Re: mips-elf-gcc -fno-delayed-branch problem

[Richard Sandiford <rdsandiford at googlemail dot com>]

Toshi Morita <tm314159@yahoo.com> writes:
> Maybe GAS could recognize -fno-delayed-branch to selectively disable
> branch slot filling?

I'd agree if it was -mno-delayed-branch.  I think -f* options are
generally compiler options, while -m* options are target options that
could in principle be passed down to either the assembler or the linker.

> Is there a list of optimizations performed by MIPS GAS listed somewhere?
> It would be nice if these could be selectively enabled.

The only other optimisation (if it can even be called that) is increased
accuracy regarding nop insertion.  Suppose we have something like:

	.text
	lw	$4,foo
	addiu	$5,$5,1
	jr	$31
	.data
foo:
	.word	1

When GAS sees the LW, it doesn't know whether the LW should use a
HI/LO pair or a GP-relative access.  It therefore creates a variant
"frag" that describes both possibilities.  As far as GAS is concerned,
the following ADDIU starts a new subblock of code.

With -Wa,-O0, GAS doesn't try to handle dependencies between these subblocks,
and just assumes the worst.  So if you assemble with -mips1, GAS has to
assume that the next subblock after the LW might use $4 straight away,
and that a nop is needed:

00000000 <.text>:
   0:   3c040000        lui     a0,0x0
                        0: R_MIPS_HI16  .data
   4:   8c840000        lw      a0,0(a0)
                        4: R_MIPS_LO16  .data
   8:   00000000        nop
   c:   24a50001        addiu   a1,a1,1
  10:   03e00008        jr      ra
  14:   00000000        nop

At -Wa,-O1 and above it does the sensible thing:

00000000 <.text>:
   0:   3c040000        lui     a0,0x0
                        0: R_MIPS_HI16  .data
   4:   8c840000        lw      a0,0(a0)
                        4: R_MIPS_LO16  .data
   8:   24a50001        addiu   a1,a1,1
   c:   03e00008        jr      ra
  10:   00000000        nop

TBH, I think the cases where you'd want the -O0 behaviour are
vanishingly rare.  It does in principle need less memory, and does
in principle assemble slightly quicker, but I don't think anyone would
notice unless they looked hard.

So -Wa,-O1 is better than the -Wa,-O0 that I mentioned previously.

Richard