On Thu, Mar 31, 2011 at 01:35:17PM -0500, Peter Bergner wrote:
> Alan,
>
> Mainline seems to be broken on bootstrap when GCC is configured with
> LD=/path/to/ld/that/supports/Large-TOC. I'm attaching a preprocessed
> source file from the build and it compiles fine with -mcmodel=small,
> but ICE's when using -mcmodel={medium,large}. Can you please have
> a look? Thanks.
So, dwarf2out.c:gen_subprogram_die calls mem_loc_descriptor with the
lo_sum in the following
(gdb) p debug_rtx(call_arg_locations->call_arg_loc_note)
(note 38 22 37 (expr_list:REG_DEP_TRUE (concat:DI (reg:DI 3 3)
(lo_sum:DI (plus:DI (reg:DI 2 2)
(high:DI (unspec:DI [
(const_int 0 [0])
] UNSPEC_TLSLD)))
(unspec:DI [
(const_int 0 [0])
] UNSPEC_TLSLD)))
(nil)) NOTE_INSN_CALL_ARG_LOCATION)
which is no doubt from the rs6000 tls_ld patterns. mem_loc_descriptor
ignores the first arg of the lo_sum (fair enough), and passes the
second arg, the unspec, to const_ok_for_output. const_ok_for_output
says it is ok, because rs6000 does not wrap the unspec in a const..
It seems to me that mem_loc_descriptor is making unwarranted
assumptions about the contents of a lo_sum. Easily fixed.
OK to apply mainline, assuming my bootstrap and regression test
passes?
* dwarf2out.c (mem_loc_descriptor): Recurse on LO_SUM.
Index: gcc/dwarf2out.c
===================================================================
--- gcc/dwarf2out.c (revision 171817)
+++ gcc/dwarf2out.c (working copy)
@@ -13698,9 +13698,7 @@ mem_loc_descriptor (rtx rtl, enum machin
break;
case LO_SUM:
- rtl = XEXP (rtl, 1);
-
- /* ... fall through ... */
+ return mem_loc_descriptor (XEXP (rtl, 1), mode, initialized);
case LABEL_REF:
/* Some ports can transform a symbol ref into a label ref, because
--
Alan Modra
Australia Development Lab, IBM
