compiler.cfg.instructions: ##shuffle-vector-halves-imm insn to map to SSE's two-input SHUFPS/SHUFPD

basis/compiler/cfg/instructions/instructions.factor

@@ -298,6 +298,11 @@ def: dst
 use: src shuffle
 literal: rep ;
 
+PURE-INSN: ##shuffle-vector-halves-imm
+def: dst
+use: src1 src2
+literal: shuffle rep ;
+
 PURE-INSN: ##shuffle-vector-imm
 def: dst
 use: src

basis/compiler/codegen/codegen.factor

@@ -182,6 +182,7 @@ CODEGEN: ##fill-vector %fill-vector
 CODEGEN: ##gather-vector-2 %gather-vector-2
 CODEGEN: ##gather-vector-4 %gather-vector-4
 CODEGEN: ##shuffle-vector-imm %shuffle-vector-imm
+CODEGEN: ##shuffle-vector-halves-imm %shuffle-vector-halves-imm
 CODEGEN: ##shuffle-vector %shuffle-vector
 CODEGEN: ##tail>head-vector %tail>head-vector
 CODEGEN: ##merge-vector-head %merge-vector-head

basis/cpu/architecture/architecture.factor

@@ -301,6 +301,7 @@ 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: %shuffle-vector-imm cpu ( dst src shuffle rep -- )
+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 -- )
@@ -356,6 +357,7 @@ HOOK: %gather-vector-4-reps cpu ( -- reps )
 HOOK: %alien-vector-reps cpu ( -- reps )
 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: %signed-pack-vector-reps cpu ( -- reps )
 HOOK: %unsigned-pack-vector-reps cpu ( -- reps )
@@ -404,6 +406,7 @@ M: object %gather-vector-4-reps { } ;
 M: object %alien-vector-reps { } ;
 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 %signed-pack-vector-reps { } ;
 M: object %unsigned-pack-vector-reps { } ;

basis/cpu/x86/x86.factor

@@ -798,6 +798,19 @@ M: x86 %shuffle-vector-imm-reps
         { sse2? { double-2-rep int-4-rep uint-4-rep longlong-2-rep ulonglong-2-rep } }
     } available-reps ;
 
+M:: x86 %shuffle-vector-halves-imm ( dst src1 src2 shuffle rep -- )
+    dst src1 src2 rep two-operand
+    shuffle rep {
+        { double-2-rep [ >float-4-shuffle SHUFPS ] }
+        { float-4-rep [ SHUFPS ] }
+    } case ;
+
+M: x86 %shuffle-vector-halves-imm-reps
+    {
+        { sse? { float-4-rep } }
+        { sse2? { double-2-rep } }
+    } available-reps ;
+
 M: x86 %shuffle-vector ( dst src shuffle rep -- )
     two-operand PSHUFB ;