Performance regression related to inlining

[Michael Kappert]

The following code (without any inline/notinline declarations) takes 0.5s on SBCL 2.1.7 but 16.5s on 2.1.8 - 2.2.1 (current):

(defun rad (x)
  (declare (double-float x))
  (* (* 2d0 pi) (* x (/ 1 360d0))))

(time
 (dotimes (k 100000000)
   (cis (rad 90d0))))

Declaiming RAD inline achieves the performance of 2.1.7 without inline declarations.
Surprisingly, declaiming RAD notinline is even faster.

Repeatable test case:

Save to test-inlining.cl and execute on 2.1.7 an 2.1.8 with
sbcl --no-userinit --eval '(progn (compile-file "test-inlining.cl") (load "test-inlining"))'
Add inline declarations and repeat.

;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
;;; Run with
;;; /opt/sbcl-2.2.1/bin/sbcl --no-userinit --eval '(progn (compile-file "test-inlining.cl") (load "test-inlining"))'
;;; OS: Fedora 34, Fedora 35

(defpackage :regression
  (:use :common-lisp))

(in-package :regression)

;; -- Performance regression when not using any inline declarations
;; -- Inlining RAD eliminates the performance regression
;; (declaim (inline rad))
;; -- Declaim notinline is faster than decaim inline ?!
;; (declaim (notinline rad))
(defun rad (x)
  (declare (double-float x))
  (* (* 2d0 pi) (* x (/ 1 360d0))))

;; SBCL 2.1.4, 2.1.6, 2.1.7:
;; no inline/notinline 1.9s
;; SBCL 2.2.0, 2.2.1, 2.1.8:
;; no inline/notinline: 16.5s
;; declaim inline rad: 1.5s
;; declaim notinline rad: 0.5s
(time
 (dotimes (k 100000000)
   (cis (rad 90d0))))

;;; EOF
;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

CL-USER> *features*
(:SWANK :CL-FAD :BORDEAUX-THREADS :THREAD-SUPPORT CFFI-FEATURES:FLAT-NAMESPACE
 CFFI-FEATURES:X86-64 CFFI-FEATURES:UNIX :CFFI CFFI-SYS::FLAT-NAMESPACE
 :SPLIT-SEQUENCE :SBCL-USES-SB-ROTATE-BYTE :QUICKLISP :ASDF3.3 :ASDF3.2
 :ASDF3.1 :ASDF3 :ASDF2 :ASDF :OS-UNIX :NON-BASE-CHARS-EXIST-P :ASDF-UNICODE
 :X86-64 :GENCGC :64-BIT :ANSI-CL :COMMON-LISP :ELF :IEEE-FLOATING-POINT :LINUX
 :LITTLE-ENDIAN :PACKAGE-LOCAL-NICKNAMES :SB-LDB :SB-PACKAGE-LOCKS :SB-THREAD
 :SB-UNICODE :SBCL :UNIX)
--------------------------------------------------------------------------------
uname -a
Linux aguas-13 5.15.18-100.fc34.x86_64 #1 SMP Sat Jan 29 13:00:44 UTC 2022 x86_64 x86_64 x86_64 GNU/Linux
--------------------------------------------------------------------------------

[Stas Boukarev (stassats)]

Is the better performance of 2.1.7 because it throws away the body of dotimes?

[Michael Kappert (mak08)]

(defvar *acc* 0d0)
(defun rad (x)
  (declare (double-float x))
  (incf *acc*
        (* (* 2d0 pi) (* x (/ 1 360d0))))
(time
 (dotimes (k 100000000)
   (cis (rad 90d0))))
(format t "~a~%" *acc*)

==> Prints 1.57079632812136d8 in both 2.1.7 and 2.1.8 but it takes 6s in 2.1.7 vs. 24s in 2.1.8.
Am I still overlooking something?

[Stas Boukarev (stassats)]

You are not using the result of CIS.

[Stas Boukarev (stassats)]

This seems to be some weird AVX2 register interaction.