[ARM] Allow MOV/MOV.W to accept all possible immediates

[Szabolcs Nagy]

On 05/01/17 14:16, Jiong Wang wrote:
> On 05/01/17 13:59, Jiong Wang wrote:
>> On 05/01/17 10:49, Szabolcs Nagy wrote:
>>> adding the reporter to cc and rich because it broke musl.
>>>
>>> On 05/01/17 10:45, Szabolcs Nagy wrote:
>>>> On 01/11/16 12:21, Jiong Wang wrote:
>>>>> Hi,
>>>>>
>>>>>    AArch32 MOV Rd, #imm should accept all immediates which can be encoded into
>>>>> available encoding, A1/A2 for ARM, T1/T2/T3 for Thumb/Thumb2, and MOV is the
>>>>> preferred disassembly syntax.  See latest ARM Architecture Reference Manual for
>>>>> ARMv8-A, section F6.1.108 and for ARMv8-M, section C2.4.89.
>>>>>
>>>>>    The same for MOV.W under Thumb mode.  It should try all possible 32-bit Thumb
>>>>> encoding instead of T2 only.
>>>>>
>>>>>    This patch let MOV/MOV.W accept more immediate formats while their currently
>>>>> supported immediate formats are not affected, so there is no backward compatibility
>>>>> issue, also this patch haven't touched the disassembler.
>>>>>
>>>>>    I think this patch brings GAS closer to ARM Architecture Reference Manual.
>>>>>
>>>>>    OK for master?
>>>>>
>>>>> gas/
>>>>> 2016-11-01  Jiong Wang  <jiong.wang@arm.com>
>>>>>                            * config/tc-arm.c (SBIT_SHIFT): New.
>>>>>          (T2_SBIT_SHIFT): Likewise.
>>>>>          (t32_insn_ok): Return TRUE for MOV in ARMv8-M Baseline.
>>>>>          (md_apply_fix): Try UINT16 encoding when ARM/Thumb modified immediate
>>>>>          encoding failed.
>>>>>          * testsuite/gas/arm/archv6t2-bad.s: New error case.
>>>>>          * testsuite/gas/arm/archv6t2-bad.l: New error match.
>>>>>          * testsuite/gas/arm/archv6t2.s: New testcase.
>>>>>          * testsuite/gas/arm/archv6t2.d: New expected result.
>>>>>          * testsuite/gas/arm/archv8m.s: New testcase
>>>>>          * testsuite/gas/arm/archv8m-base.d: New expected result.
>>>>>          * testsuite/gas/arm/archv8m-main.d: Likewise.
>>>>>          * testsuite/gas/arm/archv8m-main-dsp-1.d: Likewise.
>>>> this caused
>>>>
>>>> $ cat bug.s
>>>> .syntax unified
>>>> .text
>>>> .arch armv7-a
>>>>     mov r3,#1
>>>>
>>>> .arch armv4t
>>>> .eabi_attribute 6,2
>>>>     bx lr
>>>>
>>>> $ arm-none-linux-gnueabihf-as -mthumb bug.s
>>>> bug.s: Assembler messages:
>>>> bug.s:4: Error: invalid constant (1) after fixup
>>>>
>>>> previously this worked, but i'm not sure if the mix
>>>> of .arch directives and .syntax unified is valid here.
>>
>> It looks to me the failure is caused by the second ".arch armv4t" overrides the
>> first ".arch armv7-a", therefore the arch feature is lowerd to armv4t in
>> assembler fixup stage.  While "mov r3, #1" under thumb mode is only allowed
>> with ARMv6T2.
>>
>> This patch tightend the check from
>>
>>       if (fixP->fx_r_type == BFD_RELOC_ARM_T32_IMMEDIATE
>>           || fixP->fx_r_type == BFD_RELOC_ARM_T32_ADD_IMM)
>>
>> into:
>>
>>       if ((fixP->fx_r_type == BFD_RELOC_ARM_T32_IMMEDIATE
>>            && ARM_CPU_HAS_FEATURE (cpu_variant, arm_ext_v6t2))
>>           || fixP->fx_r_type == BFD_RELOC_ARM_T32_ADD_IMM)
>>
>>
>> thus the sequences in bug.s is not allowed.
> 
> Does .object_arch works for you? It seems you want the final object arch adjusted.
> 
> https://sourceware.org/ml/binutils/2006-09/msg00054.html

musl got fixed to use .object_arch:

http://git.musl-libc.org/cgit/musl/commit/?id=b261a24256792177a5f0531dbb25cc6267220ca5

thanks.