qemu_arm_defconfig with LTO fails due to unaligned access

[Tom Rini]

On Wed, Mar 08, 2023 at 05:12:26AM +0100, Heinrich Schuchardt wrote:
> Hello Ilias, hello Tom,
> 
> Tom tried to run qemu_arm_defconfig with CONFIG_LTO=y in gitlab. This
> failed as shown in protocol
> https://source.denx.de/u-boot/u-boot/-/jobs/589913/raw
> 
> Executing 'HII database protocols'
> test_hii_database_new_package_list:
> data abort
> pc : [<7ff39b98>]          lr : [<7ff87328>]
> reloc pc : [<00000b98>]    lr : [<0004e328>]
> sp : 7edf8cc0  ip : 0000000c     fp : 7ffe60ec
> r10: 00000000  r9 : 7eef8eb0     r8 : 7ffe0d02
> r7 : 00000000  r6 : 7ef0f8c8     r5 : 7ffe0cf0  r4 : 7ffe0cb4
> r3 : 7ffe0cef  r2 : 00000000     r1 : ffffffff  r0 : 00000000
> Flags: nzcv  IRQs off  FIQs off  Mode SVC_32
> Code: e2403002 e3a00000 e1500001 012fff1e (e1f320b2)
> UEFI image [0x00000000:0xffffffff] '/\selftest'
> Resetting CPU ..
> 
> Debugging shows:
> 
> efi_hii_sibt_string_ucs2_block_next() calls u16_strnlen() for an
> unaligned u16 string. Here "ldrh r2, [r3, #2]!" is executed for
> unaligned r3. This should be allowable for SCTLR.A = 0.
> 
> When the crash occurs SCRLR has value 0xc5187f. SCTLR.A is bit 1 with
> value 1.
> 
> The implementation of allow_unaligned() in
> arch/arm/cpu/armv7 /sctlr.S should have set the flag to 0.
> arch/arm/cpu/armv7/sctlr.S is compiled (as demonstrated by adding #error
> to the code).
> 
> If I remove the weak implementation of allow_unaligned() in
> lib/efi_loader/efi_setup.c, the error does not occur.
> 
> Shouldn't building with LTO ignore the weak implementation?
> 
> If I add a printf() statement to the weak implemenation, the printf()
> command is not executed but
> 
> SCTLR 0xc5187d, SCTLR.A=0
> 
> The test passes as unaligned access is allowable.
> 
> I was building inside the Docker image with the GCC downloaded by
> buildman (gcc-12.2.0-nolibc/arm-linux-gnueabi).
> 
> To me this looks like a compiler issue.

Interesting, yes. It seems like it shouldn't be too hard to come up with
a condensed example where the assembly function isn't used but instead
the weak C function is.

And as a work-around, re-doing the code so that path_to_uefi() just
checks for ARM && !ARM64 before calling allow_unaligned() and not doing
the weak function trick should also be fine.

-- 
Tom
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 659 bytes
Desc: not available
URL: <https://lists.denx.de/pipermail/u-boot/attachments/20230308/42a71502/attachment.sig>