This is the mail archive of the binutils@sourceware.org mailing list for the binutils project.


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]
Other format: [Raw text]

Re: [PATCH 06/15] MIPS/GAS/test: Adjust LD for multi-target testing


"Maciej W. Rozycki" <macro@linux-mips.org> writes:
> 2010-10-03  Maciej W. Rozycki  <macro@linux-mips.org>
>
> 	gas/testsuite/
> 	* gas/mips/ld.d: Spell out section offsets and addends
> 	explicitly.  Clean up some regexps.
> 	* gas/mips/ld-ilocks.d: Likewise.  Add missing "$" prefixes to
> 	the names of FP registers.
> 	* gas/mips/ld-ilocks-addr32.d: Likewise.
> 	* gas/mips/ld.s: Align sections to 4k, adjust padding.

The patch seems to contain some parts that make the matching more strict
and some that make it less so.  The former look good, but I'm less sure
about the latter.  For example, with something like:

[0-9a-f]+ <[^>]*> lw   a0,(0|4096)\(at\)
[      ]*[0-9a-f]+: [A-Z0-9_]*LO[A-Z0-9_]*     \.data(\+0xfffff000)?

the difference seems relatively important: on some targets the addend
must be in-place, on some it mustn't be.  It might make more sense
to have separate dumps for the two cases.  However, that sort of
difference is probably picked up by other tests too, so the risk
might not be great.

Also, how confident are we that the more relaxed tests are correct
for ECOFF?  You say:

> 2. Consequently section alignment is requested at 4k so that offsets are 
>    round and do not change for ECOFF when instructions are added or 
>    removed (ECOFF seems to make data relocations relative to .text).

but is that intentional?  The point here, and here:

> 3. Missing "$" prefixes are added to the names of FP registers (shows how 
>    frequently some of the targets are tested, hmm...).

is precisely that mips*-ecoff has severely bitrotted from lack of
active interest, so who knows how good the current output actually is?
There's a danger in simply making the testsuite match what the current
ECOFF assembler does.

In a similar vein, the difference between endiannesses is something
that could easily go wrong, but there are separate tests that should
pick that up.  I'm much more comfortable with the endianness bits.

Overall the patch is improvement, so it's OK to apply despite
the concerns above.

Richard


Index Nav: [Date Index] [Subject Index] [Author Index] [Thread Index]
Message Nav: [Date Prev] [Date Next] [Thread Prev] [Thread Next]