RE: [PING] PR middle-end/95126: Expand small const structs as immediate constants

["Roger Sayle" <roger@nextmovesoftware.com>]

Hi Rainer,

> > The one experiment I'd like to be able to try, to investigate the
> > cause/cure of this, is:
> >
> > diff --git a/gcc/calls.cc b/gcc/calls.cc index a4336c1..05fdd24 100644
> > --- a/gcc/calls.cc
> > +++ b/gcc/calls.cc
> > @@ -2177,7 +2177,7 @@ load_register_parameters (struct arg_data *args,
> > int num_a
> >              VAR_DECL with a simple constructor, expand that constructor
> >              via a pseudo rather than read from (possibly misaligned)
> >              memory.  PR middle-end/95126.  */
> > -         else if (nregs == 1
> > +         else if (0 && nregs == 1
> >                    && partial == 0
> >                    && !args[i].pass_on_stack
> >                    && VAR_P (tree_value)
> >
> > My "small const structs" patch affected code generation in three
> > places, and identifying which of these is causing the miscompilation
> > issue for gnat will help narrow down the problem, or worst case allow
> > reverting less of the problematic patch.
> 
> I just tried this on i386-pc-solaris2.11: unfortunately, it made no
difference.

Awesome!  Very many thanks for trying this.  Alas it confirms that the patch
I'm currently spinning won't have any affect.  So by a process of
elimination,
the miscompilation must be triggered by one of the other two changes:

diff --git a/gcc/expr.cc b/gcc/expr.cc
index fb062dc..37f1405 100644
--- a/gcc/expr.cc
+++ b/gcc/expr.cc
@@ -4871,7 +4871,7 @@ emit_push_insn (rtx x, machine_mode mode, tree type,
rtx s
          /* If source is a constant VAR_DECL with a simple constructor,
              store the constructor to the stack instead of moving it.  */
          const_tree decl;
-         if (partial == 0
+         if (0 && partial == 0
              && MEM_P (xinner)
              && SYMBOL_REF_P (XEXP (xinner, 0))
              && (decl = SYMBOL_REF_DECL (XEXP (xinner, 0))) != NULL_TREE
@@ -10603,7 +10603,7 @@ expand_expr_real_1 (tree exp, rtx target,
machine_mode t
        }
       /* Expand const VAR_DECLs with CONSTRUCTOR initializers that
         have scalar integer modes to a reg via store_constructor.  */
-      if (TREE_READONLY (exp)
+      if (0 && TREE_READONLY (exp)
          && !TREE_SIDE_EFFECTS (exp)
          && (modifier == EXPAND_NORMAL || modifier == EXPAND_STACK_PARM)
          && immediate_const_ctor_p (DECL_INITIAL (exp))


p.s.  I've just set up a build environment on another machine where I'm
using 
GNAT 11.3.1 20220421 (Red Hat 11.3.1-2)
as the bootstrap compiler, but it'll be a few hours before I'm investigating
in
gdb.

Many thanks again for your help.  Sorry again for the
breakage/inconvenience.

Roger
==