This is the mail archive of the
gdb-patches@sources.redhat.com
mailing list for the GDB project.
[preliminary patch] sim/h8300/compile.c: Fix cmp.w.
- From: Kazu Hirata <kazu at cs dot umass dot edu>
- To: gdb-patches at sources dot redhat dot com
- Cc: msnyder at redhat dot com, rsandifo at redhat dot com, aoliva at redhat dot com, vinschen at redhat dot com, dvenkat at noida dot hcltech dot com
- Date: Tue, 24 Jun 2003 10:43:22 -0400 (EDT)
- Subject: [preliminary patch] sim/h8300/compile.c: Fix cmp.w.
Hi,
Attached is a preliminary patch to fix cmp.w.
The following testcase, derived from
gcc.c-torture/execute/20020307-1.c, fails.
.h8300h
.section .text
.align 1
.global _main
_main:
sub.l er0,er0
cmp.w #0x8000,r0
bls .L1
jsr @_abort
.L1:
jsr @_exit
.end
This is equivalent to
if ((unsigned short) 0x8000 >= (unsigned short) 0)
exit ();
else
abort ();
Inside the simulator, 0x8000 is somehow sign-extended to 0xffff8000,
and cmp.w computes the difference like so.
0 - 0xffff8000 = 0x00008000
This result won't set the carry flag, so bls jumps in the wrong way.
I am trying to track down where this strange sign extension occurs,
but I thought it may be a good idea to share the problem early. Any
thought?
Kazu Hirata
2003-06-24 Kazu Hirata <kazu@cs.umass.edu>
* compile.c (sim_resume): Mask the upper half of operands when
simulating cmp.w.
Index: compile.c
===================================================================
RCS file: /cvs/src/src/sim/h8300/compile.c,v
retrieving revision 1.31
diff -u -r1.31 compile.c
--- compile.c 19 Jun 2003 02:14:14 -0000 1.31
+++ compile.c 24 Jun 2003 14:29:51 -0000
@@ -2355,6 +2355,8 @@
goto end;
if (fetch (sd, &code->src, &ea))
goto end;
+ rd &= 0xffff;
+ ea &= 0xffff;
ea = -ea;
res = rd + ea;
goto just_flags_alu16;