Skip to content

Re-organize LLVM JIT mode implementation #1450

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

Jump to bottom
Merged
merged 3 commits into from
Sep 7, 2022
Merged

Re-organize LLVM JIT mode implementation #1450

merged 3 commits into from
Sep 7, 2022

Conversation

@lum1n0us
Copy link
Collaborator

@lum1n0us lum1n0us commented Sep 4, 2022

It includes:

  • Remove code with MCJIT APIs. Embrance ORC.
    • WASM_ENABLE_LAZY_JIT only represent lazy compilation. It won't specify
      ORC or MCJIT
    • Remove runtime ORC compilation threads and move all aot functions
      into one module
    • Current LLVM ORC JIT implements eagel compilation. Will add lazy
      compilation later
  • JIT and AOT both apply the new pass manager

@lum1n0us lum1n0us force-pushed the remove_mcjit branch 3 times, most recently from d672ec9 to d3035ab Compare September 5, 2022 02:57
wenyongh
wenyongh reviewed Sep 5, 2022
View reviewed changes
@lum1n0us lum1n0us force-pushed the remove_mcjit branch 3 times, most recently from d10091e to ea04105 Compare September 5, 2022 07:01
wenyongh
wenyongh reviewed Sep 5, 2022
View reviewed changes
core/iwasm/compilation/aot_llvm.h
/* LLVM execution engine required by JIT */
#if WASM_ENABLE_LAZY_JIT != 0
LLVMOrcLLJITRef orc_lazyjit;
LLVMOrcLLJITRef orc_jit;
Copy link
Contributor

@wenyongh wenyongh Sep 5, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Seems that we don't use the WAMR_BUILD_JIT macro to control the fields and the code in core/iwasm/compilation?

wenyongh
wenyongh reviewed Sep 5, 2022
View reviewed changes
@lum1n0us lum1n0us force-pushed the remove_mcjit branch 3 times, most recently from 11a468d to 79273ef Compare September 6, 2022 00:26
@lum1n0us
Re-organize LLVM JIT mode implementation
772f6ee 
It includes:
- Remove code with MCJIT APIs. Embrance ORC.
  - WASM_ENABLE_LAZY_JIT only represent lazy compilation. It won't specify
    ORC or MCJIT
  - Remove runtime ORC compilation threads and move all aot functions
    into one module
  - Current LLVM ORC JIT implements eagel compilation. Will add lazy
    compilation later
- JIT and AOT both apply the new pass manager
  - All JIT passes are in IRTransferLayer

P.S.
| Varient impl.           | Norm. Score | Norm. (ms) |
| ----------------------- | ----------- | ---------- |
| Original                | 1.0         | 1.0        |
| Apply the new PM        | 1.01        | 1.007      |
| Opt in IRTransformLayer | 0.997       | 1.04       |
wenyongh
wenyongh reviewed Sep 6, 2022
View reviewed changes
core/iwasm/compilation/aot_emit_function.c
}
}
else {
#if WASM_ENABLE_LAZY_JIT == 0
Copy link
Contributor

@wenyongh wenyongh Sep 6, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should not remove the code of if branch, which is the operation of AOT compiler mode and eager JIT mode. Removing the code may impact the performance of AOT.

How about adding a flag:

bool is_lazy_jit = false;
#if WASM_ENABLE_LAZY_JIT != 0
if (comp_ctx->is_jit_mode) {
    is_lazy_jit = true;
}
#endif

if (!is_lazy_jit) {
    func = func_ctxes[func_idx - import_func_count]->func;
}
else {
    code of original #else branch
}

Copy link
Collaborator Author

@lum1n0us lum1n0us Sep 6, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Right.
Second thoughts, there will be only one module even in 'lazy compilation' and the code of original #else branch may not be necessary.

Copy link
Contributor

@wenyongh wenyongh Sep 6, 2022

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Not very sure, how about keeping the code currently and testing it when implementing ORCv2?

@lum1n0us
Performance Tuning
d5f55e1 
1. Register the `AAManager` and `buildDefaultAAPipeline`
2. Use function pointers in `aot_compile_op_call` and `aot_compile_op_indirect_call`
   instead of `lookup_orcjit_func`

CoreMark Score:
| Varient impl.           | Norm. Score | Norm. (ms) |
| ----------------------- | ----------- | ---------- |
| Original                | 1.0         | 1.0        |
| Apply the new PM        | 1.01        | 1.007      |
| Opt in IRTransformLayer | 0.997       | 1.04       |
| Tuning#1                | 1.01        | 0.83       |
@lum1n0us
Copy link
Collaborator Author

lum1n0us commented Sep 6, 2022

I have uploaded a new improvement which includes:

  1. Register the AAManager and buildDefaultAAPipeline
  2. Use function pointers in aot_compile_op_call and aot_compile_op_indirect_call instead of lookup_orcjit_func

If it is OK, I will remove lookup_orcjit_func related stuff. For now, the branch is commented for now.

@lum1n0us lum1n0us mentioned this pull request Sep 6, 2022
Merged
@wenyongh wenyongh merged commit 5db3412 into bytecodealliance:dev/refactor_llvm_jit Sep 7, 2022
@lum1n0us lum1n0us deleted the remove_mcjit branch October 1, 2022 11:41
@wenyongh wenyongh mentioned this pull request Oct 11, 2022
Closed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Reviewers

@wenyongh wenyongh wenyongh left review comments

Assignees

No one assigned

Labels

None yet

Projects

None yet

Milestone

No milestone

Development

Successfully merging this pull request may close these issues.

2 participants

@lum1n0us @wenyongh