This is an ICE for g++.dg/pr48660.C on HP-PA/Linux introduced by:
http://gcc.gnu.org/ml/gcc-patches/2012-10/msg00745.html
When I wrote it, I convinced myself that the HP-PA wasn't using the BLKmode
registers path any more and wrapped everything in PARALLELs for return values,
so it tested the change on MIPS instead. Now it turns out that, for sub-word
structures, the HP-PA still uses BLKmode registers and the C++ testcase shows
that the code doesn't correctly handle assignments from incoming return values
to outgoing return values. Not clear why this didn't show up on MIPS...
Tested on x86-64, HP-PA and MIPS/Linux, applied on the mainline.
2012-12-16 Eric Botcazou <ebotca...@adacore.com>
PR rtl-optimization/55630
* expr.c (expand_assignment): Do not call copy_blkmode_to_reg to move
BLKmode return values to the return register.
--
Eric Botcazou
Index: expr.c
===================================================================
--- expr.c (revision 194517)
+++ expr.c (working copy)
@@ -4920,7 +4920,12 @@ expand_assignment (tree to, tree from, b
rtx temp;
push_temp_slots ();
- if (REG_P (to_rtx) && TYPE_MODE (TREE_TYPE (from)) == BLKmode)
+
+ /* If the source is itself a return value, it still is in a pseudo at
+ this point so we can move it back to the return register directly. */
+ if (REG_P (to_rtx)
+ && TYPE_MODE (TREE_TYPE (from)) == BLKmode
+ && TREE_CODE (from) != CALL_EXPR)
temp = copy_blkmode_to_reg (GET_MODE (to_rtx), from);
else
temp = expand_expr (from, NULL_RTX, GET_MODE (to_rtx), EXPAND_NORMAL);