fold test-vector/branch sequences into a test-vector-branch instruction

diff --git a/basis/compiler/cfg/comparisons/comparisons.factor b/basis/compiler/cfg/comparisons/comparisons.factor
index d538ee818ce87c82a8a2cdfa8fb025f22238f824..0b4a6f2f02e561c1d3d6e87efa5279964f6121f6 100644 (file)
--- a/basis/compiler/cfg/comparisons/comparisons.factor
+++ b/basis/compiler/cfg/comparisons/comparisons.factor
@@ -10,7 +10,7 @@ SYMBOLS:
     cc/< cc/<= cc/= cc/> cc/>= cc/<> cc/<>= ;
 
 SYMBOLS:
-    vcc-all vcc-any vcc-none ;
+    vcc-all vcc-notall vcc-any vcc-none ;
 
 : negate-cc ( cc -- cc' )
     H{
@@ -30,6 +30,14 @@ SYMBOLS:
         { cc/<>= cc<>=  }
     } at ;
 
+: negate-vcc ( cc -- cc' )
+    H{
+        { vcc-all vcc-notall }
+        { vcc-any vcc-none }
+        { vcc-none vcc-any }
+        { vcc-notall vcc-all }
+    } at ;
+
 : swap-cc ( cc -- cc' )
     H{
         { cc<   cc> }

diff --git a/basis/compiler/cfg/instructions/instructions.factor b/basis/compiler/cfg/instructions/instructions.factor
index 09e773623584b3b0360a756ac45b33aa2409025e..d40aabcb1992b798929fff3ac20b24e4043eadca 100644 (file)
--- a/basis/compiler/cfg/instructions/instructions.factor
+++ b/basis/compiler/cfg/instructions/instructions.factor
@@ -304,7 +304,18 @@ literal: rep cc ;
 
 PURE-INSN: ##test-vector
 def: dst/int-rep
-use: src
+use: src1
+temp: temp/int-rep
+literal: rep vcc ;
+
+INSN: ##test-vector-branch
+use: src1
+temp: temp/int-rep
+literal: rep vcc ;
+
+INSN: _test-vector-branch
+literal: label
+use: src1
 temp: temp/int-rep
 literal: rep vcc ;
 

diff --git a/basis/compiler/cfg/linearization/linearization.factor b/basis/compiler/cfg/linearization/linearization.factor
index 66ac1addb0987b3928b299aa56367ebaba4c299b..31a424720620aa1962e162bf7617d0ebcb80d732 100755 (executable)
--- a/basis/compiler/cfg/linearization/linearization.factor
+++ b/basis/compiler/cfg/linearization/linearization.factor
@@ -2,7 +2,7 @@
 ! See http://factorcode.org/license.txt for BSD license.
 USING: kernel math accessors sequences namespaces make
 combinators assocs arrays locals layouts hashtables
-cpu.architecture
+cpu.architecture generalizations
 compiler.cfg
 compiler.cfg.comparisons
 compiler.cfg.stack-frame
@@ -42,14 +42,26 @@ M: ##branch linearize-insn
 
 : successors ( bb -- first second ) successors>> first2 ; inline
 
+:: conditional ( bb insn n conditional-quot negate-cc-quot -- bb successor label ... )
+    bb insn
+    conditional-quot
+    [ drop dup successors>> second useless-branch? ] 2bi
+    [ [ swap block-number ] n ndip ]
+    [ [ block-number ] n ndip negate-cc-quot call ] if ; inline
+
 : (binary-conditional) ( bb insn -- bb successor1 successor2 src1 src2 cc )
     [ dup successors ]
     [ [ src1>> ] [ src2>> ] [ cc>> ] tri ] bi* ; inline
 
 : binary-conditional ( bb insn -- bb successor label2 src1 src2 cc )
-    [ (binary-conditional) ]
-    [ drop dup successors>> second useless-branch? ] 2bi
-    [ [ swap block-number ] 3dip ] [ [ block-number ] 3dip negate-cc ] if ;
+    3 [ (binary-conditional) ] [ negate-cc ] conditional ;
+
+: (test-vector-conditional) ( bb insn -- bb successor1 successor2 src1 temp rep vcc )
+    [ dup successors ]
+    [ { [ src1>> ] [ temp>> ] [ rep>> ] [ vcc>> ] } cleave ] bi* ; inline
+
+: test-vector-conditional ( bb insn -- bb successor label src1 temp rep vcc )
+    4 [ (test-vector-conditional) ] [ negate-vcc ] conditional ;
 
 M: ##compare-branch linearize-insn
     binary-conditional _compare-branch emit-branch ;
@@ -63,6 +75,9 @@ M: ##compare-float-ordered-branch linearize-insn
 M: ##compare-float-unordered-branch linearize-insn
     binary-conditional _compare-float-unordered-branch emit-branch ;
 
+M: ##test-vector-branch linearize-insn
+    test-vector-conditional _test-vector-branch emit-branch ;
+
 : overflow-conditional ( bb insn -- bb successor label2 dst src1 src2 )
     [ dup successors block-number ]
     [ [ dst>> ] [ src1>> ] [ src2>> ] tri ] bi* ; inline

diff --git a/basis/compiler/cfg/value-numbering/rewrite/rewrite.factor b/basis/compiler/cfg/value-numbering/rewrite/rewrite.factor
index 8e5e013606d39761744ebf10a5276a99dbed1f4d..9827e02bf545d09b647952adc2573720687b75e6 100755 (executable)
--- a/basis/compiler/cfg/value-numbering/rewrite/rewrite.factor
+++ b/basis/compiler/cfg/value-numbering/rewrite/rewrite.factor
@@ -40,6 +40,7 @@ M: insn rewrite drop f ;
         [ compare-imm-expr? ]
         [ compare-float-unordered-expr? ]
         [ compare-float-ordered-expr? ]
+        [ test-vector-expr? ]
     } 1|| ;
 
 : rewrite-boolean-comparison? ( insn -- ? )
@@ -53,12 +54,21 @@ M: insn rewrite drop f ;
 : >compare-imm-expr< ( expr -- in1 in2 cc )
     [ src1>> vn>vreg ] [ src2>> vn>constant ] [ cc>> ] tri ; inline
 
+: >test-vector-expr< ( expr -- src1 temp rep vcc )
+    {
+        [ src1>> vn>vreg ]
+        [ drop next-vreg ]
+        [ rep>> ]
+        [ vcc>> ]
+    } cleave ; inline
+
 : rewrite-boolean-comparison ( expr -- insn )
     src1>> vreg>expr {
         { [ dup compare-expr? ] [ >compare-expr< \ ##compare-branch new-insn ] }
         { [ dup compare-imm-expr? ] [ >compare-imm-expr< \ ##compare-imm-branch new-insn ] }
         { [ dup compare-float-unordered-expr? ] [ >compare-expr< \ ##compare-float-unordered-branch new-insn ] }
         { [ dup compare-float-ordered-expr? ] [ >compare-expr< \ ##compare-float-ordered-branch new-insn ] }
+        { [ dup test-vector-expr? ] [ >test-vector-expr< \ ##test-vector-branch new-insn ] }
     } cond ;
 
 : tag-fixnum-expr? ( expr -- ? )

diff --git a/basis/compiler/cfg/value-numbering/value-numbering-tests.factor b/basis/compiler/cfg/value-numbering/value-numbering-tests.factor
index b2750da3faf49efcb6bf3493a69f1caba477e4e6..ba33885aacc13ef703c7b3eaee851a578d12a62f 100644 (file)
--- a/basis/compiler/cfg/value-numbering/value-numbering-tests.factor
+++ b/basis/compiler/cfg/value-numbering/value-numbering-tests.factor
@@ -14,6 +14,8 @@ IN: compiler.cfg.value-numbering.tests
             [ ##compare-imm? ]
             [ ##compare-float-unordered? ]
             [ ##compare-float-ordered? ]
+            [ ##test-vector? ]
+            [ ##test-vector-branch? ]
         } 1|| [ f >>temp ] when
     ] map ;
 
@@ -141,6 +143,22 @@ IN: compiler.cfg.value-numbering.tests
     } value-numbering-step trim-temps
 ] unit-test
 
+[
+    {
+        T{ ##peek f 1 D -1 }
+        T{ ##unbox-vector f 1111 1 float-4-rep }
+        T{ ##test-vector f 1 1111 f float-4-rep vcc-any }
+        T{ ##test-vector-branch f 1111 f float-4-rep vcc-any }
+    }
+] [
+    {
+        T{ ##peek f 1 D -1 }
+        T{ ##unbox-vector f 1111 1 float-4-rep }
+        T{ ##test-vector f 1 1111 2 float-4-rep vcc-any }
+        T{ ##compare-imm-branch f 1 5 cc/= }
+    } value-numbering-step trim-temps
+] unit-test
+
 ! Immediate operand conversion
 [
     {

diff --git a/basis/compiler/codegen/codegen.factor b/basis/compiler/codegen/codegen.factor
index 7c4c593d1624930f6de96af8cbe1e3ac410dd2ef..839e1aef05a40b0940b175d712a8c63a95153b14 100755 (executable)
--- a/basis/compiler/codegen/codegen.factor
+++ b/basis/compiler/codegen/codegen.factor
@@ -235,6 +235,7 @@ CODEGEN: _compare-branch %compare-branch
 CODEGEN: _compare-imm-branch %compare-imm-branch
 CODEGEN: _compare-float-ordered-branch %compare-float-ordered-branch
 CODEGEN: _compare-float-unordered-branch %compare-float-unordered-branch
+CODEGEN: _test-vector-branch %test-vector-branch
 CODEGEN: _dispatch %dispatch
 CODEGEN: _spill %spill
 CODEGEN: _reload %reload

diff --git a/basis/cpu/architecture/architecture.factor b/basis/cpu/architecture/architecture.factor
index ffe77671a747ad5e934f19e78512c54c6fe5e212..41b3fed08d819b7fe9d11cad0de3380007bf96ca 100644 (file)
--- a/basis/cpu/architecture/architecture.factor
+++ b/basis/cpu/architecture/architecture.factor
@@ -224,7 +224,8 @@ HOOK: %gather-vector-2 cpu ( dst src1 src2 rep -- )
 HOOK: %gather-vector-4 cpu ( dst src1 src2 src3 src4 rep -- )
 HOOK: %shuffle-vector cpu ( dst src shuffle rep -- )
 HOOK: %compare-vector cpu ( dst src1 src2 rep cc -- )
-HOOK: %test-vector cpu ( dst src temp rep vcc -- )
+HOOK: %test-vector cpu ( dst src1 temp rep vcc -- )
+HOOK: %test-vector-branch cpu ( label src1 temp rep vcc -- )
 HOOK: %add-vector cpu ( dst src1 src2 rep -- )
 HOOK: %saturated-add-vector cpu ( dst src1 src2 rep -- )
 HOOK: %add-sub-vector cpu ( dst src1 src2 rep -- )

diff --git a/basis/cpu/x86/x86.factor b/basis/cpu/x86/x86.factor
index 877afaa390d4a86e028abbb1338b16482b41ce9c..cca8e617ca297cd3fa7918243bc002a5942e3c5b 100644 (file)
--- a/basis/cpu/x86/x86.factor
+++ b/basis/cpu/x86/x86.factor
@@ -747,20 +747,36 @@ M: x86 %compare-vector-reps
         { sse4.1? { longlong-2-rep ulonglong-2-rep } }
     } available-reps ;
 
-:: (%test-vector) ( dst temp mask vcc -- )
+:: %test-vector-mask ( dst temp mask vcc -- )
     vcc {
-        { vcc-any  [ dst dst TEST dst temp \ CMOVNE %boolean ] }
-        { vcc-none [ dst dst TEST dst temp \ CMOVE  %boolean ] }
-        { vcc-all  [ dst mask CMP dst temp \ CMOVE  %boolean ] }
+        { vcc-any    [ dst dst TEST dst temp \ CMOVNE %boolean ] }
+        { vcc-none   [ dst dst TEST dst temp \ CMOVE  %boolean ] }
+        { vcc-all    [ dst mask CMP dst temp \ CMOVE  %boolean ] }
+        { vcc-notall [ dst mask CMP dst temp \ CMOVNE %boolean ] }
     } case ;
 
-M:: x86 %test-vector ( dst src temp rep vcc -- )
-    dst src rep {
+: %move-vector-mask ( dst src rep -- mask )
+    {
         { double-2-rep [ MOVMSKPD HEX: 3 ] }
         { float-4-rep  [ MOVMSKPS HEX: f ] }
         [ drop PMOVMSKB HEX: ffff ]
-    } case :> mask
-    dst temp mask vcc (%test-vector) ;
+    } case ;
+
+M:: x86 %test-vector ( dst src temp rep vcc -- )
+    dst src rep %move-vector-mask :> mask
+    dst temp mask vcc %test-vector-mask ;
+
+:: %test-vector-mask-branch ( label temp mask vcc -- )
+    vcc {
+        { vcc-any    [ temp temp TEST label JNE ] }
+        { vcc-none   [ temp temp TEST label JE ] }
+        { vcc-all    [ temp mask CMP label JE ] }
+        { vcc-notall [ temp mask CMP label JNE ] }
+    } case ;
+
+M:: x86 %test-vector-branch ( label src temp rep vcc -- )
+    temp src rep %move-vector-mask :> mask
+    label temp mask vcc %test-vector-mask-branch ;
 
 M: x86 %test-vector-reps
     {