Re: [Qemu-devel] [PULL 0/5] tcg queued patches

[Richard Henderson <rth@twiddle.net>]

On 02/16/2016 12:04 AM, Peter Maydell wrote:
> On 15 February 2016 at 11:29, Richard Henderson <rth@twiddle.net> wrote:
>> ----------------------------------------------------------------
>> Richard Henderson (5):
>>        tcg: Work around clang bug wrt enum ranges, part 2
>>        tcg: Implement indirect memory registers
>>        tcg: Allocate indirect_base temporaries in a different order
>>        target-sparc: Tidy global register initialization
>>        target-sparc: Use global registers for the register window
>>
>>   target-sparc/translate.c | 196 ++++++++++++++++++++++-------------------------
>>   tcg/tcg.c                | 138 +++++++++++++++++++++++----------
>>   tcg/tcg.h                |   2 +
>>   3 files changed, 192 insertions(+), 144 deletions(-)
>
> This failed to build on the i686-w64-mingw32 compiler
> ("i686-w64-mingw32-gcc (GCC) 4.8.2", from the Ubuntu
> gcc-mingw-w64-i686 package version 4.8.2-10ubuntu2+12):
>
> target-sparc/translate.c: In function ‘gen_intermediate_code’:
> target-sparc/translate.c:299:24: error: array subscript is above array
> bounds [-Werror=array-bounds]
>           return cpu_regs[reg];
>                          ^
>
> Fiddling around with the source file to see which call to
> gen_dest_gpr() is provoking this shows that it's the one
> in gen_ldda_asi() at line 2157 -- if I change the second
> argument from 'rd + 1' to 'rd' it compiles OK.
>
> Changing the call site of gen_ldda_asi at line 4727 so its last
> argument is 'rd & ~1' rather than 'rd' also suppresses the
> error. (That can't possibly change the semantics because we've
> just done "if (rd & 1) goto illegal_insn;"...)
>
> I'm generally reluctant to suggest compiler bugs, but this does
> look rather like a compiler bug...

There are at least 5 such bugs open against gcc at the moment.

     https://gcc.gnu.org/bugzilla/show_bug.cgi?id=56456

I couldn't reproduce this quickly with a freshly built gcc 4.8 branch on 
i686-linux.  Could you please file a gcc bug with your preprocessed source? 
There's a chance it isn't a duplicate, but...

In the meantime... hmm.  I don't suppose removing the inline helps?  Probably 
not, since there's only one caller...  Otherwise I guess we should go with your 
"rd & ~1" workaround.


r~