On Wed, Dec 20, 2017 at 11:28:27AM -0800, Junaid Shahid wrote:
> > > + # adjust the shuffle mask pointer to be able to shift either 0 or 8
> > > + # bytes depending on whether the last block is <8 bytes or not
> > > + mov \DLEN, \TMP1
> > > + and $8, \TMP1
> > > + lea SHIFT_MASK(%rip), %rax
> > > + sub \TMP1, %rax
> > > + movdqu (%rax), \XMM2 # get the appropriate shuffle mask
> > > + PSHUFB_XMM \XMM2, \XMM1 # shift left either 0 or 8 bytes
> >
> > Is there any way this can be simplified to use pslldq? The shift is just 8
> > bytes or nothing, and we already had to branch earlier depending on whether
> > we
> > needed to read the 8 bytes or not.
>
> I am not sure if we can use a simple pslldq without either introducing another
> branch, or duplicating the _read_last_lt8 block for each case of the original
> branch. Do you think that it is better to just duplicate the _read_last_lt8
> block instead of using pshufb? Or do you have any other suggestion about how
> to do it?
>
The best I can come up with now is just duplicating the "read one byte at a
time" instructions (see below). One way to avoid the duplication would be to
read the 1-7 byte remainder (end of the block) first and the 8-byte word
(beginning of the block) second, but it would be a bit weird.
# read the last <16 byte block
# Clobbers %rax, TMP1 and XMM1
.macro READ_PARTIAL_BLOCK DPTR DLEN TMP1 XMM1 XMMDst
mov \DLEN, \TMP1
cmp $8, \DLEN
jl _read_partial_lt8_\@
mov (\DPTR), %rax
MOVQ_R64_XMM %rax, \XMMDst
sub $8, \TMP1
jz _done_read_partial_\@
xor %rax, %rax
1:
shl $8, %rax
mov 7(\DPTR, \TMP1, 1), %al
dec \TMP1
jnz 1b
MOVQ_R64_XMM %rax, \XMM1
pslldq $8, \XMM1
por \XMM1, \XMMDst
jmp _done_read_partial_\@
_read_partial_lt8_\@:
xor %rax, %rax
1:
shl $8, %rax
mov -1(\DPTR, \TMP1, 1), %al
dec \TMP1
jnz 1b
MOVQ_R64_XMM %rax, \XMMDst
_done_read_partial_\@:
.endm
