compiler.cfg.value-numbering: new optimizations; reassociation for shifts and redistribution for shifts/multiplies over additions/subtractions

db4
Slava Pestov 2010-04-23 04:17:41 -04:00
parent 2ca8d543f1
commit 2ce926594d
3 changed files with 257 additions and 7 deletions

View File

@ -1,7 +1,8 @@
! Copyright (C) 2010 Slava Pestov.
! See http://factorcode.org/license.txt for BSD license.
USING: accessors combinators cpu.architecture fry kernel layouts
math sequences compiler.cfg.instructions
locals make math sequences compiler.cfg.instructions
compiler.cfg.registers
compiler.cfg.value-numbering.expressions
compiler.cfg.value-numbering.folding
compiler.cfg.value-numbering.graph
@ -9,10 +10,12 @@ compiler.cfg.value-numbering.rewrite
compiler.cfg.value-numbering.simplify ;
IN: compiler.cfg.value-numbering.math
: f-expr? ( expr -- ? ) T{ reference-expr f f } = ;
M: ##tagged>integer rewrite
[ dst>> ] [ src>> vreg>expr ] bi {
{ [ dup integer-expr? ] [ value>> tag-fixnum \ ##load-integer new-insn ] }
{ [ dup reference-expr? ] [ value>> [ drop f ] [ \ f type-number \ ##load-integer new-insn ] if ] }
{ [ dup f-expr? ] [ \ f type-number \ ##load-integer new-insn ] }
[ 2drop f ]
} cond ;
@ -22,13 +25,22 @@ M: ##neg rewrite
M: ##not rewrite
dup unary-constant-fold? [ unary-constant-fold ] [ drop f ] if ;
: reassociate ( insn -- dst src1 src2 )
! Reassociation converts
! ## *-imm 2 1 X
! ## *-imm 3 2 Y
! into
! ## *-imm 3 1 (X $ Y)
! If * is associative, then $ is the same operation as *.
! In the case of shifts, $ is addition.
: (reassociate) ( insn -- dst src1 src2' src2'' )
{
[ dst>> ]
[ src1>> vreg>expr [ src1>> vn>vreg ] [ src2>> vn>integer ] bi ]
[ src2>> ]
[ ]
} cleave binary-constant-fold* ;
} cleave ; inline
: reassociate ( insn -- dst src1 src2 )
[ (reassociate) ] keep binary-constant-fold* ;
: ?new-insn ( dst src1 src2 ? class -- insn/f )
'[ _ new-insn ] [ 3drop f ] if ; inline
@ -39,6 +51,9 @@ M: ##not rewrite
: reassociate-bitwise ( insn new-insn -- insn/f )
[ reassociate dup immediate-bitwise? ] dip ?new-insn ; inline
: reassociate-shift ( insn new-insn -- insn/f )
[ (reassociate) + dup immediate-shift-count? ] dip ?new-insn ; inline
M: ##add-imm rewrite
{
{ [ dup binary-constant-fold? ] [ binary-constant-fold ] }
@ -56,24 +71,57 @@ M: ##sub-imm rewrite
[ sub-imm>add-imm ]
} cond ;
! Convert ##mul-imm -1 => ##neg
: mul-to-neg? ( insn -- ? )
src2>> -1 = ;
: mul-to-neg ( insn -- insn' )
[ dst>> ] [ src1>> ] bi \ ##neg new-insn ;
! Convert ##mul-imm 2^X => ##shl-imm X
: mul-to-shl? ( insn -- ? )
src2>> power-of-2? ;
: mul-to-shl ( insn -- insn' )
[ [ dst>> ] [ src1>> ] bi ] [ src2>> log2 ] bi \ ##shl-imm new-insn ;
! Distribution converts
! ##+-imm 2 1 X
! ##*-imm 3 2 Y
! Into
! ##*-imm 4 1 Y
! ##+-imm 3 4 X*Y
! Where * is mul or shl, + is add or sub
! Have to make sure that X*Y fits in an immediate
:: (distribute) ( insn expr imm temp add-op mul-op -- new-insns/f )
imm immediate-arithmetic? [
[
temp expr src1>> vn>vreg insn src2>> mul-op execute
insn dst>> temp imm add-op execute
] { } make
] [ f ] if ;
: distribute-over-add? ( insn -- ? )
src1>> vreg>expr add-imm-expr? ;
: distribute-over-sub? ( insn -- ? )
src1>> vreg>expr sub-imm-expr? ;
: distribute ( insn add-op mul-op -- new-insns/f )
[
dup src1>> vreg>expr
2dup src2>> vn>integer swap [ src2>> ] keep binary-constant-fold*
next-vreg
] 2dip (distribute) ;
M: ##mul-imm rewrite
{
{ [ dup binary-constant-fold? ] [ binary-constant-fold ] }
{ [ dup mul-to-neg? ] [ mul-to-neg ] }
{ [ dup mul-to-shl? ] [ mul-to-shl ] }
{ [ dup src1>> vreg>expr mul-imm-expr? ] [ \ ##mul-imm reassociate-arithmetic ] }
{ [ dup distribute-over-add? ] [ \ ##add-imm \ ##mul-imm distribute ] }
{ [ dup distribute-over-sub? ] [ \ ##sub-imm \ ##mul-imm distribute ] }
[ drop f ]
} cond ;
@ -101,21 +149,31 @@ M: ##xor-imm rewrite
M: ##shl-imm rewrite
{
{ [ dup binary-constant-fold? ] [ binary-constant-fold ] }
{ [ dup src1>> vreg>expr shl-imm-expr? ] [ \ ##shl-imm reassociate-shift ] }
{ [ dup distribute-over-add? ] [ \ ##add-imm \ ##shl-imm distribute ] }
{ [ dup distribute-over-sub? ] [ \ ##sub-imm \ ##shl-imm distribute ] }
[ drop f ]
} cond ;
M: ##shr-imm rewrite
{
{ [ dup binary-constant-fold? ] [ binary-constant-fold ] }
{ [ dup src1>> vreg>expr shr-imm-expr? ] [ \ ##shr-imm reassociate-shift ] }
[ drop f ]
} cond ;
M: ##sar-imm rewrite
{
{ [ dup binary-constant-fold? ] [ binary-constant-fold ] }
{ [ dup src1>> vreg>expr sar-imm-expr? ] [ \ ##sar-imm reassociate-shift ] }
[ drop f ]
} cond ;
! Convert
! ##load-integer 2 X
! ##* 3 1 2
! Where * is an operation with an -imm equivalent into
! ##*-imm 3 1 X
: insn>imm-insn ( insn op swap? -- new-insn )
swap [
[ [ dst>> ] [ src1>> ] [ src2>> ] tri ] dip
@ -129,12 +187,17 @@ M: ##add rewrite
[ drop f ]
} cond ;
! ##sub 2 1 1 => ##load-integer 2 0
: subtraction-identity? ( insn -- ? )
[ src1>> ] [ src2>> ] bi [ vreg>vn ] bi@ eq? ;
: rewrite-subtraction-identity ( insn -- insn' )
dst>> 0 \ ##load-integer new-insn ;
! ##load-integer 1 0
! ##sub 3 1 2
! =>
! ##neg 3 2
: sub-to-neg? ( ##sub -- ? )
src1>> vn>expr expr-zero? ;

View File

@ -1125,6 +1125,189 @@ cpu x86.32? [
} value-numbering-step
] unit-test
[
{
T{ ##peek f 0 D 0 }
T{ ##shl-imm f 1 0 10 }
T{ ##shl-imm f 2 0 21 }
T{ ##replace f 2 D 0 }
}
] [
{
T{ ##peek f 0 D 0 }
T{ ##shl-imm f 1 0 10 }
T{ ##shl-imm f 2 1 11 }
T{ ##replace f 2 D 0 }
} value-numbering-step
] unit-test
[
{
T{ ##peek f 0 D 0 }
T{ ##shl-imm f 1 0 10 }
T{ ##shl-imm f 2 1 $[ cell-bits 1 - ] }
T{ ##replace f 2 D 0 }
}
] [
{
T{ ##peek f 0 D 0 }
T{ ##shl-imm f 1 0 10 }
T{ ##shl-imm f 2 1 $[ cell-bits 1 - ] }
T{ ##replace f 2 D 0 }
} value-numbering-step
] unit-test
[
{
T{ ##peek f 0 D 0 }
T{ ##sar-imm f 1 0 10 }
T{ ##sar-imm f 2 0 21 }
T{ ##replace f 2 D 0 }
}
] [
{
T{ ##peek f 0 D 0 }
T{ ##sar-imm f 1 0 10 }
T{ ##sar-imm f 2 1 11 }
T{ ##replace f 2 D 0 }
} value-numbering-step
] unit-test
[
{
T{ ##peek f 0 D 0 }
T{ ##sar-imm f 1 0 10 }
T{ ##sar-imm f 2 1 $[ cell-bits 1 - ] }
T{ ##replace f 2 D 0 }
}
] [
{
T{ ##peek f 0 D 0 }
T{ ##sar-imm f 1 0 10 }
T{ ##sar-imm f 2 1 $[ cell-bits 1 - ] }
T{ ##replace f 2 D 0 }
} value-numbering-step
] unit-test
[
{
T{ ##peek f 0 D 0 }
T{ ##shr-imm f 1 0 10 }
T{ ##shr-imm f 2 0 21 }
T{ ##replace f 2 D 0 }
}
] [
{
T{ ##peek f 0 D 0 }
T{ ##shr-imm f 1 0 10 }
T{ ##shr-imm f 2 1 11 }
T{ ##replace f 2 D 0 }
} value-numbering-step
] unit-test
[
{
T{ ##peek f 0 D 0 }
T{ ##shr-imm f 1 0 10 }
T{ ##shr-imm f 2 1 $[ cell-bits 1 - ] }
T{ ##replace f 2 D 0 }
}
] [
{
T{ ##peek f 0 D 0 }
T{ ##shr-imm f 1 0 10 }
T{ ##shr-imm f 2 1 $[ cell-bits 1 - ] }
T{ ##replace f 2 D 0 }
} value-numbering-step
] unit-test
[
{
T{ ##peek f 0 D 0 }
T{ ##shr-imm f 1 0 10 }
T{ ##sar-imm f 2 1 11 }
T{ ##replace f 2 D 0 }
}
] [
{
T{ ##peek f 0 D 0 }
T{ ##shr-imm f 1 0 10 }
T{ ##sar-imm f 2 1 11 }
T{ ##replace f 2 D 0 }
} value-numbering-step
] unit-test
! Distributive law
2 \ vreg-counter set-global
[
{
T{ ##peek f 0 D 0 }
T{ ##add-imm f 1 0 10 }
T{ ##shl-imm f 3 0 2 }
T{ ##add-imm f 2 3 40 }
T{ ##replace f 2 D 0 }
}
] [
{
T{ ##peek f 0 D 0 }
T{ ##add-imm f 1 0 10 }
T{ ##shl-imm f 2 1 2 }
T{ ##replace f 2 D 0 }
} value-numbering-step
] unit-test
[
{
T{ ##peek f 0 D 0 }
T{ ##add-imm f 1 0 10 }
T{ ##mul-imm f 4 0 3 }
T{ ##add-imm f 2 4 30 }
T{ ##replace f 2 D 0 }
}
] [
{
T{ ##peek f 0 D 0 }
T{ ##add-imm f 1 0 10 }
T{ ##mul-imm f 2 1 3 }
T{ ##replace f 2 D 0 }
} value-numbering-step
] unit-test
[
{
T{ ##peek f 0 D 0 }
T{ ##add-imm f 1 0 -10 }
T{ ##shl-imm f 5 0 2 }
T{ ##add-imm f 2 5 -40 }
T{ ##replace f 2 D 0 }
}
] [
{
T{ ##peek f 0 D 0 }
T{ ##sub-imm f 1 0 10 }
T{ ##shl-imm f 2 1 2 }
T{ ##replace f 2 D 0 }
} value-numbering-step
] unit-test
[
{
T{ ##peek f 0 D 0 }
T{ ##add-imm f 1 0 -10 }
T{ ##mul-imm f 6 0 3 }
T{ ##add-imm f 2 6 -30 }
T{ ##replace f 2 D 0 }
}
] [
{
T{ ##peek f 0 D 0 }
T{ ##sub-imm f 1 0 10 }
T{ ##mul-imm f 2 1 3 }
T{ ##replace f 2 D 0 }
} value-numbering-step
] unit-test
! Simplification
[
{

View File

@ -1,8 +1,9 @@
! Copyright (C) 2006, 2010 Slava Pestov.
! See http://factorcode.org/license.txt for BSD license.
USING: accessors arrays assocs generic kernel kernel.private
math memory namespaces make sequences layouts system hashtables
classes alien byte-arrays combinators words sets fry ;
math math.order memory namespaces make sequences layouts system
hashtables classes alien byte-arrays combinators words sets fry
;
IN: cpu.architecture
! Representations -- these are like low-level types
@ -523,6 +524,9 @@ M: object immediate-comparand? ( n -- ? )
[ drop f ]
} cond ;
: immediate-shift-count? ( n -- ? )
0 cell-bits 1 - between? ;
! What c-type describes the implicit struct return pointer for
! large structs?
HOOK: struct-return-pointer-type cpu ( -- c-type )