sb-simd-avx:f64.4-reverse is broken

[ari pro]

There are a handful of sb-simd functions defined in terms of undefined fake VOPs.

(require "SB-SIMD")
(loop for fn in (sb-simd-internals:filter-available-function-records #'sb-simd-internals:instruction-record-p)
      unless (fboundp (sb-simd-internals:instruction-record-vop fn))
 collect (sb-simd-internals:function-record-name fn))
=> (SB-SIMD::U64-FROM-F32
    SB-SIMD::U64-FROM-F64
    SB-SIMD::F32-RECIPROCAL
    SB-SIMD::F32-RSQRT
    SB-SIMD::F64-RECIPROCAL
    SB-SIMD::F64-RSQRT
    SB-SIMD-SSE2::F64.2-HSUM
    SB-SIMD-SSE2::F64.2-HPROD
    SB-SIMD-AVX:F64.4-REVERSE)

(sb-simd-avx:f64.4-reverse (sb-simd-avx:make-f64.4 1.0 2.0 3.0 4.0))
=>
debugger invoked on a UNDEFINED-FUNCTION @21A208E4 in thread
#<THREAD "main thread" RUNNING {1004858073}>:
  The function SB-SIMD-AVX::%F64.4-REVERSE is undefined.

Type HELP for debugger help, or (SB-EXT:EXIT) to exit from SBCL.

restarts (invokable by number or by possibly-abbreviated name):
  0: [CONTINUE ] Retry calling SB-SIMD-AVX::%F64.4-REVERSE.
  1: [USE-VALUE ] Call specified function.
  2: [RETURN-VALUE ] Return specified values.
  3: [RETURN-NOTHING] Return zero values.
  4: [ABORT ] Exit debugger, returning to top level.

("undefined function" #<SIMD-PACK-256 1.0000000000000d+0 2.0000000000000d+0 3.0000000000000d+0 4.0000000000000d+0>)

There is a test named check-package for which checks for exported symbols with no attached definition in contrib/sb-simd/test-suite/test-suite.lisp, but it does not catch this type of error, because sb-simd-avx:f64.4-reverse is fbound.

sb-simd-avx:f64.4-reverse was originally defined in https://github.com/marcoheisig/sb-simd/blob/e2f6bbaf19701e7db2aee9a535a73e36df889e75/code/define-pseudo-vops.lisp#L110-L111 as
(sb-simd::define-pseudo-vop f64.4-reverse (a)
  (f64.4-permute (f64.4-permute2f128 a a 1) 5))

It was commented out in https://github.com/marcoheisig/sb-simd/commit/1c5b07d383d798e1c580e5680f699501cf60bb88 and deleted in https://github.com/marcoheisig/sb-simd/commit/b8ba0d3d1fd3f9cb7d1202c1d3a34bc2863be425 .

This patch adds the fake VOP
(define-fake-vop f64.4-reverse (x)
  (%f64.4-permute (%f64.4-permute128 x x #4r01) #4r11))
to contrib/sb-simd/code/define-fake-vops.lisp and changes the mnemonic in contrib/sb-simd/code/instruction-sets/avx.lisp from #:vpermilpd to nil to reflect that the :encoding is :fake-vop.

I left the :cost at 2, because f64.4-permute and f64.4-permute128 both have a cost of 1.

Note that an equivalent operation is currently used in https://github.com/sbcl/sbcl/blob/master/contrib/sb-simd/code/define-fake-vops.lisp#L906-L907

As for the other symbols, which are internal:

In contrib/sb-simd/code/instruction-sets/sse2.lisp:
SB-SIMD-SSE2::F64.2-HSUM and SB-SIMD-SSE2::F64.2-HPROD were added in the commit https://github.com/marcoheisig/sb-simd/commit/d0238545af9536decec87edc622f0f343852541e .
Their corresponding fake VOPs were defined in a commented-out section of code/define-horizontals.lisp in that commit:
https://github.com/marcoheisig/sb-simd/blob/d0238545af9536decec87edc622f0f343852541e/code/define-horizontals.lisp .
The file code/define-horizontals.lisp was deleted in
https://github.com/marcoheisig/sb-simd/commit/ac744b48387a62b876e000a5818a36cc401e20aa

This patch deletes those two symbols, since I believe they were intended to have been deleted along with the other -hsum and -hprod functions that used to exist for other instruction sets.

In contrib/sb-simd/code/instruction-sets/sb-simd.lisp, this patch comments out SB-SIMD::U64-FROM-F32, SB-SIMD::U64-FROM-F64, SB-SIMD::F32-RECIPROCAL, SB-SIMD::F32-RSQRT, SB-SIMD::F64-RECIPROCAL, and SB-SIMD::F64-RSQRT.

SB-SIMD::U64-FROM-F32 and SB-SIMD::U64-FROM-F64 were added in
https://github.com/marcoheisig/sb-simd/commit/892ecdda2a157b8ef6e6723dec796b2017b823d6
but as far as I can tell, they never had their fake VOPs implemented.

The fake VOPs for SB-SIMD::F32-RECIPROCAL, SB-SIMD::F32-RSQRT, SB-SIMD::F64-RECIPROCAL, and SB-SIMD::F64-RSQRT are commented-out in contrib/sb-simd/code/define-fake-vops.lisp

I added a test similar to the loop above called check-instruction-record-vops. It signals an error when it finds instructions with unbound instruction-record-vops, which can happen when the instruction-record VOP is not defined or when the contrib is built on a machine that doesn't support a particular instruction set that the host machine does support (see https://sourceforge.net/p/sbcl/mailman/message/37795412/). The test may need a better error message or a better name.

[Marco Heisig (marco-heisig-h)]

You are right, you found several remnants from the past that should either be removed, or modernized.

I am working on a solution for all of these that also incorporates your patch. Stay tuned.

[Marco Heisig (marco-heisig-h)]

The attached patch fixes this bug, removes a few exports with no attached definition, and adds a few exports for functions that I simply forgot in the initial release.

It includes all changes from Ari's patch.