Re: PATCH: Improve debuggability at -O0 and fix PR 29609

["Richard Guenther" <richard dot guenther at gmail dot com>]

On 9/3/07, Tristan Gingold <gingold@adacore.com> wrote:
> Hi,
>
> this patch improve debuggability at -O0 and fixes PR 29609.
>
> For break/continue/goto statements directly inside an if statement the
> code generated is too optimized since tree-ssa.  On x86:
>
>      if (!f3())
>          goto failure; /* Line 23 */
>
> is translated to:
>
>          call    f3
>          testl   %eax, %eax
>          je      .L6
>
> and therefore there is no possible breakpoint at line 23 which is a mess
> at -O0.
>
> This patch first add a new flag to GOTO_EXPR: GOTO_EXPLICIT_P
> (field static_flag) so that gotos coming explicitely from user code
> can be
> tracked.  During gimple lower pass and only at -O0 the explicit gotos
> are
> not merged with the conditionnal jump created by an if statement.
>
> This is just enough to generate correct code but not yet enough to
> have to
> correct debug line opcode generated.  The jump insn may be removed by
> the
> into_cfglayout pass.  To avoid this issue a new flag is added to edge
> flags:
> EDGE_EXPLICIT (only at -O0).  When this flag is set cfgrtl doesn't
> remove
> the jump and cfgcleanup does not try to remove this forwarding edge.
>
> Now the code generated at -O0 -g is:
>
>          .loc 1 22 0
>          call    f3
>          testl   %eax, %eax
>          jne     .L6
>          .loc 1 23 0
>          jmp     .L7
>

I don't like this approach.  From a quick check I see that gimple
lowering retains
location information for both the if and the then branch correctly in
lower_cond_expr
and the CFG created has the correct locus set on the edge already.  So where
is this information lost?  I suppose it may be during expand where we call
do_jump via jumpif but the complete jump sequence has one location.  So the
proper fix would probably to teach do_jump to take an alternate location for the
jump.

Richard.