Hi,
I found a bug while testing latest gcc sources in the picochip port.
It was small enough that i thought i could fix it myself.
In expr.c, function emit_push_insn, with strict_alignment, if "align"
was not as strong as what is needed for the mode, we now try move the
value to a properly aligned temp (gathered using assign_temp). When
this change was made, calls to emit_push_insn from calls.c have been
changed to properly emit "align". But, there is still a place in
calls.c, in function emit_library_call_value_1, lines ~3720, where we
have a call to emit_push_insn with PARM_BOUNDARY as the alignment.
Note that we pass a NULL_TREE as well. In usual targets,
PARM_BOUNDARY(generally 4 bytes) would probably suffice for parameters
passed on stack. Ours is a bit unusual in that we only *require* 2
byte alignment, but we enforce STRICT_ALIGNMENT.
In my opinion, the call to emit_push_insn should get the proper
alignment for the parameter and then use that for the parameter
"align", instead of the current PARM_BOUNDARY.
I see that Ian raised this issue when the original change was made.
http://gcc.gnu.org/ml/gcc-patches/2007-02/msg00716.html
I am submitting a patch with the change. I have tested it on our port
(which is quite unusual in its own respects) and i686-pc-linux-gnu
bootstrap and tested it.
Regards
Hari
Index: gcc/calls.c
===================================================================
--- gcc/calls.c (revision 132323)
+++ gcc/calls.c (working copy)
@@ -3660,6 +3660,7 @@
rtx val = argvec[argnum].value;
rtx reg = argvec[argnum].reg;
int partial = argvec[argnum].partial;
+ unsigned int parm_align = argvec[argnum].locate.boundary;
int lower_bound = 0, upper_bound = 0, i;
if (! (reg != 0 && partial == 0))
@@ -3721,7 +3722,7 @@
}
}
- emit_push_insn (val, mode, NULL_TREE, NULL_RTX, PARM_BOUNDARY,
+ emit_push_insn (val, mode, NULL_TREE, NULL_RTX, parm_align,
partial, reg, 0, argblock,
GEN_INT (argvec[argnum].locate.offset.constant),
reg_parm_stack_space,