Micro-optimization: Avoid byte register writes on x86-64 in LOAD-TYPE.

The optimization guide for AMD's x86-64 processors recommends not
to write a partial register but instead to use MOVZX to write the
corresponding 32/64-bit register. Otherwise the instruction would have
an unnecessary dependency on the most recent write to the register,
reducing the available instruction level parallelism. On Intel's
processors this is not necessary but doesn't hurt.

To follow this recommendation, modify LOAD-TYPE to use MOVZX instead of
a byte MOV and adapt the VOPs that call it: FUN-SUBTYPE, SET-FDEFN-FUN,
and WIDETAG-OF. This additionally spares a temporary register in
FUN-SUBTYPE and allows to shorten all paths through WIDETAG-OF by one
instruction.

The effect on code size is small and mixed.

diff --git a/src/compiler/x86-64/cell.lisp b/src/compiler/x86-64/cell.lisp
index 8744c26..0263435 100644 (file)
--- a/src/compiler/x86-64/cell.lisp
+++ b/src/compiler/x86-64/cell.lisp
@@ -263,7 +263,7 @@
   (:args (function :scs (descriptor-reg) :target result)
          (fdefn :scs (descriptor-reg)))
   (:temporary (:sc unsigned-reg) raw)
-  (:temporary (:sc byte-reg) type)
+  (:temporary (:sc unsigned-reg) type)
   (:results (result :scs (descriptor-reg)))
   (:generator 38
     (load-type type function (- fun-pointer-lowtag))
@@ -271,7 +271,7 @@
           (make-ea :byte :base function
                    :disp (- (* simple-fun-code-offset n-word-bytes)
                             fun-pointer-lowtag)))
-    (inst cmp type simple-fun-header-widetag)
+    (inst cmp (reg-in-size type :byte) simple-fun-header-widetag)
     (inst jmp :e NORMAL-FUN)
     (inst lea raw (make-fixup "closure_tramp" :foreign))
     NORMAL-FUN

diff --git a/src/compiler/x86-64/macros.lisp b/src/compiler/x86-64/macros.lisp
index 801c905..13deab4 100644 (file)
--- a/src/compiler/x86-64/macros.lisp
+++ b/src/compiler/x86-64/macros.lisp
@@ -120,10 +120,10 @@
               (n-offset offset))
     (ecase *backend-byte-order*
       (:little-endian
-       `(inst mov ,n-target
+       `(inst movzx ,n-target
               (make-ea :byte :base ,n-source :disp ,n-offset)))
       (:big-endian
-       `(inst mov ,n-target
+       `(inst movzx ,n-target
               (make-ea :byte :base ,n-source
                              :disp (+ ,n-offset (1- n-word-bytes))))))))
 \f

diff --git a/src/compiler/x86-64/system.lisp b/src/compiler/x86-64/system.lisp
index 7cb54a4..ae57c4c 100644 (file)
--- a/src/compiler/x86-64/system.lisp
+++ b/src/compiler/x86-64/system.lisp
@@ -28,11 +28,12 @@
   (:translate widetag-of)
   (:policy :fast-safe)
   (:args (object :scs (descriptor-reg)))
-  (:temporary (:sc unsigned-reg :offset eax-offset :to (:result 0)) rax)
+  (:temporary (:sc unsigned-reg :offset rax-offset :target result
+                   :to (:result 0)) rax)
   (:results (result :scs (unsigned-reg)))
   (:result-types positive-fixnum)
   (:generator 6
-    (inst mov rax object)
+    (inst movzx rax (reg-in-size object :byte))
     (inst and al-tn lowtag-mask)
     (inst cmp al-tn other-pointer-lowtag)
     (inst jmp :e OTHER-PTR)
@@ -48,29 +49,27 @@
     (inst jmp :ne DONE)
 
     ;; must be an other immediate
-    (inst mov rax object)
+    (inst movzx rax (reg-in-size object :byte))
     (inst jmp DONE)
 
     FUNCTION-PTR
-    (load-type al-tn object (- fun-pointer-lowtag))
+    (load-type rax object (- fun-pointer-lowtag))
     (inst jmp DONE)
 
     OTHER-PTR
-    (load-type al-tn object (- other-pointer-lowtag))
+    (load-type rax object (- other-pointer-lowtag))
 
     DONE
-    (inst movzx result al-tn)))
+    (move result rax)))
 \f
 (define-vop (fun-subtype)
   (:translate fun-subtype)
   (:policy :fast-safe)
   (:args (function :scs (descriptor-reg)))
-  (:temporary (:sc byte-reg :from (:eval 0) :to (:eval 1)) temp)
   (:results (result :scs (unsigned-reg)))
   (:result-types positive-fixnum)
   (:generator 6
-    (load-type temp function (- fun-pointer-lowtag))
-    (inst movzx result temp)))
+    (load-type result function (- fun-pointer-lowtag))))
 
 (define-vop (set-fun-subtype)
   (:translate (setf fun-subtype))