Improved SIMD-PACK manipulation VOPs on x86-64

 * Tighten naive (and, in one case, wrong-looking) lifetime
   specifications to enable more register coalescing.

 * Exploit the coalescing by replacing explicit move instructions
   with MOVE, and specify the SC to get MOVAPS for float data.

 * Microoptimise away some PSRLDQ by 0.

diff --git a/src/compiler/x86-64/simd-pack.lisp b/src/compiler/x86-64/simd-pack.lisp
index 717265c..cbb48e2 100644 (file)
--- a/src/compiler/x86-64/simd-pack.lisp
+++ b/src/compiler/x86-64/simd-pack.lisp
@@ -159,14 +159,15 @@
 
 (define-vop (%simd-pack-high)
   (:translate %simd-pack-high)
-  (:args (x :scs (int-sse-reg double-sse-reg single-sse-reg)))
+  (:args (x :scs (int-sse-reg double-sse-reg single-sse-reg)
+            :target tmp))
   (:arg-types simd-pack)
-  (:temporary (:sc sse-reg) tmp)
+  (:temporary (:sc sse-reg :from (:argument 0)) tmp)
   (:results (dst :scs (unsigned-reg)))
   (:result-types unsigned-num)
   (:policy :fast-safe)
   (:generator 3
-    (inst movdqa tmp x)
+    (move tmp x)
     (inst psrldq tmp 8)
     (inst movd dst tmp)))
 
@@ -208,7 +209,7 @@
   (:args (lo :scs (unsigned-reg))
          (hi :scs (unsigned-reg)))
   (:arg-types unsigned-num unsigned-num)
-  (:temporary (:sc sse-reg) tmp)
+  (:temporary (:sc int-sse-reg) tmp)
   (:results (dst :scs (int-sse-reg)))
   (:result-types simd-pack-int)
   (:generator 5
@@ -245,15 +246,14 @@
     (values (%simd-pack-low pack)
             (%simd-pack-high pack))))
 
-
 (define-vop (%make-simd-pack-double)
   (:translate %make-simd-pack-double)
   (:policy :fast-safe)
-  (:args (lo :scs (double-reg))
-         (hi :scs (double-reg)))
+  (:args (lo :scs (double-reg) :target dst)
+         (hi :scs (double-reg) :target tmp))
   (:arg-types double-float double-float)
-  (:temporary (:sc double-sse-reg) tmp)
-  (:results (dst :scs (double-sse-reg)))
+  (:temporary (:sc double-sse-reg :from (:argument 1)) tmp)
+  (:results (dst :scs (double-sse-reg) :from (:argument 0)))
   (:result-types simd-pack-double)
   (:generator 5
     (move dst lo)
@@ -267,13 +267,13 @@
 (define-vop (%make-simd-pack-single)
   (:translate %make-simd-pack-single)
   (:policy :fast-safe)
-  (:args (x :scs (single-reg))
-         (y :scs (single-reg))
+  (:args (x :scs (single-reg) :target dst)
+         (y :scs (single-reg) :target tmp)
          (z :scs (single-reg))
          (w :scs (single-reg)))
   (:arg-types single-float single-float single-float single-float)
-  (:temporary (:sc sse-reg) tmp)
-  (:results (dst :scs (single-sse-reg)))
+  (:temporary (:sc single-sse-reg :from (:argument 1)) tmp)
+  (:results (dst :scs (single-sse-reg) :from (:argument 0)))
   (:result-types simd-pack-single)
   (:generator 5
     (move dst x)
@@ -290,18 +290,25 @@
   (%simd-pack-tag pack))
 
 (define-vop (%simd-pack-single-item)
-  (:args (x :scs (int-sse-reg double-sse-reg single-sse-reg)))
+  (:args (x :scs (int-sse-reg double-sse-reg single-sse-reg)
+            :target tmp))
   (:arg-types simd-pack)
   (:info index)
   (:results (dst :scs (single-reg)))
   (:result-types single-float)
-  (:temporary (:sc sse-reg) tmp)
+  (:temporary (:sc single-sse-reg :from (:argument 0)) tmp)
   (:policy :fast-safe)
   (:generator 3
-    (inst movdqa tmp x)
-    (inst psrldq tmp (* 4 index))
-    (inst xorps dst dst)
-    (inst movss dst tmp)))
+    (cond ((and (zerop index)
+                (not (location= x dst)))
+           (inst xorps dst dst)
+           (inst movss dst x))
+          (t
+           (move tmp x)
+           (when (plusp index)
+             (inst psrldq tmp (* 4 index)))
+           (inst xorps dst dst)
+           (inst movss dst tmp)))))
 
 #-sb-xc-host
 (declaim (inline %simd-pack-singles))
@@ -314,18 +321,25 @@
           (%primitive %simd-pack-single-item pack 3)))
 
 (define-vop (%simd-pack-double-item)
-  (:args (x :scs (int-sse-reg double-sse-reg single-sse-reg)))
+  (:args (x :scs (int-sse-reg double-sse-reg single-sse-reg)
+            :target tmp))
   (:info index)
   (:arg-types simd-pack)
   (:results (dst :scs (double-reg)))
   (:result-types double-float)
-  (:temporary (:sc sse-reg) tmp)
+  (:temporary (:sc double-sse-reg :from (:argument 0)) tmp)
   (:policy :fast-safe)
   (:generator 3
-    (inst movdqa tmp x)
-    (inst psrldq tmp (* 8 index))
-    (inst xorpd dst dst)
-    (inst movsd dst tmp)))
+    (cond ((and (zerop index)
+                (not (location= x dst)))
+           (inst xorpd dst dst)
+           (inst movsd dst x))
+          (t
+           (move tmp x)
+           (when (plusp index)
+             (inst psrldq tmp (* 8 index)))
+           (inst xorpd dst dst)
+           (inst movsd dst tmp)))))
 
 #-sb-xc-host
 (declaim (inline %simd-pack-doubles))