Unboxed 32-bit arithmetic on 64-bit processor

[Guillaume LE VAILLANT]

When calling a sequence of arithmetic functions working on a few 32-bit unsigned integers on a 64-bit processor, the integers are not unboxed.

(require 'sb-rotate-byte)

(declaim (inline mod32+ rol32)
         (ftype (function ((unsigned-byte 32) (unsigned-byte 32)) (unsigned-byte 32)) mod32+)
         (ftype (function ((unsigned-byte 32) (unsigned-byte 5)) (unsigned-byte 32)) rol32))

(defun mod32+ (a b)
  (declare (type (unsigned-byte 32) a b))
  (ldb (byte 32 0) (+ a b)))

(defun rol32 (a s)
  (declare (type (unsigned-byte 32) a) (type (integer 0 32) s))
  (sb-rotate-byte:rotate-byte s (byte 32 0) a))

(defun func32 (x y z)
  (declare (type (unsigned-byte 32) x y z)
           (optimize (speed 3) (space 0) (safety 0) (debug 0)))
  (mod32+ (rol32 (mod32+ x y) 5) z))

(disassemble 'func32)
; disassembly for FUNC32
; Size: 38 bytes. Origin: #x1003847822
; 22: 4801F9 ADD RCX, RDI ; no-arg-parsing entry point
; 25: 48230DA4FFFFFF AND RCX, [RIP-92] ; [#x10038477D0] = 1FFFFFFFE
; 2C: 48D1F9 SAR RCX, 1
; 2F: C1C105 ROL ECX, 5
; 32: 48D1E1 SHL RCX, 1
; 35: 4801F1 ADD RCX, RSI
; 38: 48230D91FFFFFF AND RCX, [RIP-111] ; [#x10038477D0] = 1FFFFFFFE
; 3F: 488BD1 MOV RDX, RCX
; 42: 488BE5 MOV RSP, RBP
; 45: F8 CLC
; 46: 5D POP RBP
; 47: C3 RET

Unboxing the integers would allow getting rid of "AND RCX #x1FFFFFFFE", "SAR RCX 1" and "SHL RCX 1". In functions where a lot of arithmetic instructions are applied to a few integers (like hash functions or ciphers), it could divide the number of assembly instructions by 2 or 3.

$ uname -a
Linux 4.7.3-gentoo #1 SMP Fri Sep 9 11:21:14 CEST 2016 x86_64 Intel(R) Core(TM) i7-3630QM CPU @ 2.40GHz GenuineIntel GNU/Linux

$ sbcl --version
SBCL 1.3.9

*features*
(:SWANK :QUICKLISP :SB-BSD-SOCKETS-ADDRINFO :ASDF-PACKAGE-SYSTEM :ASDF3.1
 :ASDF3 :ASDF2 :ASDF :OS-UNIX :NON-BASE-CHARS-EXIST-P :ASDF-UNICODE :64-BIT
 :64-BIT-REGISTERS :ALIEN-CALLBACKS :ANSI-CL :ASH-RIGHT-VOPS
 :C-STACK-IS-CONTROL-STACK :COMMON-LISP :COMPARE-AND-SWAP-VOPS
 :COMPLEX-FLOAT-VOPS :CYCLE-COUNTER :ELF :FLOAT-EQL-VOPS
 :FP-AND-PC-STANDARD-SAVE :GENCGC :IEEE-FLOATING-POINT :INLINE-CONSTANTS
 :INTEGER-EQL-VOP :LARGEFILE :LINKAGE-TABLE :LINUX :LITTLE-ENDIAN
 :MEMORY-BARRIER-VOPS :MULTIPLY-HIGH-VOPS :OS-PROVIDES-BLKSIZE-T
 :OS-PROVIDES-DLADDR :OS-PROVIDES-DLOPEN :OS-PROVIDES-GETPROTOBY-R
 :OS-PROVIDES-POLL :OS-PROVIDES-PUTWC :OS-PROVIDES-SUSECONDS-T
 :PACKAGE-LOCAL-NICKNAMES :PRECISE-ARG-COUNT-ERROR :RAW-INSTANCE-INIT-VOPS
 :SB-CORE-COMPRESSION :SB-DOC :SB-EVAL :SB-FUTEX :SB-LDB :SB-PACKAGE-LOCKS
 :SB-SIMD-PACK :SB-SOURCE-LOCATIONS :SB-THREAD :SB-UNICODE :SBCL
 :STACK-ALLOCATABLE-CLOSURES :STACK-ALLOCATABLE-FIXED-OBJECTS
 :STACK-ALLOCATABLE-LISTS :STACK-ALLOCATABLE-VECTORS
 :STACK-GROWS-DOWNWARD-NOT-UPWARD :SYMBOL-INFO-VOPS :UNBIND-N-VOP :UNIX
 :UNWIND-TO-FRAME-AND-CALL-VOP :X86-64)