From e74b0b2a7b21996bfc9026fa0fbb82909d4403ee Mon Sep 17 00:00:00 2001
From: Joe Groff <arcata@gmail.com>
Date: Sat, 15 May 2010 23:48:22 -0700
Subject: [PATCH] compiler.cfg.instructions: ##gather-int-vector-2/4
 instructions that map to SSE4 PINSR/PEXTR

---
 .../cfg/instructions/instructions.factor      | 10 +++++++++
 basis/compiler/codegen/codegen.factor         |  2 ++
 basis/cpu/architecture/architecture.factor    |  6 +++++
 basis/cpu/x86/x86.factor                      | 22 +++++++++++++++++++
 4 files changed, 40 insertions(+)

diff --git a/basis/compiler/cfg/instructions/instructions.factor b/basis/compiler/cfg/instructions/instructions.factor
index 38ca3a93b0..b469866668 100644
--- a/basis/compiler/cfg/instructions/instructions.factor
+++ b/basis/compiler/cfg/instructions/instructions.factor
@@ -295,11 +295,21 @@ def: dst
 use: src1/scalar-rep src2/scalar-rep
 literal: rep ;
 
+PURE-INSN: ##gather-int-vector-2
+def: dst
+use: src1/int-rep src2/int-rep
+literal: rep ;
+
 PURE-INSN: ##gather-vector-4
 def: dst
 use: src1/scalar-rep src2/scalar-rep src3/scalar-rep src4/scalar-rep
 literal: rep ;
 
+PURE-INSN: ##gather-int-vector-4
+def: dst
+use: src1/int-rep src2/int-rep src3/int-rep src4/int-rep
+literal: rep ;
+
 PURE-INSN: ##shuffle-vector
 def: dst
 use: src shuffle
diff --git a/basis/compiler/codegen/codegen.factor b/basis/compiler/codegen/codegen.factor
index d5ab84bf45..b787220b56 100755
--- a/basis/compiler/codegen/codegen.factor
+++ b/basis/compiler/codegen/codegen.factor
@@ -182,6 +182,8 @@ CODEGEN: ##zero-vector %zero-vector
 CODEGEN: ##fill-vector %fill-vector
 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: ##shuffle-vector-imm %shuffle-vector-imm
 CODEGEN: ##shuffle-vector-halves-imm %shuffle-vector-halves-imm
 CODEGEN: ##shuffle-vector %shuffle-vector
diff --git a/basis/cpu/architecture/architecture.factor b/basis/cpu/architecture/architecture.factor
index d3638dd377..b0d2747ce0 100644
--- a/basis/cpu/architecture/architecture.factor
+++ b/basis/cpu/architecture/architecture.factor
@@ -299,7 +299,9 @@ HOOK: %float>integer cpu ( dst src -- )
 HOOK: %zero-vector cpu ( dst rep -- )
 HOOK: %fill-vector cpu ( dst rep -- )
 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: %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 -- )
@@ -354,7 +356,9 @@ HOOK: %scalar>vector cpu ( dst src rep -- )
 HOOK: %zero-vector-reps cpu ( -- reps )
 HOOK: %fill-vector-reps cpu ( -- reps )
 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: %alien-vector-reps cpu ( -- reps )
 HOOK: %shuffle-vector-reps cpu ( -- reps )
 HOOK: %shuffle-vector-imm-reps cpu ( -- reps )
@@ -403,7 +407,9 @@ HOOK: %horizontal-shr-vector-imm-reps cpu ( -- reps )
 M: object %zero-vector-reps { } ;
 M: object %fill-vector-reps { } ;
 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 %alien-vector-reps { } ;
 M: object %shuffle-vector-reps { } ;
 M: object %shuffle-vector-imm-reps { } ;
diff --git a/basis/cpu/x86/x86.factor b/basis/cpu/x86/x86.factor
index 78c80264ed..9790f6e7dd 100644
--- a/basis/cpu/x86/x86.factor
+++ b/basis/cpu/x86/x86.factor
@@ -744,6 +744,18 @@ M: x86 %gather-vector-4-reps
         { sse2? { float-4-rep int-4-rep uint-4-rep } }
     } available-reps ;
 
+M:: x86 %gather-int-vector-4 ( dst src1 src2 src3 src4 rep -- )
+    dst rep %zero-vector
+    dst src1 32-bit-version-of 0 PINSRD
+    dst src2 32-bit-version-of 1 PINSRD
+    dst src3 32-bit-version-of 2 PINSRD
+    dst src4 32-bit-version-of 3 PINSRD ;
+
+M: x86 %gather-int-vector-4-reps
+    {
+        { sse4.1? { int-4-rep uint-4-rep } }
+    } available-reps ;
+
 M:: x86 %gather-vector-2 ( dst src1 src2 rep -- )
     rep signed-rep {
         { double-2-rep [
@@ -761,6 +773,16 @@ M: x86 %gather-vector-2-reps
         { sse2? { double-2-rep longlong-2-rep ulonglong-2-rep } }
     } available-reps ;
 
+M:: x86.64 %gather-int-vector-2 ( dst src1 src2 rep -- )
+    dst rep %zero-vector
+    dst src1 0 PINSRQ
+    dst src2 1 PINSRQ ;
+
+M: x86.64 %gather-int-vector-2-reps
+    {
+        { sse4.1? { longlong-2-rep ulonglong-2-rep } }
+    } available-reps ;
+
 : sse1-float-4-shuffle ( dst shuffle -- )
     {
         { { 0 1 2 3 } [ drop ] }