Keyword argument type info appears to be lost

[Andrew Kravchuk]

Hello. It looks like the type info from &KEY FTYPE declaration is somehow lost in the process of type inference with optimization cranked up to maximum. Here are two examples to illustrate:

(proclaim '(optimize (speed 3) (safety 0) (debug 0) (compilation-speed 0)))
(declaim (type (simple-array single-float) a))
(defparameter a (make-array 1 :element-type 'single-float))

(declaim (ftype (function (single-float)) test1))
(defun test1 (x) (equalp x (aref a 0)))
(disassemble #'test1)
; disassembly for TEST1
; Size: 52 bytes. Origin: #x5366950F ; TEST1
; 0F: 488B05CAFFFFFF MOV RAX, [RIP-54] ; 'A
; 16: 8B48F5 MOV ECX, [RAX-11]
; 19: 4A8B0C29 MOV RCX, [RCX+R13]
; 1D: 4883F9FF CMP RCX, -1
; 21: 480F444801 CMOVEQ RCX, [RAX+1]
; 26: F30F104901 MOVSS XMM1, [RCX+1]
; 2B: 0F2FC1 COMISS XMM0, XMM1
; 2E: BA4F010050 MOV EDX, #x5000014F ; T
; 33: B817010050 MOV EAX, #x50000117 ; NIL
; 38: 480F4AD0 CMOVP RDX, RAX
; 3C: 480F45D0 CMOVNE RDX, RAX
; 40: C9 LEAVE
; 41: F8 CLC
; 42: C3 RET
NIL

;; we can see the call to EQUALP was nicely optimized out to single COMISS

(declaim (ftype (function (&key (:x single-float))) test2))
(defun test2 (&key x) (equalp x (aref a 0)))
(disassemble #'test2)
; disassembly for TEST2
; Size: 55 bytes. Origin: #x53669651 ; TEST2
; 51: 488B0540FFFFFF MOV RAX, [RIP-192] ; 'A
; 58: 8B48F5 MOV ECX, [RAX-11]
; 5B: 4A8B0C29 MOV RCX, [RCX+R13]
; 5F: 4883F9FF CMP RCX, -1
; 63: 480F444801 CMOVEQ RCX, [RAX+1]
; 68: F30F104101 MOVSS XMM0, [RCX+1]
; 6D: 660F7EC7 MOVD EDI, XMM0
; 71: 48C1E720 SHL RDI, 32
; 75: 4080CF19 OR DIL, 25
; 79: B904000000 MOV ECX, 4
; 7E: FF7508 PUSH QWORD PTR [RBP+8]
; 81: FF2425C8040050 JMP [#x500004C8] ; EQUALP
NIL

;; ...and here we see the full call to EQUALP although compiler still has the information that X is a SINGLE-FLOAT

;; additional info:

(lisp-implementation-version) ;; => "2.3.5"

*features*
(:SLYNK :OCICL :CLPM-CLIENT :QUICKLISP :ASDF3.3 :ASDF3.2 :ASDF3.1 :ASDF3 :ASDF2
 :ASDF :OS-UNIX :NON-BASE-CHARS-EXIST-P :ASDF-UNICODE :ARENA-ALLOCATOR :X86-64
 :GENCGC :64-BIT :ANSI-CL :COMMON-LISP :ELF :IEEE-FLOATING-POINT :LINUX
 :LITTLE-ENDIAN :PACKAGE-LOCAL-NICKNAMES :SB-CORE-COMPRESSION :SB-LDB
 :SB-PACKAGE-LOCKS :SB-THREAD :SB-UNICODE :SBCL :UNIX)

$ uname -a
Linux melchior 6.2.13-zen #1 ZEN SMP PREEMPT Sun Jun 11 00:28:58 CEST 2023 x86_64 AMD Ryzen 5 3600X 6-Core Processor AuthenticAMD GNU/Linux

[Andrew Kravchuk (awkravchuk)]

It just occured to me that keyword argument might be NIL, in which case that would require EQUALP, so no bug here. Sorry for the fuss.

[Stas Boukarev (stassats)]

I suppose (when x (equalp x float)) could be a valid optimization.

[Andrew Kravchuk (awkravchuk)]

Yep, that does the trick. Thank you Stas!

[Stas Boukarev (stassats)]

But SBCL could do that automatically.