Inline assembly - how to get gcc to clear the full rcx register in x86-64 mode

Inline assembly - how to get gcc to clear the full rcx register in x86-64 mode

[Jeroen van Bemmel]

Hi,

I have ported an SSE4 strcmp function from 
http://www.strchr.com/strcmp_and_strlen_using_sse_4.2
to GCC inline assembly:

long __res;
__asm__ __volatile__(
         "sub        $16, %4                    \n"
         "1:\n"
         "add        $16, %4                     \n"
         "movdqu        (%4), %%xmm0 \n"    // Could use any XMM, using 
register constraint "x"
         // ".byte 0x48                           \n"    // REX prefix 
with REX.w=1, to get result in RCX
         "pcmpistri    $0x18, (%4,%0), %%xmm0  \n"    // 
EQUAL_EACH(0x08) + NEGATIVE_POLARITY(0x10)
         "ja 1b                                \n"
         "jc 2f                                 \n"
         "xor %0, %0                     \n"
         "jmp 3f                              \n"    // XXX Extra jump 
could be avoided in pure asm
         "2:\n"
         "add %4, %0                     \n"
         "movzxb (%0,%1), %0      \n"
         "movzxb (%4,%1), %4      \n"
         "sub %4, %0                     \n"
         "3:\n"
     : "=a"(__res), "=c"(cs) : "0"(cs-ct), "1"(0L), "r"(ct) : "xmm0" );

     return (int) __res;

The problem with this code is that "pcmpistri" returns its result in ECX 
(i.e. the lower 32 bits of RCX), while the "movzxb" instructions use the 
full RCX register.
One solution is to insert a REX prefix with REX.w bit set ( any gas 
directive for this? )

However, I'd prefer to have gcc clear RCX at the beginning of the 
function. The above code loads the "c" register with 0, but the 
resulting asm code is
"xorl    ecx, ecx"

Is this a bug in GCC? Or how do I get it to clear the full RCX, without 
doing it 'manually' in the asm block?

Thanks,
Jeroen

Re: Inline assembly - how to get gcc to clear the full rcx register in x86-64 mode

[Ian Lance Taylor]

Jeroen van Bemmel <jbemmel@zonnet.nl> writes:

> I have ported an SSE4 strcmp function from
> http://www.strchr.com/strcmp_and_strlen_using_sse_4.2
> to GCC inline assembly:
>
> long __res;
> __asm__ __volatile__(
>         "sub        $16, %4                    \n"
>         "1:\n"
>         "add        $16, %4                     \n"
>         "movdqu        (%4), %%xmm0 \n"    // Could use any XMM, using
> register constraint "x"
>         // ".byte 0x48                           \n"    // REX prefix
> with REX.w=1, to get result in RCX
>         "pcmpistri    $0x18, (%4,%0), %%xmm0  \n"    //
> EQUAL_EACH(0x08) + NEGATIVE_POLARITY(0x10)
>         "ja 1b                                \n"
>         "jc 2f                                 \n"
>         "xor %0, %0                     \n"
>         "jmp 3f                              \n"    // XXX Extra jump
> could be avoided in pure asm
>         "2:\n"
>         "add %4, %0                     \n"
>         "movzxb (%0,%1), %0      \n"
>         "movzxb (%4,%1), %4      \n"
>         "sub %4, %0                     \n"
>         "3:\n"
>     : "=a"(__res), "=c"(cs) : "0"(cs-ct), "1"(0L), "r"(ct) : "xmm0" );
>
>     return (int) __res;
>
> The problem with this code is that "pcmpistri" returns its result in
> ECX (i.e. the lower 32 bits of RCX), while the "movzxb" instructions
> use the full RCX register.
> One solution is to insert a REX prefix with REX.w bit set ( any gas
> directive for this? )

Normally setting the low 32 bits of an x86 register will zero out the
upper 32 bits.  Is that not true for pcmpistri?

Otherwise, it sounds like you want the addressing mode (%rax,%ecx).
Does x86 really have that addressing mode?

Why not just zero extend %ecx to %rcx?

> However, I'd prefer to have gcc clear RCX at the beginning of the
> function. The above code loads the "c" register with 0, but the
> resulting asm code is
> "xorl    ecx, ecx"

That instruction will indeed set %ecx to zero.  Think about it.

Ian

Re: Inline assembly - how to get gcc to clear the full rcx register in x86-64 mode

[Jeroen van Bemmel]

Hi Ian,

I was confused by the Intel documentation, which states that pcmpistri 
provides its result in ECX. I stepped through a debugger, and verified 
that the upper 32 bits in RCX indeed get set to 0.

I also found that while I cannot distinguish between 32-bit and 64-bit 
forms of registers in constraints, GCC will choose the optimal flavor 
for me.

Thanks,
Jeroen

On 06/13/12 18:37, Ian Lance Taylor wrote:
> Jeroen van Bemmel<jbemmel@zonnet.nl>  writes:
>
>> I have ported an SSE4 strcmp function from
>> http://www.strchr.com/strcmp_and_strlen_using_sse_4.2
>> to GCC inline assembly:
>>
>> long __res;
>> __asm__ __volatile__(
>>          "sub        $16, %4                    \n"
>>          "1:\n"
>>          "add        $16, %4                     \n"
>>          "movdqu        (%4), %%xmm0 \n"    // Could use any XMM, using
>> register constraint "x"
>>          // ".byte 0x48                           \n"    // REX prefix
>> with REX.w=1, to get result in RCX
>>          "pcmpistri    $0x18, (%4,%0), %%xmm0  \n"    //
>> EQUAL_EACH(0x08) + NEGATIVE_POLARITY(0x10)
>>          "ja 1b                                \n"
>>          "jc 2f                                 \n"
>>          "xor %0, %0                     \n"
>>          "jmp 3f                              \n"    // XXX Extra jump
>> could be avoided in pure asm
>>          "2:\n"
>>          "add %4, %0                     \n"
>>          "movzxb (%0,%1), %0      \n"
>>          "movzxb (%4,%1), %4      \n"
>>          "sub %4, %0                     \n"
>>          "3:\n"
>>      : "=a"(__res), "=c"(cs) : "0"(cs-ct), "1"(0L), "r"(ct) : "xmm0" );
>>
>>      return (int) __res;
>>
>> The problem with this code is that "pcmpistri" returns its result in
>> ECX (i.e. the lower 32 bits of RCX), while the "movzxb" instructions
>> use the full RCX register.
>> One solution is to insert a REX prefix with REX.w bit set ( any gas
>> directive for this? )
> Normally setting the low 32 bits of an x86 register will zero out the
> upper 32 bits.  Is that not true for pcmpistri?
>
> Otherwise, it sounds like you want the addressing mode (%rax,%ecx).
> Does x86 really have that addressing mode?
>
> Why not just zero extend %ecx to %rcx?
>
>> However, I'd prefer to have gcc clear RCX at the beginning of the
>> function. The above code loads the "c" register with 0, but the
>> resulting asm code is
>> "xorl    ecx, ecx"
> That instruction will indeed set %ecx to zero.  Think about it.
>
> Ian
>