Re: armv5 userland broken

[Rin Okuyama <rokuyama.rk%gmail.com@localhost>]

On 2021/06/22 13:50, Martin Husemann wrote:
> On Tue, Jun 22, 2021 at 06:28:27AM +0900, Rin Okuyama wrote:
> > Oops... Is this v5 or v5hf? Also, does userland fail even if
> > only kernel is compiled by GCC9?
>
> This is v5 (one of the machines is the one running regular tests, see
> https://www.NetBSD.org/~martin/evbarm-atf/dmesg.txt)
>
> I double checked with multiple clean builds and it is exactly reproducable.
>
> It works when I use base.tgz from a HAVE_GCC=9 build (kernel and everything
> else still from the gcc 10 build).
>
> I have a sh and core from the failing build, but it is note very helpfull,
> the crash seems to happen in ld.elf_so.

On WZERO3 (xscale), userland works if compiled for earmv4 even by GCC10.

One of the major differences b/w armv5 and armv4 is presence of LDRD and
STRD instructions, for which memory operands should be placed to 8-byte
boundaries.

GCC10 seems to use LDRD/STRD for unaligned operands (this is legal for
armv6 and later). Actually, if LDRD/STRD are forcibly disabled for armv5
with this patch,

----
Index: external/gpl3/gcc/dist/gcc/config/arm/arm.h
===================================================================
RCS file: /home/netbsd/src/external/gpl3/gcc/dist/gcc/config/arm/arm.h,v
retrieving revision 1.23
diff -p -u -r1.23 arm.h
--- external/gpl3/gcc/dist/gcc/config/arm/arm.h	29 Apr 2021 01:51:39 -0000	1.23
+++ external/gpl3/gcc/dist/gcc/config/arm/arm.h	22 Jun 2021 10:29:48 -0000
@@ -173,7 +173,8 @@ emission of floating point pcs attribute
 #define TARGET_THUMB1_ONLY		(TARGET_THUMB1 && !arm_arch_notm)
 
 #define TARGET_LDRD			(arm_arch5te && ARM_DOUBLEWORD_ALIGN \
-                                         && !TARGET_THUMB1)
+                                         && !TARGET_THUMB1 \
+					 && unaligned_access /* XXX */)
 
 #define TARGET_CRC32			(arm_arch_crc)
 
----

init(8) and sh(1) do not crash.

Also, as Martin pointed out, the crash is due to ld.elf_so(1); if only
ld.elf_so(1) is replaced by ``hacked GCC10 version'', userland built by
original GCC10 seems to work fine.

So, my conclusions at the moment are:

(1) GCC10 emits LDRD/STRD for unaligned memory operand for armv5,
(2) at least for -O3 (only used for ld.elf_so(1) in our tree).

I've not figured out yet which change in GCC10 causes this.

It would be nice if we have a tiny testcase which reproduces the problem.
However, I've still not found it...

Thanks,
rin