want -fomit-frame-pointer by default on PowerPC

[Quuxplusone]

Bugzilla Link	PR41094
Status	RESOLVED FIXED
Importance	P enhancement
Reported by	George Koehler (kernigh@gmail.com)
Reported on	2019-03-15 20:20:42 -0700
Last modified on	2019-07-11 19:16:59 -0700
Version	trunk
Hardware	Macintosh OpenBSD
CC	i@maskray.me, llvm-bugs@lists.llvm.org, neeilans@live.com, richard-llvm@metafoo.co.uk
Fixed by commit(s)
Attachments	llvm-9-omit-fp.diff (1096 bytes, text/plain)
Blocks
Blocked by
See also

Created attachment 21609
default -fomit-frame-pointer for PowerPC

Right now, when clang targets PowerPC with optimizations, clang defaults to -
fomit-frame-pointer for Linux and NetBSD, but not for other targets.  The
attached patch (for llvm-project.git master) changes the default to -fomit-
frame-pointer for all PowerPC targets.

I want this because I am targeting 32-bit PowerPC OpenBSD, where gcc already
uses -fomit-frame-pointer by default.  All targets except Darwin have a similar
System V ELF ABI, so if -fomit-frame-pointer works for Linux and NetBSD, it
would probably work for FreeBSD and OpenBSD.  (It might work for Darwin, but
"Darwin is no longer supported for PowerPC" in llvm's PPCTargetMachine.)  The
frame pointer r31, when it exists, is almost always an extra copy of the stack
pointer r1.

Consider this code:   void nothing(void) {}

`clang -target powerpc-openbsd -O3 -S nothing.c` now emits

        stwu 1, -16(1)  # allocate stack frame
        stw 31, 12(1)   # save old r31
        mr 31, 1        # set frame pointer r31 = stack pointer r1
        lwz 31, 12(1)   # restore old r31
        addi 1, 1, 16   # free stack frame
        blr             # return

With -fomit-frame-pointer or with the patch, clang emits only

        blr

Beware that -fno-omit-frame-pointer doesn't work; this patch doesn't fix that.
In the near future, I will backport the patch to clang 7 and compile more code.

[Quuxplusone]

Attached llvm-9-omit-fp.diff (1096 bytes, text/plain): default -fomit-frame-pointer for PowerPC

[Quuxplusone]

I think this is just the gcc default, it doesn't have the OS-varying behavior. Fixed by r365862