christoph/wamr
1
Fork 0
Code Releases Activity

Implement AOT static PGO (#2243)

Browse Source 
LLVM PGO (Profile-Guided Optimization) allows the compiler to better optimize code
for how it actually runs. This PR implements the AOT static PGO, and is tested on
Linux x86-64 and x86-32. The basic steps are:

1. Use `wamrc --enable-llvm-pgo -o <aot_file_of_pgo> <wasm_file>`
   to generate an instrumented aot file.
2. Compile iwasm with `cmake -DWAMR_BUILD_STATIC_PGO=1` and run
      `iwasm --gen-prof-file=<raw_profile_file> <aot_file_of_pgo>`
    to generate the raw profile file.
3. Run `llvm-profdata merge -output=<profile_file> <raw_profile_file>`
    to merge the raw profile file into the profile file.
4. Run `wamrc --use-prof-file=<profile_file> -o <aot_file> <wasm_file>`
    to generate the optimized aot file.
5. Run the optimized aot_file: `iwasm <aot_file>`.

The test scripts are also added for each benchmark, run `test_pgo.sh` under
each benchmark's folder to test the AOT static pgo.
This commit is contained in:
Wenyong Huang
2023-06-05 09:17:39 +08:00
committed by GitHub
parent f1e9029ebc
commit 8d88471c46
29 changed files with 2000 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

67
core/iwasm/compilation/aot_llvm_extra.cpp
View File

@ -44,6 +44,7 @@
#if LLVM_VERSION_MAJOR >= 12
#include <llvm/Analysis/AliasAnalysis.h>
#endif
#include <llvm/ProfileData/InstrProf.h>
#include <cstring>
#include "../aot/aot_runtime.h"
@ -232,14 +233,26 @@ aot_apply_llvm_new_pass_manager(AOTCompContext *comp_ctx, LLVMModuleRef module)
PTO.SLPVectorization = true;
PTO.LoopUnrolling = true;
Optional<PGOOptions> PGO = None;
if (comp_ctx->enable_llvm_pgo) {
/* Disable static counter allocation for value profiler,
it will be allocated by runtime */
const char *argv[] = { "", "-vp-static-alloc=false" };
cl::ParseCommandLineOptions(2, argv);
PGO = PGOOptions("", "", "", PGOOptions::IRInstr);
}
else if (comp_ctx->use_prof_file) {
PGO = PGOOptions(comp_ctx->use_prof_file, "", "", PGOOptions::IRUse);
}
#ifdef DEBUG_PASS
PassInstrumentationCallbacks PIC;
PassBuilder PB(TM, PTO, None, &PIC);
PassBuilder PB(TM, PTO, PGO, &PIC);
#else
#if LLVM_VERSION_MAJOR == 12
PassBuilder PB(false, TM, PTO);
PassBuilder PB(false, TM, PTO, PGO);
#else
PassBuilder PB(TM, PTO);
PassBuilder PB(TM, PTO, PGO);
#endif
#endif
@ -334,8 +347,16 @@ aot_apply_llvm_new_pass_manager(AOTCompContext *comp_ctx, LLVMModuleRef module)
FPM.addPass(SLPVectorizerPass());
FPM.addPass(LoadStoreVectorizerPass());
if (comp_ctx->enable_llvm_pgo || comp_ctx->use_prof_file) {
LICMOptions licm_opt;
/* LICM pass: loop invariant code motion, attempting to remove
as much code from the body of a loop as possible. Experiments
show it is good to enable it when pgo is enabled. */
FPM.addPass(
createFunctionToLoopPassAdaptor(LICMPass(licm_opt), true));
}
/*
FPM.addPass(createFunctionToLoopPassAdaptor(LICMPass()));
FPM.addPass(createFunctionToLoopPassAdaptor(LoopRotatePass()));
FPM.addPass(createFunctionToLoopPassAdaptor(SimpleLoopUnswitchPass()));
*/
@ -344,9 +365,10 @@ aot_apply_llvm_new_pass_manager(AOTCompContext *comp_ctx, LLVMModuleRef module)
if (!disable_llvm_lto) {
/* Apply LTO for AOT mode */
if (comp_ctx->comp_data->func_count >= 10)
/* Adds the pre-link optimizations if the func count
is large enough */
if (comp_ctx->comp_data->func_count >= 10
|| comp_ctx->enable_llvm_pgo || comp_ctx->use_prof_file)
/* Add the pre-link optimizations if the func count
is large enough or PGO is enabled */
MPM.addPass(PB.buildLTOPreLinkDefaultPipeline(OL));
else
MPM.addPass(PB.buildLTODefaultPipeline(OL, NULL));
@ -358,3 +380,34 @@ aot_apply_llvm_new_pass_manager(AOTCompContext *comp_ctx, LLVMModuleRef module)
MPM.run(*M, MAM);
}
char *
aot_compress_aot_func_names(AOTCompContext *comp_ctx, uint32 *p_size)
{
std::vector<std::string> NameStrs;
std::string Result;
char buf[32], *compressed_str;
uint32 compressed_str_len, i;
for (i = 0; i < comp_ctx->func_ctx_count; i++) {
snprintf(buf, sizeof(buf), "%s%d", AOT_FUNC_PREFIX, i);
std::string str(buf);
NameStrs.push_back(str);
}
if (collectPGOFuncNameStrings(NameStrs, true, Result)) {
aot_set_last_error("collect pgo func name strings failed");
return NULL;
}
compressed_str_len = Result.size();
if (!(compressed_str = (char *)wasm_runtime_malloc(compressed_str_len))) {
aot_set_last_error("allocate memory failed");
return NULL;
}
bh_memcpy_s(compressed_str, compressed_str_len, Result.c_str(),
compressed_str_len);
*p_size = compressed_str_len;
return compressed_str;
}