compiler.cfg.intrinsics.simd: intrinsic support for double-2>float-4 conversion

basis/compiler/cfg/instructions/instructions.factor

@@ -345,6 +345,11 @@ def: dst
 use: src1 src2
 literal: rep ;
 
+PURE-INSN: ##float-pack-vector
+def: dst
+use: src
+literal: rep ;
+
 PURE-INSN: ##signed-pack-vector
 def: dst
 use: src1 src2

basis/compiler/cfg/intrinsics/simd/backend/backend.factor

@@ -28,6 +28,7 @@ M: ##shuffle-vector-imm insn-available? rep>> %shuffle-vector-imm-reps member? ;
 M: ##shuffle-vector-halves-imm insn-available? rep>> %shuffle-vector-halves-imm-reps member? ;
 M: ##merge-vector-head insn-available? rep>> %merge-vector-reps member? ;
 M: ##merge-vector-tail insn-available? rep>> %merge-vector-reps member? ;
+M: ##float-pack-vector insn-available? rep>> %float-pack-vector-reps member? ;
 M: ##signed-pack-vector insn-available? rep>> %signed-pack-vector-reps member? ;
 M: ##unsigned-pack-vector insn-available? rep>> %unsigned-pack-vector-reps member? ;
 M: ##unpack-vector-head insn-available? rep>> %unpack-vector-head-reps member? ;

basis/compiler/cfg/intrinsics/simd/simd.factor

@@ -570,7 +570,12 @@ PREDICATE: fixnum-vector-rep < int-vector-rep
 
 : emit-simd-vpack-signed ( node -- )
     {
-        [ ^^signed-pack-vector ]
+        { double-2-rep [| src1 src2 rep |
+            src1 double-2-rep ^^float-pack-vector :> dst-head
+            src2 double-2-rep ^^float-pack-vector :> dst-tail
+            dst-head dst-tail { 0 1 0 1 } float-4-rep ^^shuffle-vector-halves-imm
+        ] }
+        { int-vector-rep [ ^^signed-pack-vector ] }
     } emit-vv-vector-op ;
 
 : emit-simd-vpack-unsigned ( node -- )

basis/compiler/codegen/codegen.factor

@@ -191,6 +191,7 @@ CODEGEN: ##shuffle-vector %shuffle-vector
 CODEGEN: ##tail>head-vector %tail>head-vector
 CODEGEN: ##merge-vector-head %merge-vector-head
 CODEGEN: ##merge-vector-tail %merge-vector-tail
+CODEGEN: ##float-pack-vector %float-pack-vector
 CODEGEN: ##signed-pack-vector %signed-pack-vector
 CODEGEN: ##unsigned-pack-vector %unsigned-pack-vector
 CODEGEN: ##unpack-vector-head %unpack-vector-head

basis/cpu/architecture/architecture.factor

@@ -314,6 +314,7 @@ HOOK: %shuffle-vector-halves-imm cpu ( dst src1 src2 shuffle rep -- )
 HOOK: %tail>head-vector cpu ( dst src rep -- )
 HOOK: %merge-vector-head cpu ( dst src1 src2 rep -- )
 HOOK: %merge-vector-tail cpu ( dst src1 src2 rep -- )
+HOOK: %float-pack-vector cpu ( dst src rep -- )
 HOOK: %signed-pack-vector cpu ( dst src1 src2 rep -- )
 HOOK: %unsigned-pack-vector cpu ( dst src1 src2 rep -- )
 HOOK: %unpack-vector-head cpu ( dst src rep -- )
@@ -371,6 +372,7 @@ HOOK: %shuffle-vector-reps cpu ( -- reps )
 HOOK: %shuffle-vector-imm-reps cpu ( -- reps )
 HOOK: %shuffle-vector-halves-imm-reps cpu ( -- reps )
 HOOK: %merge-vector-reps cpu ( -- reps )
+HOOK: %float-pack-vector-reps cpu ( -- reps )
 HOOK: %signed-pack-vector-reps cpu ( -- reps )
 HOOK: %unsigned-pack-vector-reps cpu ( -- reps )
 HOOK: %unpack-vector-head-reps cpu ( -- reps )
@@ -423,6 +425,7 @@ M: object %shuffle-vector-reps { } ;
 M: object %shuffle-vector-imm-reps { } ;
 M: object %shuffle-vector-halves-imm-reps { } ;
 M: object %merge-vector-reps { } ;
+M: object %float-pack-vector-reps { } ;
 M: object %signed-pack-vector-reps { } ;
 M: object %unsigned-pack-vector-reps { } ;
 M: object %unpack-vector-head-reps { } ;

basis/cpu/x86/sse/sse.factor

@@ -298,6 +298,14 @@ M: x86 %merge-vector-reps
         { sse2? { double-2-rep char-16-rep uchar-16-rep short-8-rep ushort-8-rep int-4-rep uint-4-rep longlong-2-rep ulonglong-2-rep } }
     } available-reps ;
 
+M: x86 %float-pack-vector
+    drop CVTPD2PS ;
+
+M: x86 %float-pack-vector-reps
+    {
+        { sse2? { double-2-rep } }
+    } available-reps ;
+
 M: x86 %signed-pack-vector
     [ two-operand ] keep
     {