functions

Other flags for best performance:

-fsplit-machine-functions

3. Collect Perf Data using your PGO LTO build and build hot list

a. Run PGO Optimized Version with perf record

perf record -e cycles:u -j any -o samples

b. Generate hotlist

hot_list_creator
    --binary="<PGO VERSION>" \
		--profile="<Sample>" \
		--output="hot_list" \
		--detail="detail" \  # for debugging
		--hot_threshold=3  # 3 counts in samples will mark the function hot

4. Build the final Binary

The last step is to build the final binary using two different groups o information:

a. The PGO profile data

b. The hot list

Build FDOIPRA

-Wl,-mllvm -fdo-ipra -Wl,-fdoipra-new-impl -Wl,-mllvm -Wl,-fdoipra-both-hot=false -Wl,-mllvm -Wl,-fdoipra-ch=1 -Wl,-mllvm -Wl,-fdoipra-hc=1 -Wl,-mllvm -Wl,-fdoipra-use-caller-reg=1

Flags for the hot list:

-Wl,-mllvm -Wl,-fdoipra-hot-list=hot_list

fdo-ipra	Enable FDOIPRA pass
fdoipra-new-impl	There are two implementation, the new one supports ThinLTO which should be good to use.
fdoipra-both-hot	False: only function hot in entry will be considered. True: function hot in entry and body will both be considered as candidates for optimization.
fdoipra-ch	Enable optimization for cold-callsite-hot-callee
fdoipra-hc	Enable optimization for hot-callsite-cold-callee
fdoipra-use-caller-reg	Enable optimization for transferring callee saved registers to caller saved.