Clean up localloc implementation #6276
Conversation
| // push_hide means don't track the stack | ||
| inst_IV(INS_push_hide, 0); | ||
| cntPtrSizedWords--; | ||
| inst_IV(INS_push_hide, 0); // push_hide means don't track the stack |
There was a problem hiding this comment.
An extra comment might be nice here - we're not tracking the stack because the stack tracking is updated separately for localloc, or ...?
|
LGTM with just a couple of comment suggestions. |
|
Looks good. |
src/jit/codegenxarch.cpp
Outdated
| // Align to STACK_ALIGN | ||
| // regCnt will be the total number of bytes to localloc | ||
| inst_RV_IV(INS_add, regCnt, (STACK_ALIGN - 1), emitActualTypeSize(type)); | ||
| // Align the number of bytes to allocate to STACK_ALIGN. |
There was a problem hiding this comment.
This comment should read more like "Round up to the number of bytes to the next higher STACK_ALIGN "
|
LGTM with comments |
|
@briansull - Is this change applicable to Arm64 implementation of genLclHeap()? More specifically, if we have a smaller instruction to decrement a register, we can use that instead of "sub reg, 16" by converting regCnt to loop-iteration count for compInitMem case. |
|
I added / improved a bunch of comments per suggestions |
BruceForstall
force-pushed
the
LocAllocCleanup
branch
from
3d8eb37 to
7a516a2
Compare
|
There aren't any "smaller" instructions on ARM64. |
…anup Clean up localloc implementation Commit migrated from dotnet/coreclr@e8bdf3e
5 participants
Reduce the differences between x86 and AMD64 codegen to reduce the number of
#ifdefs. This changes AMD64 codegen for initialized locallocs by using the x86 style "dec" loop instead of "sub rsp,16". Saves one byte of code size per instance. I ran all standard internal testing on this change.@sivarv PTAL
cc @dotnet/jit-contrib