tqchen commented on PR #634:
URL: https://github.com/apache/tvm-ffi/pull/634#issuecomment-4744873263
Side codegen probe for dynamic `RuntimeTypeIndex()` access.
Probe:
```cpp
struct DynProbeObj : public tvm::ffi::Object {
static constexpr const char* _type_key = "probe.DynProbe";
TVM_FFI_DECLARE_OBJECT_INFO_PREDEFINED_TYPE_KEY(DynProbeObj,
tvm::ffi::Object);
};
extern "C" int32_t ReadDynTypeIndex() {
return DynProbeObj::RuntimeTypeIndex();
}
```
Compiler: `c++ (Ubuntu 11.4.0-1ubuntu1~22.04.2) 11.4.0` with `-std=c++17 -O2
-fPIC`.
Hot accessor asm:
```diff
ReadDynTypeIndex:
endbr64
- push r12
- push rbp
- push rbx
- sub rsp,0x20
- mov rbp,QWORD PTR [rip+0x0]
- R_X86_64_REX_GOTPCRELX _ZGVZN...E6tindex-0x4
- ... stack canary/type_key setup ...
- movzx eax,BYTE PTR [rbp+0x0]
- test al,al
- je <guard_acquire_path>
- mov rbx,QWORD PTR [rip+0x0]
- R_X86_64_REX_GOTPCRELX _ZZN...E6tindex-0x4
- mov eax,DWORD PTR [rbx]
- ... restore/ret ...
-
-guard_acquire_path:
- call __cxa_guard_acquire
- call TVMFFITypeGetOrAllocIndex@PLT
- call __cxa_guard_release
+ mov rax,QWORD PTR [rip+0x0]
+ R_X86_64_REX_GOTPCRELX _ZN11DynProbeObj11_type_indexE-0x4
+ mov eax,DWORD PTR [rax]
+ ret
```
Symbol sizes from `nm -S --size-sort`:
| Symbol | Before | After |
|---|---:|---:|
| `ReadDynTypeIndex` | 176 B | 14 B |
| `ForceDynTypeIndexInit` | 176 B | 9 B |
| `DynProbeObj::_GetOrAllocRuntimeTypeIndex` | inlined into callers | 179 B
weak cold symbol |
| `_GLOBAL__sub_I_probe.cc` | none | 21 B startup + 20 B cold |
Section-size diff from `size -A probe.o`:
| Section | Before | After | Delta |
|---|---:|---:|---:|
| `.text` | 352 B | 14 B | -338 B |
| `.text.unlikely*` | 32 B | 208 B | +176 B |
| `.text.startup` | 0 B | 21 B | +21 B |
| `.init_array` | 0 B | 8 B | +8 B |
| `_GetOrAllocRuntimeTypeIndex` local static guard | 8 B | 8 B | 0 B |
| `_GetOrAllocRuntimeTypeIndex` local `tindex` | 4 B | 4 B | 0 B |
| new inline `_type_index` guard | 0 B | 8 B | +8 B |
| new inline `_type_index` storage | 0 B | 4 B | +4 B |
| `.rodata*` type key | 15 B | 15 B | 0 B |
| `.gcc_except_table*` | 36 B | 14 B | -22 B |
| `.eh_frame` | 208 B | 184 B | -24 B |
| total object | 769 B | 634 B | -135 B |
Takeaway: the static inline cache adds a small guard/init/data cost, but
removes the guarded allocation path from the hot accessor. The hot load becomes
direct, `_GetOrAllocRuntimeTypeIndex` moves to cold/startup code, and this
probe still shrinks overall object size.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
