steve
/
factor
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

compiler.cfg.instructions: ##select-vector instruction mapping to SSE4 PEXTR*

db4
Joe Groff 2010-05-16 00:49:12 -07:00
parent 5b48cd2a63
commit 645164188d
4 changed files with 54 additions and 0 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

5
basis/compiler/cfg/instructions/instructions.factor
View File

@ -310,6 +310,11 @@ def: dst
use: src1/int-rep src2/int-rep src3/int-rep src4/int-rep
literal: rep ;
PURE-INSN: ##select-vector
def: dst/int-rep
use: src
literal: n rep ;
PURE-INSN: ##shuffle-vector
def: dst
use: src shuffle

1
basis/compiler/codegen/codegen.factor
View File

@ -184,6 +184,7 @@ CODEGEN: ##gather-vector-2 %gather-vector-2
CODEGEN: ##gather-vector-4 %gather-vector-4
CODEGEN: ##gather-int-vector-2 %gather-int-vector-2
CODEGEN: ##gather-int-vector-4 %gather-int-vector-4
CODEGEN: ##select-vector %select-vector
CODEGEN: ##shuffle-vector-imm %shuffle-vector-imm
CODEGEN: ##shuffle-vector-halves-imm %shuffle-vector-halves-imm
CODEGEN: ##shuffle-vector %shuffle-vector

3
basis/cpu/architecture/architecture.factor
View File

@ -297,6 +297,7 @@ HOOK: %gather-vector-2 cpu ( dst src1 src2 rep -- )
HOOK: %gather-int-vector-2 cpu ( dst src1 src2 rep -- )
HOOK: %gather-vector-4 cpu ( dst src1 src2 src3 src4 rep -- )
HOOK: %gather-int-vector-4 cpu ( dst src1 src2 src3 src4 rep -- )
HOOK: %select-vector cpu ( dst src n rep -- )
HOOK: %shuffle-vector cpu ( dst src shuffle rep -- )
HOOK: %shuffle-vector-imm cpu ( dst src shuffle rep -- )
HOOK: %shuffle-vector-halves-imm cpu ( dst src1 src2 shuffle rep -- )
@ -354,6 +355,7 @@ HOOK: %gather-vector-2-reps cpu ( -- reps )
HOOK: %gather-int-vector-2-reps cpu ( -- reps )
HOOK: %gather-vector-4-reps cpu ( -- reps )
HOOK: %gather-int-vector-4-reps cpu ( -- reps )
HOOK: %select-vector-reps cpu ( -- reps )
HOOK: %alien-vector-reps cpu ( -- reps )
HOOK: %shuffle-vector-reps cpu ( -- reps )
HOOK: %shuffle-vector-imm-reps cpu ( -- reps )
@ -405,6 +407,7 @@ M: object %gather-vector-2-reps { } ;
M: object %gather-int-vector-2-reps { } ;
M: object %gather-vector-4-reps { } ;
M: object %gather-int-vector-4-reps { } ;
M: object %select-vector-reps { } ;
M: object %alien-vector-reps { } ;
M: object %shuffle-vector-reps { } ;
M: object %shuffle-vector-imm-reps { } ;

45
basis/cpu/x86/x86.factor
View File

@ -781,6 +781,51 @@ M: x86.64 %gather-int-vector-2-reps
{ sse4.1? { longlong-2-rep ulonglong-2-rep } }
} available-reps ;
:: %select-vector-32 ( dst src n rep -- )
rep {
{ char-16-rep [
dst 32-bit-version-of src n PEXTRB
dst dst 8-bit-version-of MOVSX
] }
{ uchar-16-rep [
dst 32-bit-version-of src n PEXTRB
] }
{ short-8-rep [
dst 32-bit-version-of src n PEXTRW
dst dst 16-bit-version-of MOVSX
] }
{ ushort-8-rep [
dst 32-bit-version-of src n PEXTRW
] }
{ int-4-rep [
dst 32-bit-version-of src n PEXTRD
dst dst 32-bit-version-of 2dup = [ 2drop ] [ MOVSX ] if
] }
{ uint-4-rep [
dst 32-bit-version-of src n PEXTRD
] }
} case ;
M: x86.32 %select-vector
%select-vector-32 ;
M: x86.32 %select-vector-reps
{
{ sse4.1? { uchar-16-rep char-16-rep ushort-8-rep short-8-rep uint-4-rep int-4-rep } }
} available-reps ;
M: x86.64 %select-vector
{
{ longlong-2-rep [ PEXTRQ ] }
{ ulonglong-2-rep [ PEXTRQ ] }
[ %select-vector-32 ]
} case ;
M: x86.64 %select-vector-reps
{
{ sse4.1? { uchar-16-rep char-16-rep ushort-8-rep short-8-rep uint-4-rep int-4-rep ulonglong-2-rep longlong-2-rep } }
} available-reps ;
: sse1-float-4-shuffle ( dst shuffle -- )
{
{ { 0 1 2 3 } [ drop ] }