Another failure in CASE form

[Paul F. Dietz]

(lambda (b)
  (case b
    ((4 1 2 3) 0)
    ((2 4) (mod b (min -49 0)))
    (t 0)))

==>

The value
  NIL
is not of type
  SB-C:TN-REF
   [Condition of type TYPE-ERROR]

Restarts:
 0: [ABORT] Exit debugger, returning to top level.

Backtrace:
  0: (SB-C::DELETE-TN-REF #<SB-C:TN-REF :TN #<SB-C:TN t1 :UNUSED> :WRITE-P T :VOP SB-C:CALL-NAMED>)
  1: (SB-C::DELETE-VOP #<SB-C::VOP :INFO SB-C:CALL-NAMED :ARGS #<SB-C:TN-REF :TN #<SB-C:TN t2[RSP] :NORMAL> :WRITE-P NIL :VOP SB-C:CALL-NAMED> :RESULTS #<SB-C:TN-REF :TN #<SB-C:TN t1 :UNUSED> :WRITE-P T :V..
  2: (SB-C::DELETE-UNUSED-IR2-BLOCKS #<SB-C:COMPONENT :NAME (LAMBDA (B)) {1009DFB9F3}>)
  3: (SB-C::IR2-OPTIMIZE #<SB-C:COMPONENT :NAME (LAMBDA (B)) {1009DFB9F3}>)
[…]

[Paul F. Dietz (paul-f-dietz)]

Similar failure on

(lambda (s1 s2 c)
  (case s2
    ((2) s2)
    ((2) (mod s1 (max 9 0)))
    ((1 4) 0)
    ((-1) c)
    (t 0)))

[Douglas Katzman (dougk)]

this is interesting because it essentially is performing dead-code elimination that was not otherwise performed. The disassembly prior the new feature of multiway branch contains a ton of garbage.

... serial comparisons here, fine ... but then TRUNCATE and GENERIC= ?

; 29C: L2: 48837DF002 CMP QWORD PTR [RBP-16], 2
; 2A1: 74F1 JEQ L0
; 2A3: 48837DF004 CMP QWORD PTR [RBP-16], 4
; 2A8: 74EA JEQ L0
; 2AA: 48837DF006 CMP QWORD PTR [RBP-16], 6
; 2AF: 74E3 JEQ L0
; 2B1: 48837DF004 CMP QWORD PTR [RBP-16], 4
; 2B6: 0F85F2000000 JNE L13
; 2BC: L3: 488B45F0 MOV RAX, [RBP-16]
; 2C0: 8D70F1 LEA ESI, [RAX-15]
; 2C3: 40F6C601 TEST SIL, 1
; 2C7: 751A JNE L4
; 2C9: 4080FE0A CMP SIL, 10
; 2CD: 7414 JEQ L4
; 2CF: 40F6C60F TEST SIL, 15
; 2D3: 0F85CE000000 JNE L12
; 2D9: 8078F11D CMP BYTE PTR [RAX-15], 29
; 2DD: 0F87C4000000 JNBE L12
; 2E3: L4: 488B55F0 MOV RDX, [RBP-16]
; 2E7: 4883EC10 SUB RSP, 16
; 2EB: 48C7C79EFFFFFF MOV RDI, -98
; 2F2: B904000000 MOV ECX, 4
; 2F7: 48892C24 MOV [RSP], RBP
; 2FB: 488BEC MOV RBP, RSP
; 2FE: E83D535CFF CALL #x52116640 ; #<FUNCTION TRUNCATE>
; 303: 48897DE8 MOV [RBP-24], RDI
; 307: 488B55E8 MOV RDX, [RBP-24]
; 30B: 31FF XOR EDI, EDI
; 30D: E8EEFA5AFF CALL #x52100E00 ; GENERIC-=
; etc.

As evidenced from the few uses of DELETE-VOP, it's unclear how robust this stuff is.
Are they broken, or am I "holding it wrong"? I'll dig some more into it.
Anyway, it certainly suggests that analysis of the if/else chain would be nice to do even if the backend does not support coding as anything but a bunch of serial tests, because the analysis is a precondition for finding the unreachable branches.