[Core 12] arch/amd64: let the frame pointer return #17
Conversation
https://www.brendangregg.com/blog/2024-03-17/the-return-of-the-frame-pointers.html argues that amd64 has enough registers to not be bothered by the FP. Google, Meta, Fedora, Ubuntu, and Arch seem to agree. This commit affects all arches that do not override the following variables: * CFLAGS_COMMON_OPTI * RUSTFLAGS_COMMON_OPTI
|
And, for the record, I second this change. |
|
Huh, I wasn't even aware that the original got merged. Must have slipped my mind… |
It was a gap in communication - I have noted that we should wait for Core 12 (i.e., in a month) before this change but only on Telegram. Sorry about that. |
| @@ -4,7 +4,7 @@ | |||
|
|
|||
| # C Compiler Flags. | |||
| CFLAGS_COMMON=('-pipe' '-Wno-error') | |||
| CFLAGS_COMMON_OPTI=('-O2') | |||
| CFLAGS_COMMON_OPTI=('-O2' '-fno-omit-frame-pointer' '-mno-omit-leaf-frame-pointer') | |||
There was a problem hiding this comment.
Some CPU architectures may not have the concept of "frame pointers", enabling frame pointers on those platforms will cause the compiler to use an emulated version of "frame pointers" (while sacrificing an extra register for storing emulated frame pointer data).
There was a problem hiding this comment.
This is the case for almost any architecture. The concept of a frame pointer comes from the ABI, not from the architecture. In the case of x86-64, %rbp is one of 16 GP registers chosen to act as a frame pointer by ABI convention, and using -fomit frees it to be used for other purposes.
Now if 16 GP registers is enough to not omit the one frame pointer, the many RISC architectures with 32 or more GP registers must be completely okay with not omitting.
Apologies for abruptly reverting that change but this change is best queued for Core 12 when we survey potential changes for core toolchains. cc @Artoria2e5 @jiegec
From original pull request: