llvm dev - Nov 2010 - [LLVMdev] Pesudo X86 instructions used for generating constants

Hi,

I noticed a bunch of psuedo instructions used for creation of constants without 
generating loads. e.g. pxor xmm0, xmm0
Here is an example of what i am referring to snipped from X86InstrSSE.td:

def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins), "",
                 [(set FR32:$dst, fp32imm0)]>,
                 Requires<[HasSSE1]>, TB, OpSize;

My question is why was there a need to define such a pseudo instruction?
Wouldn't it be cleaner to use a def: Pat<> which selects: 
pxor $dst, $dst 
from:
[(set FR32:$dst, fp32imm0)] ?

Thanks in advance.


      
-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20101114/6789ae65/attachment.html>

On Nov 14, 2010, at 8:20 AM, zr wrote:
> Hi,
> 
> I noticed a bunch of psuedo instructions used for creation of constants
without generating loads. e.g. pxor xmm0, xmm0
> Here is an example of what i am referring to snipped from X86InstrSSE.td:
> 
> def FsFLD0SS : I<0xEF, MRMInitReg, (outs FR32:$dst), (ins),
"",
>                  [(set FR32:$dst, fp32imm0)]>,
>                  Requires<[HasSSE1]>, TB, OpSize;
> 
> My question is why was there a need to define such a pseudo instruction?
> Wouldn't it be cleaner to use a def: Pat<> which selects: 
> pxor $dst, $dst 
> from:
> [(set FR32:$dst, fp32imm0)] ?
Yes, that would be cleaner, except that "xor" and "clear
register" have different side effects from the code generator perspective. 
"Clear register" has no input register operand, but "xor"
does.  This difference is eliminated after codegen is done and the pseudo
instruction is lowered by MCInstLowering.

-Chris 

-------------- next part --------------
An HTML attachment was scrubbed...
URL:
<http://lists.llvm.org/pipermail/llvm-dev/attachments/20101114/74df44fd/attachment.html>