>Number:         8145
>Category:       port-alpha
>Synopsis:       egcs generates incorrect code resulting in unaligned accesses
>Confidential:   no
>Severity:       non-critical
>Priority:       medium
>Responsible:    port-alpha-maintainer (NetBSD/alpha Portmaster)
>State:          open
>Class:          sw-bug
>Submitter-Id:   net
>Arrival-Date:   Wed Aug  4 15:50:01 1999
>Last-Modified:
>Originator:     Der Beagle
>Organization:
	
>Release:        1.4
>Environment:
	
System: NetBSD sopwith.uucp 1.4C NetBSD 1.4C (SOPWITH) #1: Fri Jun 4 16:11:43 PDT 1999 snoopy@machtoo.uucp:/b/usr.src/current_1999_05_25/src/sys/arch/alpha/compile/SOPWITH alpha

	Note: the kernel is -current 1999-05-25, but the compiler is
	from the 1.4 release.

	sopwith cc -v
	Using builtin specs.
	gcc version egcs-2.91.60 19981201 (egcs-1.1.1 release)
	sopwith 


>Description:
	The compiler generates code which requires greater alignment
	than it should.

>How-To-Repeat:
	Compile Ghostscript using the unix-gcc makefile.
	Run the resulting binary and observe the unaligned access
	complaints from the kernel.

	sysctl -w machdep.unaligned_sigbus=1
	Run Ghostscript from gdb.
	Observe the following results:

		pid 8939 (gs): unaligned access: va=0x12039a58a pc=0x12005ae30 ra=0x12005b89c op=ldl

		Program received signal SIGBUS, Bus error.
		0x12005ae34 in interp (pi_ctx_p=0x1203164c0, pref=0x120191266, perror_object=0x1fffff2d8)
		    at /usr/app/src/gs5.87/./src/interp.c:883
		883         switch (r_type_xe(iref_packed)) {

	The author of Ghostscript (L. Peter Deutsch, ghost@aladdin.com) says:

		This is a compiler bug.

		iref_packed is declared as register const ref_packed *.
		ref_packed is defined in iref.h as ushort (16-bit).  Also,
		r_type_xe is defined in iref.h as casting the pointer to const ushort *.
		Therefore, the compiler should be generating code that assumes no more
		than ushort (2-byte) alignment.
		Indeed, the value of iref_packed at the point of the unaligned access
		is ushort (2-byte) aligned, a.k.a. even.
		However, the trap message says the opcode was ldl, which is a 4- or 8-byte
		load.

	Peter would like to be copied on any reply regarding this problem.

>Fix:
	
>Audit-Trail:
>Unformatted: