Re: [045/nnn] poly_int: REG_ARGS_SIZE

[Richard Sandiford <richard dot sandiford at linaro dot org>]

Andreas Schwab <schwab@linux-m68k.org> writes:
> This breaks gcc.dg/tls/opt-3.c, gcc.dg/tls/pr47715-3.c and
> gcc.dg/tls/struct-1.c on m68k:
>
> /daten/aranym/gcc/gcc-20171222/gcc/testsuite/gcc.dg/tls/opt-3.c:11:3:
> internal compiler error: in add_args_size_note, at rtlanal.c:2379
> 0xae7aa9 add_args_size_note(rtx_insn*, poly_int<1u, long>)
>         ../../gcc/rtlanal.c:2379
> 0x7ea4ca fixup_args_size_notes(rtx_insn*, rtx_insn*, poly_int<1u, long>)
>         ../../gcc/expr.c:4105
> 0x7f6a02 emit_single_push_insn
>         ../../gcc/expr.c:4225
> 0x7fa412 emit_push_insn(rtx_def*, machine_mode, tree_node*, rtx_def*,
> unsigned int, int, rtx_def*, poly_int<1u, long>, rtx_def*, rtx_def*,
> int, rtx_def*, bool)
>         ../../gcc/expr.c:4561
> 0x6b8976 store_one_arg
>         ../../gcc/calls.c:5694
> 0x6c15b1 expand_call(tree_node*, rtx_def*, int)
>         ../../gcc/calls.c:4030
> 0x7f0485 expand_expr_real_1(tree_node*, rtx_def*, machine_mode,
> expand_modifier, rtx_def**, bool)
>         ../../gcc/expr.c:10927
> 0x6d6c97 expand_expr
>         ../../gcc/expr.h:276
> 0x6d6c97 expand_call_stmt
>         ../../gcc/cfgexpand.c:2690
> 0x6d6c97 expand_gimple_stmt_1
>         ../../gcc/cfgexpand.c:3624
> 0x6d6c97 expand_gimple_stmt
>         ../../gcc/cfgexpand.c:3790
> 0x6d8058 expand_gimple_tailcall
>         ../../gcc/cfgexpand.c:3836
> 0x6d8058 expand_gimple_basic_block
>         ../../gcc/cfgexpand.c:5774
> 0x6dd62e execute
>         ../../gcc/cfgexpand.c:6403

Bah.  Looks like I need to update my scripts to use --enable-tls,
since I'd ended up with emultls for the m68k targets.

I think the assert is catching a pre-existing bug here.  If we pushed
a value that needs a call to something like __tls_get_addr, we ended
up with two different REG_ARGS_SIZE notes on the same instruction.

It seems to be OK for emit_single_push_insn to push something that
needs a call to __tls_get_addr:

      /* We have to allow non-call_pop patterns for the case
	 of emit_single_push_insn of a TLS address.  */
      if (GET_CODE (pat) != PARALLEL)
	return 0;

so I think the problem is in the way this is handled rather than the fact
that it occurs at all.

If we're pushing a value X that needs a call C to calculate, we'll
add REG_ARGS_SIZE notes to the pushes and pops for C as part of the
call sequence.  Then emit_single_push_insn calls fixup_args_size_notes
on the whole push sequence (the calculation of X, including C,
and the push of X itself).  This is where the double notes came from.
But emit_single_push_insn_1 adjusted stack_pointer_delta *before* the
push, so the notes added for C were relative to the situation after
the future push of X rather than before it.

Presumably this didn't matter in practice because the note added
second tended to trump the note added first.  But code is allowed to
walk REG_NOTES without having to disregard secondary notes.

This patch seems to fix it for me, but I'm not sure how best to test it.

Thanks,
Richard


2017-12-23  Richard Sandiford  <richard.sandiford@linaro.org>

gcc/
	* expr.c (fixup_args_size_notes): Check that any existing
	REG_ARGS_SIZE notes are correct, and don't try to re-add them.
	(emit_single_push_insn_1): Move stack_pointer_delta adjustment to...
	(emit_single_push_insn): ...here.

Index: gcc/expr.c
===================================================================
--- gcc/expr.c	2017-12-23 09:29:20.226338285 +0000
+++ gcc/expr.c	2017-12-23 09:29:45.783339673 +0000
@@ -4089,6 +4089,14 @@ fixup_args_size_notes (rtx_insn *prev, r
       if (!NONDEBUG_INSN_P (insn))
 	continue;
 
+      /* We might have existing REG_ARGS_SIZE notes, e.g. when pushing
+	 a call argument containing a TLS address that itself requires
+	 a call to __tls_get_addr.  The handling of stack_pointer_delta
+	 in emit_single_push_insn is supposed to ensure that any such
+	 notes are already correct.  */
+      rtx note = find_reg_note (insn, REG_ARGS_SIZE, NULL_RTX);
+      gcc_assert (!note || known_eq (args_size, get_args_size (note)));
+
       poly_int64 this_delta = find_args_size_adjust (insn);
       if (known_eq (this_delta, 0))
 	{
@@ -4102,7 +4110,8 @@ fixup_args_size_notes (rtx_insn *prev, r
       if (known_eq (this_delta, HOST_WIDE_INT_MIN))
 	saw_unknown = true;
 
-      add_args_size_note (insn, args_size);
+      if (!note)
+	add_args_size_note (insn, args_size);
       if (STACK_GROWS_DOWNWARD)
 	this_delta = -poly_uint64 (this_delta);
 
@@ -4126,7 +4135,6 @@ emit_single_push_insn_1 (machine_mode mo
   rtx dest;
   enum insn_code icode;
 
-  stack_pointer_delta += PUSH_ROUNDING (GET_MODE_SIZE (mode));
   /* If there is push pattern, use it.  Otherwise try old way of throwing
      MEM representing push operation to move expander.  */
   icode = optab_handler (push_optab, mode);
@@ -4213,6 +4221,14 @@ emit_single_push_insn (machine_mode mode
 
   emit_single_push_insn_1 (mode, x, type);
 
+  /* Adjust stack_pointer_delta to describe the situation after the push
+     we just performed.  Note that we must do this after the push rather
+     than before the push in case calculating X needs pushes and pops of
+     its own (e.g. if calling __tls_get_addr).  The REG_ARGS_SIZE notes
+     for such pushes and pops must not include the effect of the future
+     push of X.  */
+  stack_pointer_delta += PUSH_ROUNDING (GET_MODE_SIZE (mode));
+
   last = get_last_insn ();
 
   /* Notice the common case where we emitted exactly one insn.  */