This is the mail archive of the
binutils@sourceware.org
mailing list for the binutils project.
Re: [PATCH, v2] x86-64: correct segment override prefix generation
>>> On 07.08.12 at 15:44, "H.J. Lu" <hjl.tools@gmail.com> wrote:
> On Tue, Aug 7, 2012 at 3:32 AM, Jan Beulich <JBeulich@suse.com> wrote:
>>>>> On 30.07.12 at 19:03, "H.J. Lu" <hjl.tools@gmail.com> wrote:
>>> Please provide a testcase to show the correct behavior.
>>
>> Here you go.
>>
>> Jan
>>
>> Despite them being ignored by the CPU, gas issues segment override
>> prefixes for other than FS/GS in 64-bit mode. If doing so at all, it
>> should clearly do this correctly. Determining the default segment,
>> however, requires to take into consideration RegRex (so far, RSP, RBP,
>> R12, and R13 were all treated equally here).
>>
>> gas/
>> 2012-08-07 Jan Beulich <jbeulich@suse.com>
>> * config/tc-i386-intel.c (build_modrm_byte): Split determining
>> default segment from figuring out encoding. Honor RegRex for
>> the former.
>>
>> gas/testsuite/
>> 2012-08-07 Jan Beulich <jbeulich@suse.com>
>>
>> * gas/i386/x86-64-segovr.{s,l}: New.
>> * gas/i386/i386.exp: Run new test.
>>
>> --- 2012-08-07/gas/config/tc-i386.c 2012-07-31 09:45:03.000000000 +0200
>> +++ 2012-08-07/gas/config/tc-i386.c 2012-08-07 12:13:39.000000000 +0200
>> @@ -5729,18 +5729,14 @@ build_modrm_byte (void)
>> i.sib.base = i.base_reg->reg_num;
>> /* x86-64 ignores REX prefix bit here to avoid decoder
>> complications. */
>> - if ((i.base_reg->reg_num & 7) == EBP_REG_NUM)
>> - {
>> + if (!(i.base_reg->reg_flags & RegRex)
>> + && (i.base_reg->reg_num == EBP_REG_NUM
>> + || i.base_reg->reg_num == ESP_REG_NUM))
>> default_seg = &ss;
>> - if (i.disp_operands == 0)
>> - {
>> - fake_zero_displacement = 1;
>> - i.types[op].bitfield.disp8 = 1;
>> - }
>> - }
>> - else if (i.base_reg->reg_num == ESP_REG_NUM)
>> + if (i.base_reg->reg_num == 5 && i.disp_operands == 0)
>
> Please use EBP_REG_NUM instead 5 here.
But that change was intentional - we're _not_ looking for EBP here,
we're looking for "EBP or R13". The previous use of EBP_REG_NUM
was part of why this was broken imo.
> OK with the EBP_REG_NUM change above if Linux x86-64 kernel
> compiles and runs.
Sure, that has been the case for many weeks already.
Jan