Introduce MACHINE_ABI as a list of ABI properties
ClosedPublic
Actions

Authored by brooks on Sep 1 2022, 10:25 PM.

Details

Reviewers

jrtc27
imp
jhb

Commits

rG426fc376afaf: bsd.cpu.mk: Introduce MACHINE_ABI

Summary

MACHINE_ABI is a list of properties about the ABI used for MACHINE_ARCH.
It should be used in place of long conditionals on MACHINE_ARCH where
practical.

The following properties are indicated with one of the follow values:

Floating point ABI: soft-float, hard-float
Size of long (size_t, etc): long32, long64
Pointer type: ptr32, ptr64
Size of time_t: time32, time64

For example, i386 targets will be:
MACHINE_ABI= hard-float long32 ptr32 time32

Obtained from: CheriBSD
Sponsored by: DARPA, AFRL

Diff Detail

Repository

rG FreeBSD src repository

Lint

Lint Not Applicable

Unit

Tests Not Applicable

Event Timeline

brooks created this revision.Sep 1 2022, 10:25 PM

Herald added subscribers: emaste, bdrewery. · View Herald TranscriptSep 1 2022, 10:25 PM

brooks requested review of this revision.Sep 1 2022, 10:25 PM

Harbormaster completed remote builds in B47225: Diff 110063.Sep 1 2022, 10:25 PM

brooks added a parent revision: D36420: Rename MACHINE_ABI and TARGET_ABI.Sep 1 2022, 10:25 PM

brooks added a child revision: D36422: libc: Include quad support on long32 ABIs.

I'd be willing to entertain another name like MACHINE_ABI_PROPS to avoid the need for D36420, but we use MACHINE_ABI extensively in CheriBSD and I'm not convinced anyone uses the current variable.

For example, i386 targets will be:
MACHINE_ARCH= hard-float long32 ptr62 time32

s/ptr62/ptr32/ and s/ARCH/ABI/

In D36421#827216, @jrtc27 wrote:

For example, i386 targets will be:
MACHINE_ARCH= hard-float long32 ptr62 time32

s/ptr62/ptr32/ and s/ARCH/ABI/

That was creative of me... Fixed in https://github.com/freebsd/freebsd-src/compare/main...brooksdavis:freebsd:machine_abi

brooks edited the summary of this revision. (Show Details)Sep 1 2022, 10:36 PM

Generally, I quite like it.
Just one question, though...

share/mk/bsd.cpu.mk
376	powerpcspe too?

jrtc27 added inline comments.Sep 1 2022, 10:42 PM

share/mk/bsd.cpu.mk
376	That's not soft-float, that's hard-float-in-integer-register-file?

imp added inline comments.Sep 1 2022, 10:46 PM

share/mk/bsd.cpu.mk
376	That's what 'softfloat' means on arm too... though that's a bit squishy

jrtc27 added inline comments.Sep 1 2022, 10:48 PM

share/mk/bsd.cpu.mk
376	The difference here is that powerpcspe's baseline includes floating-point instructions, and those instructions are defined to operate on the integer registers, i.e. the hardware has merged the two register files (and in fact the floating-point ones are 64-bit extensions of the integer ones). Whereas on arm it's "put your floating-point arguments in integer registers, and if you compile with floating-point instructions enabled then you can pull them out into floating-point registers and use native instructions".

Can you also add endianness?

Add endian

Harbormaster completed remote builds in B47578: Diff 111105.Sep 27 2022, 5:52 PM

We can revisit if we ever add something weird like aarch64be :).
But since that's at best speculative (and at worst manu / andy going hell no), we're good to go with this version.

This revision is now accepted and ready to land.Sep 27 2022, 6:08 PM

Add missing guard on CPUTYPE

This revision now requires review to proceed.Sep 27 2022, 6:24 PM

Harbormaster completed remote builds in B47579: Diff 111108.Sep 27 2022, 6:24 PM

eduardo added a subscriber: eduardo.Sep 27 2022, 9:09 PM

jhb added inline comments.Sep 30 2022, 8:53 PM

share/mk/bsd.cpu.mk
376	It is true though that the old CPUTYPE=soft thing for arm (which maybe we can just kill in 14.0 to make this less painful?) isn't really soft-float in that you want soft-float routines in libc or additional builtins in libcompiler_rt, at least IIRC.