misc improvements

Code/42nd-at-threadmill.asd

@@ -1,18 +1,13 @@
 (asdf:defsystem :42nd-at-threadmill
-  :depends-on (:atomics :bordeaux-threads)
+  :depends-on (:atomics :bordeaux-threads :sb-simd)
   :serial t
   :components ((:file "package")
                (:module "x86-64"
                 :components
                 ((:module "VOPs"
                   :components
                   ((:file "define-boring-vop")
-                   (:file "avx2-broadcastb"
-                    :if-feature (:not :threadmill-avx2))
-                   (:file "sse2-vops"
-                    :if-feature (:not :threadmill-avx2))
-                   (:file "avx2-vops"
-                    :if-feature :threadmill-avx2)
+                   (:file "sse2-vops" :if-feature (:not :threadmill-avx2))
                    (:file "bsf")
                    (:file "cas-bytes")))
                  (:file "sse-metadata"

Code/hash-table.lisp

@@ -1,7 +1,5 @@
 (in-package :threadmill)
 
-(defconstant +empty+ '+empty+)
-(defconstant +copied+ '+copied+)
 ;; SBCL seems to have difficulties proving that we won't go out of
 ;; bounds. Fair enough, we have a funny way of addressing the table,
 ;; but we don't want to do bounds checks.

Code/package.lisp

@@ -7,4 +7,6 @@
            #:gethash #:remhash #:clrhash #:maphash)
   (:export #:make-hash-table #:hash-table #:hash-table-p
            #:hash-table-test #:hash-table-count #:hash-table-size
-           #:gethash #:remhash #:clrhash #:maphash #:modhash))
+           #:gethash #:remhash #:clrhash #:maphash #:modhash)
+  (:local-nicknames (#:avx2 #:sb-simd-avx2)
+                    (#:sse2 #:sb-simd-sse2)))

Code/storage-vector.lisp

@@ -8,6 +8,9 @@
 ;;; 4. How many elements we are going to copy (used when resizing)
 ;;; Then the rest of the table just contains a key, then a value, and so on.
 
+(defconstant +empty+ '+empty+)
+(defconstant +copied+ '+copied+)
+
 ;;; We define macros so that we can SETF and CAS the position, without
 ;;; having to write setters and (unportable) CAS-ers.
 (defconstant +words-before-values+ 8)

Code/x86-64/VOPs/sse2-vops.lisp

@@ -1,55 +1,21 @@
 (in-package :sb-vm)
 
-(defknown threadmill::%sse2-load
-    ((simple-array (unsigned-byte 8) (*)) (unsigned-byte 64))
-    (simd-pack integer)
-    (foldable flushable)
-  :overwrite-fndb-silently t)
-
-(threadmill::define-boring-vop threadmill::%sse2-load
-    ((vector simple-array-unsigned-byte-8 :scs (descriptor-reg))
-     (index  unsigned-num :scs (unsigned-reg)))
-    (bytes simd-pack-int :scs (int-sse-reg))
-  ;; SBCL vectors are aligned to 16 bytes.
-  (inst movdqa bytes
-        (ea (- (* vector-data-offset n-word-bytes)
-               other-pointer-lowtag)
-            vector index 1)))
+; require this currently because sb-simd broadcast-byte really sucks
 (defknown threadmill::%sse2-broadcast-byte
     ((unsigned-byte 8))
-    (simd-pack integer)
-    (foldable movable flushable)
-  :overwrite-fndb-silently t)
-
-(threadmill::define-boring-vop threadmill::%sse2-broadcast-byte
-    ((byte unsigned-num :scs (unsigned-reg)))
-    (bytes simd-pack-int :scs (int-sse-reg))
-  (inst movd bytes byte)                ; xxxxxxxxxxxxxxxB
-  (inst punpcklbw bytes bytes)          ; xxxxxxxxxxxxxxBB
-  (inst punpcklbw bytes bytes)          ; xxxxxxxxxxxxBBBB
-  (inst pshufd bytes bytes #4r0000))    ; BBBBBBBBBBBBBBBB
-
-(defknown threadmill::%sse2-movemask
-    ((simd-pack integer))
-    (unsigned-byte 16)
-    (foldable movable flushable)
-  :overwrite-fndb-silently t)
-
-(threadmill::define-boring-vop threadmill::%sse2-movemask
-    ((bytes simd-pack-int :scs (int-sse-reg)))
-    (mask unsigned-num :scs (unsigned-reg))
-  (inst pmovmskb mask bytes))
-
-(defknown threadmill::%sse2=
-    ((simd-pack integer) (simd-pack integer))
-    (simd-pack integer)
+    (simd-pack (unsigned-byte 8))
     (foldable movable flushable)
   :overwrite-fndb-silently t)
 
-(threadmill::define-boring-vop threadmill::%sse2=
-    ((a simd-pack-int :scs (int-sse-reg) :target result)
-     (b simd-pack-int :scs (int-sse-reg)))
-    (result simd-pack-int :scs (int-sse-reg))
-  (unless (location= a result)
-    (inst movdqa result a))
-  (inst pcmpeqb result b))
+(sb-c:define-vop (threadmill::%sse2-broadcast-byte)
+  (:translate threadmill::%sse2-broadcast-byte)
+  (:policy :fast-safe)
+  (:args (byte :scs (unsigned-reg)))
+  (:arg-types unsigned-num)
+  (:results (bytes :scs (int-sse-reg)))
+  (:result-types simd-pack-ub8)
+  (:temporary (:scs (int-sse-reg)) zero)
+  (:generator 0
+              (inst movd bytes byte)
+              (inst pxor zero zero)
+              (inst pshufb bytes zero)))

Code/x86-64/avx2-metadata.lisp

@@ -25,14 +25,14 @@
 (defun bytes (byte group)
   "Return matches for a byte in a metadata group."
   (declare ((unsigned-byte 8) byte))
-  (%avx2-movemask
-   (%avx2= (%avx2-broadcast/256 byte) group)))
+  (avx2:u8.32-movemask
+   (avx2:u8.32= group (avx2:u8.32 byte))))
 
 (defun writable (group)
   "Return matches for metadata bytes we can put new mappings in."
   ;; movemask tests the high bit of each byte, and we want to test the
   ;; high bit, so we have nothing else to do. Magic!
-  (%avx2-movemask group))
+  (avx2:u8.32-movemask group))
 
 (defun match-union (m1 m2)
   (logior m1 m2))
@@ -77,7 +77,7 @@ Note that N has a length of an element."
            (vector-index position)
            (optimize (speed 3) (safety 0)))
   ;; Why won't SSE:AREF-PI work?
-  (%avx2-load vector position))
+  (avx2:u8.32-aref vector position))
 
 (defun metadata-groups (metadata)
   (floor (length metadata) +metadata-entries-per-group+))

Code/x86-64/sse-metadata.lisp

@@ -22,23 +22,17 @@
   (declare ((unsigned-byte 8) h2))
   (logand #x7f h2))
 
-(if (= 1 avx2-supported)
-    (defun bytes (byte group)
-      "Return matches for a byte in a metadata group."
-      (declare ((unsigned-byte 8) byte))
-      (%sse2-movemask
-       (%sse2= (%avx2-broadcast byte) group)))
-    (defun bytes (byte group)
-      "Return matches for a byte in a metadata group."
-      (declare ((unsigned-byte 8) byte))
-      (%sse2-movemask
-       (%sse2= (%sse2-broadcast-byte byte) group))))
+(defun bytes (byte group)
+  "Return matches for a byte in a metadata group."
+  (declare ((unsigned-byte 8) byte))
+  (sse2:u8.16-movemask
+   (sse2:u8.16= group (%sse2-broadcast-byte byte))))
 
 (defun writable (group)
   "Return matches for metadata bytes we can put new mappings in."
   ;; movemask tests the high bit of each byte, and we want to test the
   ;; high bit, so we have nothing else to do. Magic!
-  (%sse2-movemask group))
+  (sse2:u8.16-movemask group))
 
 (defun match-union (m1 m2)
   (logior m1 m2))
@@ -82,7 +76,7 @@ Note that N has a length of an element."
   (declare (metadata-vector vector)
            (vector-index position)
            (optimize (speed 3) (safety 0)))
-  (%sse2-load vector position))
+  (sse2:u8.16-aref vector position))
 
 (defun metadata-groups (metadata)
   (floor (length metadata) +metadata-entries-per-group+))