Re: [Patch] Fix bogus 'function does return' warning

[Richard Guenther <richard.guenther@gmail.com>]

On Thu, Oct 15, 2009 at 9:48 PM, Eric Botcazou <ebotcazou@adacore.com> wrote:
> Hi,
>
> the GIMPLE EH patch introduced a regression for the attached testcase: a
> bogus 'function does return' is now issued.
>
> The problem is that the EH pass turns
>
>      D.2142 = &"p.adb"[0];
>      .gnat_rcheck_20 (D.2142, 7);
>      <D.2137>:
>      D.2143 = &"p.adb"[0];
>      .gnat_rcheck_04 (D.2143, 9);
>      goto <D.2151>;
>    }
>  finally
>    {
>      p.error (); [static-chain: &FRAME.8]
>    }
>  <D.2151>:
>  return;
>
> into
>
>  D.2142 = &"p.adb"[0];
>  .gnat_rcheck_20 (D.2142, 7);
>  <D.2137>:
>  D.2143 = &"p.adb"[0];
>  .gnat_rcheck_04 (D.2143, 9);
>  finally_tmp.10 = 0;
>  goto <D.2156>;
>  <D.2156>:
>  p.error (); [static-chain: &FRAME.8]
>  switch (finally_tmp.10) <default: <D.2159>, default: <D.2159>, case 1:
> <D.2158>>
>  <D.2159>:
>  goto <D.2151>;
>  <D.2151>:
>  return;
>  <D.2157>:
>  finally_tmp.10 = 1;
>  goto <D.2156>;
>  <D.2158>:
>  resx 1
>
> thus masking the fact that D.2159 and D.2151 are unreachable.  So the CFG ends
> up being
>
> <bb 21>:
>  D.2143 = &"p.adb"[0];
>  .gnat_rcheck_04 (D.2143, 9);
>
> <L17>:
>  return;
>
> <L19>:
>  finally_tmp.10 = 1;
>  p.error (); [static-chain: &FRAME.8]
>  switch (finally_tmp.10) <default: <L17>, case 1: <L20>>
>
> <L20>:
>  resx 1
>
> and the EXIT block has a predecessor, hence the warning.
>
>
> The proposed fix is to remove unreachable statements after noreturn calls
> during the "lower" pass.  Tested on i586-suse-linux, OK for mainline?

Huh.  I wonder why the unreachable blocks are not removed
after CFG build?  After all noreturn calls end BBs.

Am I missing sth?

Richard.

>
> 2009-10-15  Eric Botcazou  <ebotcazou@adacore.com>
>
>        * gimple-low.c (lower_stmt) <GIMPLE_CALL>: If the call is noreturn,
>        remove subsequent regular statements in the sequence.
>
>
> 2009-10-15  Eric Botcazou  <ebotcazou@adacore.com>
>
>        * gnat.dg/noreturn1.ad[sb]: New test.
>
>
> --
> Eric Botcazou
>