Bootstrap broken for mipsel-linux: gcc/global.c:1020: error: array subscript is above array bounds

[Richard Sandiford]

Richard Sandiford <rsandifo@nildram.co.uk> writes:
> David Daney <ddaney@avtrex.com> writes:
>> For r131631 bootstrap on mipsel-linux in stage2 I am getting:
>>
>> [...]
>> ../../trunk/gcc/global.c:1020: error: array subscript is above array bounds
>> make[3]: *** [global.o] Error 1
>>
>> This is new since r131576:
>>
>> http://gcc.gnu.org/ml/gcc-testresults/2008-01/msg00758.html
>>
>> The configuration is identical to that successful bootstrap/test
>
> global.c:1020 is new code that was added during that time.  There's
> nothing intrinsically wrong with it though.  The warning is simply about
> something that can't prove is dead code.  (I.e. it's another example in
> favour of Mark's "the middle end shouldn't warn" thing.)  A reduced
> testcase is:
>
> ----------------------------------------------------------------------
> #define BIT(X, B) ((X) < ((B) ? 2 : 4) ? (X) + 4 : ~0U)
>
> void bar (unsigned int *);
> void foo (unsigned int b)
> {
>   unsigned int i, a[4];
>
>   for (i = 0; BIT (i, b) != ~0U; i++)
>     a[BIT (i, b) / 32] |= 1U << (BIT (i, b) % 32);
>   bar (a);
> }
> ----------------------------------------------------------------------
>
> which generates a warning on x86_64-linux-gnu when compiled with -O2 -Wall.
> We generate array accesses for a[(i + 4) / 32] and a[~0U / 32] and can't
> prove that the latter are never used.
>
> The attached patch uses the same construct as other E_R_D_R loops
> and avoids the warning.  I'll test it overnight and install as obvious
> if it succeeds.

Bootstrapped & regression-tested on x86_64-linux-gnu.  Tested also
by building mips64-linux-gnu (with --enable-werror and an equivalent
x86_64 CC, as usual).  Installed with this change log:

gcc/
	* global.c (find_reg): Only compute EH_RETURN_DATA_REGNO once per
	input.

> Index: gcc/global.c
> ===================================================================
> --- gcc/global.c	2008-01-19 23:54:32.000000000 +0000
> +++ gcc/global.c	2008-01-19 23:54:56.000000000 +0000
> @@ -1016,8 +1016,13 @@ find_reg (int num, HARD_REG_SET losers, 
>    if (allocno[num].no_eh_reg)
>      {
>        unsigned int j;
> -      for (j = 0; EH_RETURN_DATA_REGNO (j) != INVALID_REGNUM; j++)
> -	SET_HARD_REG_BIT (used1, EH_RETURN_DATA_REGNO (j));
> +      for (j = 0; ; ++j)
> +	{
> +	  unsigned int regno = EH_RETURN_DATA_REGNO (j);
> +	  if (regno == INVALID_REGNUM)
> +	    break;
> +	  SET_HARD_REG_BIT (used1, regno);
> +	}
>      }
>  #endif
>