cpu.x86: cleanups

db4
Slava Pestov 2009-09-28 16:38:35 -05:00
parent f59b842d9b
commit 4202211293
2 changed files with 17 additions and 18 deletions

View File

@ -618,11 +618,15 @@ ALIAS: PINSRQ PINSRD
: MOVDQA ( dest src -- ) { HEX: 6f HEX: 7f } HEX: 66 2-operand-rm-mr-sse ; : MOVDQA ( dest src -- ) { HEX: 6f HEX: 7f } HEX: 66 2-operand-rm-mr-sse ;
: MOVDQU ( dest src -- ) { HEX: 6f HEX: 7f } HEX: f3 2-operand-rm-mr-sse ; : MOVDQU ( dest src -- ) { HEX: 6f HEX: 7f } HEX: f3 2-operand-rm-mr-sse ;
<PRIVATE
: 2shuffler ( indexes/mask -- mask ) : 2shuffler ( indexes/mask -- mask )
dup integer? [ first2 { 1 0 } bitfield ] unless ; dup integer? [ first2 { 1 0 } bitfield ] unless ;
: 4shuffler ( indexes/mask -- mask ) : 4shuffler ( indexes/mask -- mask )
dup integer? [ first4 { 6 4 2 0 } bitfield ] unless ; dup integer? [ first4 { 6 4 2 0 } bitfield ] unless ;
PRIVATE>
: PSHUFD ( dest src imm -- ) 4shuffler HEX: 70 HEX: 66 3-operand-rm-sse ; : PSHUFD ( dest src imm -- ) 4shuffler HEX: 70 HEX: 66 3-operand-rm-sse ;
: PSHUFLW ( dest src imm -- ) 4shuffler HEX: 70 HEX: f2 3-operand-rm-sse ; : PSHUFLW ( dest src imm -- ) 4shuffler HEX: 70 HEX: f2 3-operand-rm-sse ;
: PSHUFHW ( dest src imm -- ) 4shuffler HEX: 70 HEX: f3 3-operand-rm-sse ; : PSHUFHW ( dest src imm -- ) 4shuffler HEX: 70 HEX: f3 3-operand-rm-sse ;

View File

@ -579,12 +579,12 @@ MACRO: available-reps ( alist -- )
'[ _ cond ] ; '[ _ cond ] ;
: unsign-rep ( rep -- rep' ) : unsign-rep ( rep -- rep' )
dup { {
{ uint-4-rep int-4-rep } { uint-4-rep int-4-rep }
{ ulonglong-2-rep longlong-2-rep } { ulonglong-2-rep longlong-2-rep }
{ ushort-8-rep short-8-rep } { ushort-8-rep short-8-rep }
{ uchar-16-rep char-16-rep } { uchar-16-rep char-16-rep }
} at* [ nip ] [ drop ] if ; } ?at drop ;
M:: x86 %broadcast-vector ( dst src rep -- ) M:: x86 %broadcast-vector ( dst src rep -- )
rep unsign-rep { rep unsign-rep {
@ -592,22 +592,23 @@ M:: x86 %broadcast-vector ( dst src rep -- )
dst src float-4-rep %copy dst src float-4-rep %copy
dst dst { 0 0 0 0 } SHUFPS dst dst { 0 0 0 0 } SHUFPS
] } ] }
{ double-2-rep [ { double-2-rep [
dst src MOVDDUP dst src MOVDDUP
] } ] }
{ longlong-2-rep [ { longlong-2-rep [
dst src = [ dst src =
dst dst PUNPCKLQDQ [ dst dst PUNPCKLQDQ ]
] [ [ dst src { 0 1 0 1 } PSHUFD ]
dst src { 0 1 0 1 } PSHUFD if
] if
] } ] }
{ int-4-rep [ dst src { 0 0 0 0 } PSHUFD ] } { int-4-rep [
{ short-8-rep [ dst src { 0 0 0 0 } PSHUFD
] }
{ short-8-rep [
dst src { 0 0 0 0 } PSHUFLW dst src { 0 0 0 0 } PSHUFLW
dst dst PUNPCKLQDQ dst dst PUNPCKLQDQ
] } ] }
{ char-16-rep [ { char-16-rep [
dst src char-16-rep %copy dst src char-16-rep %copy
dst dst PUNPCKLBW dst dst PUNPCKLBW
dst dst { 0 0 0 0 } PSHUFLW dst dst { 0 0 0 0 } PSHUFLW
@ -619,13 +620,7 @@ M: x86 %broadcast-vector-reps
{ {
! Can't do this with sse1 since it will want to unbox ! Can't do this with sse1 since it will want to unbox
! a double-precision float and convert to single precision ! a double-precision float and convert to single precision
{ sse2? { { sse2? { float-4-rep double-2-rep longlong-2-rep ulonglong-2-rep int-4-rep uint-4-rep short-8-rep ushort-8-rep char-16-rep uchar-16-rep } }
float-4-rep double-2-rep
longlong-2-rep ulonglong-2-rep
int-4-rep uint-4-rep
short-8-rep ushort-8-rep
char-16-rep uchar-16-rep
} }
} available-reps ; } available-reps ;
M:: x86 %gather-vector-4 ( dst src1 src2 src3 src4 rep -- ) M:: x86 %gather-vector-4 ( dst src1 src2 src3 src4 rep -- )