New ABI for stack layout and frame pointer scheme

[kito-cheng]

This is a placeholder for discuss the issue around Zcmp frame pointer issue.

The original thread from RVI mailing list https://lists.riscv.org/g/tech-psabi/message/154

Link for history:

• reversed push/pop order breaks de facto ABI for -fno-omit-frame-pointer riscvarchive/riscv-code-size-reduction#194
• https://lists.riscv.org/g/tech-code-size/topic/84238872

Slide from Qcom:
Qualcomm RISC-V Push&Pop&FP Proposal.pdf

Current stack layout and frame pointer scheme:

And proposed stack layout and frame pointer scheme:

[kito-cheng]

cc. @cmuellner @apazos @topperc @asb @aswaterman

[apazos]

Thanks @kito-cheng for posting this issue.

A few remarks about the clipped figures from the PDF file posted in the discussion:

• The current ABI states that “the frame pointer register will point to the Canonical Frame Address or CFA, which is the stack pointer value on entry to the current procedure.”
• In the alternative ABI "the frame pointer points to the saved return address in the CM.PUSH/CM.POP stack frame".
• If we opt for an alternative ABI for no-omit-frame-pointer + Zcmp and rely on the linker to do compatibility checks for ABI mix, I think we would need an ELF attribute that indicates that the frame pointer is enabled so the linker can implement the proper checks.

[james-ball-qualcomm]

The previously attached PDF of the slides is out of date. Please use the latest version at:
https://docs.google.com/presentation/d/1Qk-pNJ5svY8trYW3Smk073ok9pcEJFc5hsx4rbOoj5M

[sorear]

I would suggest to point the fp within each stack frame one word after the saved s0 value, so that -XLEN/8(fp) is the saved fp and -2*XLEN/8(fp) is the saved ra. This allows stack trace code to support both conventions (mixed, even) in the common case that stack bounds are known:

[unwind psuedocode]
while ((fp - stack_base) < stack_size) {
    if ((fp[-1] - stack_base) < stack_size) {
        ra = fp[-2]; fp = fp[-1]; /* Zcmp optimized convention */
    } else {
        ra = fp[-1]; fp = fp[-2]; /* ABI 1.1 convention */
    }
    print(ra);
}

This works because valid stack addresses are disjoint from valid instruction addresses.

A preprocessor define will be needed to communicate the use of the new ABI to unwinders when there is not enough information at runtime to distinguish stack and instruction addresses. It might be worth defining an ELF attribute for future-proofing, although mixing code using both frame pointer conventions may not be an error.