Compiler trouble with sb-simd-sse2:f64.2-horizonal functions

[ari pro]

(require "SB-SIMD")
(setf sb-ext:*evaluator-mode* :compile)
;; This works as expected:
(sb-simd-sse2:f64.2-horizontal* (sb-simd-sse2:make-f64.2 1.2d0 3.9d0)) ;=> 4.68d0
;; This does not:
(let ((x (sb-simd-sse2:make-f64.2 1.2d0 3.9d0)))
  (sb-simd-sse2:f64.2-horizontal* x))
;;=>0.0d0

(defmacro baz (body)
  `(let ((x (sb-simd-sse2:make-f64.2 1.2d0 3.9d0)))
     ,@(cdr body)
     (list
      (sb-simd-sse2:f64.2-horizontal* x)
      (sb-simd-sse2:f64.2-horizontal+ x)
      (sb-simd-sse2:f64.2-horizontal-and x)
      (sb-simd-sse2:f64.2-horizontal-max x)
      (sb-simd-sse2:f64.2-horizontal-min x)
      (sb-simd-sse2:f64.2-horizontal-or x)
      (sb-simd-sse2:f64.2-horizontal-xor x)
      )))

(baz '(declare (optimize debug))) ;=> (4.68d0 5.1d0 4.450147717014402d-309 3.9d0 1.2d0 #<DOUBLE-FLOAT quiet NaN> #<DOUBLE-FLOAT quiet NaN>)
(baz '(declare (optimize (debug 2)))) ;=> (4.68d0 5.1d0 4.450147717014402d-309 3.9d0 1.2d0 #<DOUBLE-FLOAT quiet NaN> #<DOUBLE-FLOAT quiet NaN>)
(baz '(declare (optimize (debug 1)))) ;=> (0.0d0 1.2d0 0.0d0 1.2d0 0.0d0 1.2d0 1.2d0)
(baz '(declare (optimize (speed 0) (safety 3) (compilation-speed 0) (space 0) (debug 1)))) ;=> (0.0d0 1.2d0 0.0d0 1.2d0 0.0d0 1.2d0 1.2d0)
(baz '(declare (special x))) ;=> (4.68d0 5.1d0 4.450147717014402d-309 3.9d0 1.2d0 #<DOUBLE-FLOAT quiet NaN> #<DOUBLE-FLOAT quiet NaN>)

(let ((x (sb-simd-sse2:make-f64.2 1.2d0 3.9d0)))
  (loop for i in (apropos-list "f64.2-horizontal" (find-package "SB-SIMD-SSE2") t)
 collect (funcall i x)))
;;=> (4.68d0 5.1d0 4.450147717014402d-309 3.9d0 1.2d0 #<DOUBLE-FLOAT quiet NaN> #<DOUBLE-FLOAT quiet NaN>)

;; sb-eval and sb-fasteval both function correctly
(setf sb-ext:*evaluator-mode* :interpret)
(baz '(declare (optimize debug))) ;=> (4.68d0 5.1d0 4.450147717014402d-309 3.9d0 1.2d0 #<DOUBLE-FLOAT quiet NaN> #<DOUBLE-FLOAT quiet NaN>)
(baz '(declare (optimize (debug 1)))) ;=> (4.68d0 5.1d0 4.450147717014402d-309 3.9d0 1.2d0 #<DOUBLE-FLOAT quiet NaN> #<DOUBLE-FLOAT quiet NaN>)
(baz '(declare (special x))) ;=> (4.68d0 5.1d0 4.450147717014402d-309 3.9d0 1.2d0 #<DOUBLE-FLOAT quiet NaN> #<DOUBLE-FLOAT quiet NaN>)

;; Other things that work as expected:
(setf sb-ext:*evaluator-mode* :compile)
(let ((x (sb-simd-avx2:make-f64.2 1.2d0 3.9d0)))
  (sb-simd-avx2:f64.2-horizontal* x))
;;=>4.68d0

(let ((x (sb-kernel:%make-simd-pack-double 1.2d0 3.9d0)))
  (sb-simd-sse2:f64.2+ x x))
;;=> #<SIMD-PACK 2.4000000000000d+0 7.8000000000000d+0>

(lisp-implementation-version) ;=>2.3.2.116-f4e98cbf2
The behavior described above remains unchanged when tested in a previous release and has probably been the same since sb-simd was introduced.

[ari pro (aripro)]

I think the issue might be with f64.2-shuffle, which is used by f64.2-values, which in turn is used to define the horizontal functions mentioned above.

(require "SB-SIMD")
(setf sb-ext:*evaluator-mode* :compile)

(sb-simd-sse2:f64.2-shuffle (sb-simd-sse2:make-f64.2 1.2d0 3.9d0) 1)
=> #<SIMD-PACK 3.9000000000000d+0 1.2000000000000d+0>

(let ((x (sb-simd-sse2:make-f64.2 1.2d0 3.9d0)))
  (sb-simd-sse2:f64.2-shuffle x 1))
=> #<SIMD-PACK 0.0000000000000d+0 1.2000000000000d+0>

(let ((x (sb-simd-sse2:make-f64.2 1.2d0 3.9d0)))
  (declare (optimize (debug 2)))
  (sb-simd-sse2:f64.2-shuffle x 1))
=> #<SIMD-PACK 3.9000000000000d+0 1.2000000000000d+0>

As above, f64.2-shuffle works fine with sb-eval and sb-fasteval.

You could "fix" f64.2-values and the horizontal functions with the following, but it would be better to figure out the cause of the f64.2-shuffle trouble.

From d4a01c7c1d10087e33c025023b727cd41cbd1441 Mon Sep 17 00:00:00 2001
From: me <email address hidden>
Date: Wed, 29 Mar 2023 01:32:58 +0000
Subject: [PATCH] Adjust sb-simd-sse2:f64.2-values

---
 contrib/sb-simd/code/define-fake-vops.lisp | 2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/contrib/sb-simd/code/define-fake-vops.lisp b/contrib/sb-simd/code/define-fake-vops.lisp
index 4432791c4..e0099c63e 100644
--- a/contrib/sb-simd/code/define-fake-vops.lisp
+++ b/contrib/sb-simd/code/define-fake-vops.lisp
@@ -420,7 +420,7 @@
 (define-fake-vop f64.2-values (x)
   (values
    (%f64!-from-p128 x)
- (%f64!-from-p128 (%f64.2-shuffle x 1))))
+ (%f64!-from-p128 (%f64.2-unpackhi x x))))

 (define-fake-vop f64.2-broadcast (x)
   (let ((v (%f64.2!-from-f64 x)))
--
2.39.1

[Marco Heisig (marco-heisig-h)]

You are right, the problem is sb-simd-sse2:f64.2-shuffle. It should take three arguments (the two SIMD packs to shuffle, and the bits indicating how to shuffle), but takes only two.

The attached patch fixes the issue. It also fixes a minor issue about the AVX shuffles.

[ari pro (aripro)]

Edit: this comment will be moved elsewhere, in accordance with Marco's request below.

[Marco Heisig (marco-heisig-h)]

This is a new bug that is not fixed by repairing f64.2-shuffle. Can you file it as a different bug, so that it gets its own number?