Re: [PATCH][PR66013] Update address_taken after ifn_va_arg expansion

[Richard Biener <richard dot guenther at gmail dot com>]

On Fri, May 8, 2015 at 5:13 PM, Tom de Vries <Tom_deVries@mentor.com> wrote:
> Hi,
>
> this patch fixes PR66013.
>
>
> I.
>
> Consider this test-case, with a va_list passed from f2 to f2_1:
> ...
> #include <stdarg.h>
>
> inline int __attribute__((always_inline))
> f2_1 (va_list ap)
> {
>   return va_arg (ap, int);
> }
>
> int
> f2 (int i, ...)
> {
>   int res;
>   va_list ap;
>
>   va_start (ap, i);
>   res = f2_1 (ap);
>   va_end (ap);
>
>   return res;
> }
> ...
>
> When compiling at -O2 with -m32, before pass_stdarg we see that va_start and
> va_arg (in the same function after inlining) use different aps (with the
> same value though):
> ...
>   # .MEM_2 = VDEF <.MEM_1(D)>
>   # USE = nonlocal escaped
>   # CLB = nonlocal escaped { D.1809 }
>   __builtin_va_startD.1021 (&apD.1809, 0);
>
>   # VUSE <.MEM_2>
>   # PT = nonlocal
>   ap.0_3 = apD.1809;
>
>   # .MEM_4 = VDEF <.MEM_2>
>   apD.1820 = ap.0_3;
>
>   # .MEM_8 = VDEF <.MEM_4>
>   # USE = nonlocal null { D.1820 } (escaped)
>   # CLB = nonlocal null { D.1820 } (escaped)
>   _7 = VA_ARG (&apD.1820, 0B, 1);
> ...
>
> After expand_ifn_va_arg_1, we have this representation, and that's the one
> the pass_stdarg optimization operates on:
> ...
>   # .MEM_2 = VDEF <.MEM_1(D)>
>   # USE = nonlocal escaped
>   # CLB = nonlocal escaped { D.1809 }
>   __builtin_va_startD.1021 (&apD.1809, 0);
>
>   # VUSE <.MEM_2>
>   # PT = nonlocal
>   ap.0_3 = apD.1809;
>
>   # .MEM_4 = VDEF <.MEM_2>
>   apD.1820 = ap.0_3;
>
>   # VUSE <.MEM_4>
>   ap.4_9 = apD.1820;
>
>   ap.5_10 = ap.4_9 + 4;
>
>   # .MEM_11 = VDEF <.MEM_4>
>   apD.1820 = ap.5_10;
>
>   # VUSE <.MEM_11>
>   _7 = MEM[(intD.1 *)ap.4_9];
> ...
>
> The optimization in pass_stdarg fails:
> ...
> f2: va_list escapes 1, needs to save all GPR units and all FPR units.
> ...
>
> The optimization fails because this assignment makes the va_list escape:
> ...
> va_list escapes in # .MEM_4 = VDEF <.MEM_2>
> apD.1820 = ap.0_3;
> ...
>
>
> II.
>
> By recalculating address_taken after expanding the ifn_va_arg, we get
> instead:
> ...
>   # .MEM_2 = VDEF <.MEM_1(D)>
>   # USE = nonlocal escaped
>   # CLB = nonlocal escaped { D.1809 }
>   __builtin_va_startD.1021 (&apD.1809, 0);
>
>   # VUSE <.MEM_2>
>   # PT = nonlocal
>   ap.0_3 = apD.1809;
>
>   ap_11 = ap.0_3;
>
>   ap.4_9 = ap_11;
>
>   ap.5_10 = ap.4_9 + 4;
>
>   ap_4 = ap.5_10;
>
>   # VUSE <.MEM_2>
>   _7 = MEM[(intD.1 *)ap.4_9];
> ...
>
> and the pass_stdarg optimization succeeds now:
> ...
> f2: va_list escapes 0, needs to save 4 GPR units and all FPR units.
> ...
>
> Bootstrapped and reg-tested on x86_64 with and without -m32.
>
> OK for trunk?

As noted in one of the PRs I think that it is the proper time to
re-implement the stdarg optimization on the un-lowered form which
should also fix this.

Thanks,
Richard.

>
> Thanks,
> - Tom