Skip to content

[WIP] Alternative attempt at compile-on-demand layer integration #44886

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Closed
wants to merge 11 commits into from
Closed

[WIP] Alternative attempt at compile-on-demand layer integration #44886

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
11 commits
Select commit Hold shift + click to select a range
c45f574
Constify various JIT variables and guard RLST with a mutex
pchintalapudi Apr 8, 2022
ae83fbd
Remove non-const access to target machine
pchintalapudi Apr 8, 2022
cd83e25
Lock around stream operations
pchintalapudi Apr 11, 2022
6b4b14a
Explicitly initialize RLST_mutex
pchintalapudi Apr 13, 2022
b2eed69
Initialize RLST before layers to prevent uninitialized mutex errors
pchintalapudi Apr 19, 2022
7132816
Move streams to execution engine to guarantee initialization order
pchintalapudi Apr 19, 2022
5d9a104
Change context pooling to use a queue instead of a stack
pchintalapudi Apr 9, 2022
c3b3bf1
Use pooled memory managers
pchintalapudi Apr 15, 2022
39b8153
Compile-on-demand layer
pchintalapudi Apr 15, 2022
87b25d4
Set partition function
pchintalapudi Apr 16, 2022
300d7c2
Fix aarch64 build and remove broken dump_method_asm_impl
pchintalapudi Apr 19, 2022
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 2 additions & 0 deletions doc/src/devdocs/locks.md
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -29,6 +29,8 @@ The following are definitely leaf locks (level 1), and must not try to acquire a
> * flisp
> * jl_in_stackwalk (Win32)
> * ResourcePool<?>::mutex
> * RLST_mutex
> * jl_locked_stream::mutex
>
> > flisp itself is already threadsafe, this lock only protects the `jl_ast_context_list_t` pool
> > likewise, the ResourcePool<?>::mutexes just protect the associated resource pool
Expand Down
28 changes: 14 additions & 14 deletions src/aotcompile.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -460,11 +460,11 @@ void jl_dump_native_impl(void *native_code,
TheTriple.setOS(llvm::Triple::MacOSX);
#endif
std::unique_ptr<TargetMachine> TM(
jl_ExecutionEngine->getTargetMachine().getTarget().createTargetMachine(
jl_ExecutionEngine->getTarget().createTargetMachine(
TheTriple.getTriple(),
jl_ExecutionEngine->getTargetMachine().getTargetCPU(),
jl_ExecutionEngine->getTargetMachine().getTargetFeatureString(),
jl_ExecutionEngine->getTargetMachine().Options,
jl_ExecutionEngine->getTargetCPU(),
jl_ExecutionEngine->getTargetFeatureString(),
jl_ExecutionEngine->getTargetOptions(),
#if defined(_OS_LINUX_) || defined(_OS_FREEBSD_)
Reloc::PIC_,
#else
Expand All @@ -481,7 +481,7 @@ void jl_dump_native_impl(void *native_code,
));

legacy::PassManager PM;
addTargetPasses(&PM, TM.get());
addTargetPasses(&PM, TM->getTargetTriple(), TM->getTargetIRAnalysis());

// set up optimization passes
SmallVector<char, 0> bc_Buffer;
Expand All @@ -502,7 +502,7 @@ void jl_dump_native_impl(void *native_code,
PM.add(createBitcodeWriterPass(unopt_bc_OS));
if (bc_fname || obj_fname || asm_fname) {
addOptimizationPasses(&PM, jl_options.opt_level, true, true);
addMachinePasses(&PM, TM.get(), jl_options.opt_level);
addMachinePasses(&PM, jl_options.opt_level);
}
if (bc_fname)
PM.add(createBitcodeWriterPass(bc_OS));
Expand Down Expand Up @@ -595,14 +595,14 @@ void jl_dump_native_impl(void *native_code,
delete data;
}

void addTargetPasses(legacy::PassManagerBase *PM, TargetMachine *TM)
void addTargetPasses(legacy::PassManagerBase *PM, const Triple &triple, TargetIRAnalysis analysis)
{
PM->add(new TargetLibraryInfoWrapperPass(Triple(TM->getTargetTriple())));
PM->add(createTargetTransformInfoWrapperPass(TM->getTargetIRAnalysis()));
PM->add(new TargetLibraryInfoWrapperPass(triple));
PM->add(createTargetTransformInfoWrapperPass(std::move(analysis)));
}


void addMachinePasses(legacy::PassManagerBase *PM, TargetMachine *TM, int optlevel)
void addMachinePasses(legacy::PassManagerBase *PM, int optlevel)
{
// TODO: don't do this on CPUs that natively support Float16
PM->add(createDemoteFloat16Pass());
Expand Down Expand Up @@ -857,9 +857,9 @@ class JuliaPipeline : public Pass {
(void)jl_init_llvm();
PMTopLevelManager *TPM = Stack.top()->getTopLevelManager();
TPMAdapter Adapter(TPM);
addTargetPasses(&Adapter, &jl_ExecutionEngine->getTargetMachine());
addTargetPasses(&Adapter, jl_ExecutionEngine->getTargetTriple(), jl_ExecutionEngine->getTargetIRAnalysis());
addOptimizationPasses(&Adapter, OptLevel, true, dump_native, true);
addMachinePasses(&Adapter, &jl_ExecutionEngine->getTargetMachine(), OptLevel);
addMachinePasses(&Adapter, OptLevel);
}
JuliaPipeline() : Pass(PT_PassManager, ID) {}
Pass *createPrinterPass(raw_ostream &O, const std::string &Banner) const override {
Expand Down Expand Up @@ -993,9 +993,9 @@ void *jl_get_llvmf_defn_impl(jl_method_instance_t *mi, size_t world, char getwra
static legacy::PassManager *PM;
if (!PM) {
PM = new legacy::PassManager();
addTargetPasses(PM, &jl_ExecutionEngine->getTargetMachine());
addTargetPasses(PM, jl_ExecutionEngine->getTargetTriple(), jl_ExecutionEngine->getTargetIRAnalysis());
addOptimizationPasses(PM, jl_options.opt_level);
addMachinePasses(PM, &jl_ExecutionEngine->getTargetMachine(), jl_options.opt_level);
addMachinePasses(PM, jl_options.opt_level);
}

// get the source code for this function
Expand Down
12 changes: 6 additions & 6 deletions src/codegen.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -186,11 +186,10 @@ typedef Instruction TerminatorInst;
#include "processor.h"
#include "julia_assert.h"

JL_STREAM *dump_emitted_mi_name_stream = NULL;
extern "C" JL_DLLEXPORT
void jl_dump_emitted_mi_name_impl(void *s)
{
dump_emitted_mi_name_stream = (JL_STREAM*)s;
**jl_ExecutionEngine->get_dump_emitted_mi_name_stream() = (JL_STREAM*)s;
}

extern "C" {
Expand Down Expand Up @@ -7978,15 +7977,16 @@ jl_llvm_functions_t jl_emit_code(
"functions compiled with custom codegen params must not be cached");
JL_TRY {
decls = emit_function(m, li, src, jlrettype, params);
if (dump_emitted_mi_name_stream != NULL) {
jl_printf(dump_emitted_mi_name_stream, "%s\t", decls.specFunctionObject.c_str());
auto stream = *jl_ExecutionEngine->get_dump_emitted_mi_name_stream();
if (stream) {
jl_printf(stream, "%s\t", decls.specFunctionObject.c_str());
// NOTE: We print the Type Tuple without surrounding quotes, because the quotes
// break CSV parsing if there are any internal quotes in the Type name (e.g. in
// Symbol("...")). The \t delineator should be enough to ensure whitespace is
// handled correctly. (And we don't need to worry about any tabs in the printed
// string, because tabs are printed as "\t" by `show`.)
jl_static_show(dump_emitted_mi_name_stream, li->specTypes);
jl_printf(dump_emitted_mi_name_stream, "\n");
jl_static_show(stream, li->specTypes);
jl_printf(stream, "\n");
}
}
JL_CATCH {
Expand Down
5 changes: 3 additions & 2 deletions src/disasm.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1208,9 +1208,10 @@ jl_value_t *jl_dump_function_asm_impl(void *F, char raw_mc, const char* asm_vari
f2.deleteBody();
}
});
LLVMTargetMachine *TM = static_cast<LLVMTargetMachine*>(&jl_ExecutionEngine->getTargetMachine());
auto TMBase = jl_ExecutionEngine->cloneTargetMachine();
LLVMTargetMachine *TM = static_cast<LLVMTargetMachine*>(TMBase.get());
legacy::PassManager PM;
addTargetPasses(&PM, TM);
addTargetPasses(&PM, TM->getTargetTriple(), TM->getTargetIRAnalysis());
if (raw_mc) {
raw_svector_ostream obj_OS(ObjBufferSV);
if (TM->addPassesToEmitFile(PM, obj_OS, nullptr, CGFT_ObjectFile, false, nullptr))
Expand Down
Loading