Incorrect implementation of pattern matching for exceptions #223
Description
Version
110.99.1
Operating System
- All
- Linux
- macOS
- Windows
- Other Unix
OS Version
No response
Processor
Any
Component
Core system
Severity
Minor
Description of the problem
The example comes from the coresml regression tests (test d001a-ac.sml).
With FLINT printing on, we see the source of the problem is that the order of cases has changed (B appears before A): For exception constructors, which can be aliases, this change results in incorrect behavior.
[After Translate ...]
FN(v85:
STR([TYC(TCAP(PRIM(ETG), [{}])),TYC(TCAP(PRIM(ETG), [{}])),
TYC(TCAP(PRIM(ETG), [{}]))]),
v73 = v85[0]
v72 = v85[1]
v74 = v85[2]
v70 =
v71 =
CON((B,EXNLVAR(72),TYC(AR[rr]([{}], [PRIM(EXN)]))), [],
RCD())
v75 = FN(v76: TYC({}),(I63)1)
v77 = FN(v78: TYC({}),(I63)2)
v79 = FN(v80: TYC({}),(I63)3)
v81 =
FN(v82:
TYC({}),
RAISE(TYC(PRIM(I63)),
APP(PRM(markexn,
TYC(AR[rr]([PRIM(EXN),PRIM(STR)],
[PRIM(EXN)])), []),
RCD(CON((Match,EXNLVAR(74),TYC(AR[rr]([{}],
[PRIM(EXN)]))), [], RCD()),
"stdIn:4.36"))))
SWIv71
of B.v83 => APP(v77, RCD())
A.v84 => APP(v75, RCD())
_ => APP(v79, RCD())
SRCD(v70))
Transcript
<jhr@Froh> sml
Standard ML of New Jersey (64-bit) v110.99.1 [built: Mon Apr 12 18:45:14 2021]
- exception A;
exception A
- exception B = A;
exception B = A
- (case B of A => 1 | B => 2 | _ => 3);
val it = 2 : int
Expected Behavior
No response
Steps to Reproduce
exception A;
exception B = A;
(case B of A => 1 | B => 2 | _ => 3);
Additional Information
No response
Email address
No response
Comments from smlnj-gforge
Original smlnj-gforge bug number 290
Submitted on 2021-57-10 at 20:5700
Keywords: pattern matching, exceptions
comment by @dmacqueen on 2021-06-10 03:0600 +000 UTC
A spurious "rev" operation in the SWITCHexp case of Translate.mkExp was removed, which restored the proper order of rules in the example (the A => rule ~ now comes first. However, this was not enough to fix the bug. Deeper analysis of the issues around identity exception declarations will be required.
comment by @dmacqueen on 2022-23-17 19:2300 +000 UTC
The mkExp rule transposition problem still applies the the legacy branch, but with the new match compiler, the bug persisted. The fix is to suppress the SWITCH contraction (FLINT/opt/lcontract.sml: swiInfo; FLINT/opt/fcontract.sml: fcSwitch) when the subject is an exception constructor, because these functions use equality of conreps to identify the exception constructor, which is not correct when one of the exceptions is defined by an exception identity declaration as in the bug test case. The bug remains open in the legacy branch.