This is the mail archive of the mailing list for the GCC 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: Add OPTION_MASK_ISA_X86_64 and support TARGET_BI_ARCH == 2

On 8 Apr 2012, at 15:54, H.J. Lu wrote:
Despite the fact that bootstrap is restored, there remain problems with this
patch and some more work is needed.

(a) [trivial] the option 'mx32' is in i386.opt, which means it is exposed to
all sub-targets, even if they don't support it.

$ ./gcc/xgcc -Bgcc ../tests/hello.c -mx32 -o hc
register name `%rbp'
register name `%rsi'
etc. etc.

It is useful for bootstrap and allows you check out what X32 code looks like even if your OS doesn't support it.

well, OK, but I'd have thought that to be something for developers to experiment with, rather than an 'end user' flag.
However, it is, as stated 'trivial' I'm not going to lose sleep over it ;-).

(b) [serious] the m64 ObjC multi-lib is broken on i?86-darwin* (and likely
there are other more subtle effects).

This is because the code in config/darwin.c that Joseph pointed out (earlier
in this thread) is called for SUBSUBTARGET_OVERRIDE_OPTIONS.

That code sets defaults for, and checks errors for, flags that apply to
*-*-darwin* (and are needed for LTO as well as c-family).

In the case of the ObjC ABI (fobjc-abi-version=) we need to default it to
"2" @ m64 (and default it to 0 or 1 depending on the darwin version @ m32).

I accept that some of this could possibly be done in driver-self- specs;
however, we allow m32/m64 to be unspecified on the c/l and to default for
the target. I'm also not yet sure whether %:version-compare() would be
applicable to fobjc-abi-version.

Thus, the current trunk implementation is broken by your patch and we need
to address that pending other solutions (I'm also very short of free time
for Darwin right now - to experiment with the specs solution).

Please try this


This is not enough to solve the problem, because there are decisions made earlier (e.g. in c-family/c-opts.c re. exceptions) that depend on flag states.

It would be possible to move/repeat some of those (which I did in the proposed fix attached to my last post).


It is possible that there is an options handling issue, (although I might
also have misunderstood) viz:

(current) i386.opt:

;; ISA support

m32 Target RejectNegative Negative(m64) Report InverseMask(ISA_64BIT) Var(ix86_isa_flags) Save Generate 32bit i386 code

Target RejectNegative Negative(mx32) Report Mask(ABI_64) Var(ix86_isa_flags)

Generate 64bit x86-64 code

Target RejectNegative Negative(m32) Report Mask(ABI_X32) Var(ix86_isa_flags)
Generate 32bit x86-64 code

from gccint.pdf (section 8.2):

Negative(othername )
The option will turn oï another option othername, which is the option name
with the leading â-â removed. This chain action will propagate through the
Negative property of the option to be turned oï.
As a consequence, if you have a group of mutually-exclusive options, their
Negative properties should form a circular chain. For example, if options
â-a â, â-b â and â-c â are mutually exclusive, their respective Negative
should be âNegative(b )â, âNegative(c )â and âNegative(a )â.


I read this as "if the User specifies -a on the command line the inverse of
-b *and* the inverse of -c will be applied".

so that when -m64 is issued, the *inverse* of Mask(ABI_X32) should be
applied and then the *inverse* of InverseMask(ISA_64BIT) - which would
(correctly, for the case we're considering) set MASK_ISA_64BIT.

However, in the example above this does NOT happen - if I set a breakpoint
at the entry of x86_internal_override_options - ISA_64BIT ends up as 0 when
-m64 is specified on the c/l for i?86-darwin*.

The x86_isa_explicit stuff doesn't appear to get set either, although
global_options_set.x_x86_isa... does, which I've used in the attached patch.

so is this an options bug or misunderstanding on my part?

There is no problem here. ISA_64BIT is turned on by ABI_64:

Well, I accept that your code in i386.c enforces the assumption above - my point is that the documentation implies (at least to me) that the enforcement should be done in options processing.

The i386.c enforcement happens *after* some points in the code that (currently) assume the behavior as I describe above (or as per the code before your patch).

I am sure that a hybrid of the two patches can be made to work - my concern is to be clear about what is *supposed* to happen at options parsing time.


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