Tcl Source Code

View Ticket
Login
2022-11-10
13:56 • Closed ticket [85ce4bf928]: binary format R Inf stores FLT_MAX plus 7 othe... artifact: 0d0318ebc8 user: jan.nijtmans
2022-11-05
10:08
Fix for [85ce4bf928]: binary format R Inf] stores FLT_MAX check-in: 48706a2425 user: jan.nijtmans tags: core-8-branch
2022-11-02
15:06 • Ticket [85ce4bf928] binary format R Inf stores FLT_MAX status still Open w... artifact: 93e86f2f9a user: jan.nijtmans
15:00
Proposed fix for [85ce4bf928]: Fix for problems with storing Inf with [binary format R] closed check-in: 20959e0da2 user: jan.nijtmans tags: bug-85ce4bf928
2016-11-10
13:01 • Ticket [85ce4bf928] binary format R Inf stores FLT_MAX status still Open w... artifact: c56e1999c0 user: aspect
2015-05-16
05:51 • Ticket [1670783fff] binary man page missing Inf/NaN ref status still Open wi... artifact: c3ee43e2b2 user: dkf
2015-05-15
22:17 • Ticket [85ce4bf928] binary format R Inf stores FLT_MAX status still Open w... artifact: 7bd484521b user: dkf
22:14
[85ce4bf928] Fix for problems with storing Inf with [binary format R]. check-in: 21866fab39 user: dkf tags: bug-85ce4bf928
2015-03-17
12:01 • Ticket [85ce4bf928] binary format R Inf stores FLT_MAX status still Open w... artifact: d9e89537bb user: dkf
2015-02-20
23:02 • Ticket [85ce4bf928]: 4 changes artifact: 6a072f6b9e user: anonymous
23:00 • Add attachment aspect-binary-float-inf.fbun to ticket [85ce4bf928cd9a40... artifact: d63c3c2716 user: anonymous
2015-02-04
18:18 • Ticket [85ce4bf928] binary format R Inf stores FLT_MAX status still Open w... artifact: 6e6905b4e2 user: dgp
18:14 • Ticket [85ce4bf928]: 4 changes artifact: b87cfdf45c user: dgp
18:10 • Ticket [85ce4bf928]: 4 changes artifact: 035c4feada user: dgp
18:03 • Ticket [85ce4bf928]: 4 changes artifact: d1f53302e6 user: dgp
15:43 • Ticket [85ce4bf928]: 4 changes artifact: b1718f14da user: dgp
15:21 • Ticket [85ce4bf928]: 4 changes artifact: 089b6e5d66 user: dgp
15:15 • Ticket [85ce4bf928]: 4 changes artifact: 8616f05c52 user: dgp
15:12 • Ticket [85ce4bf928]: 4 changes artifact: ee2071a2ce user: dgp
2015-02-02
08:21 • New ticket [85ce4bf928]. artifact: 26ff2520d3 user: anonymous

Ticket UUID: 85ce4bf928cd9a40064aa5b36e7033184dd5379a
Title: [binary format R Inf] stores FLT_MAX
Type: Bug Created on: 2015-02-02 08:21:54
Submitter: anonymous Assigned to: jan.nijtmans
Subsystem: 48. Number Handling Severity: Minor
Priority: 5 Medium Last modified: 2022-11-10 13:56:50
Status: Closed Closed by: jan.nijtmans
Resolution: Fixed Closed on: 2022-11-10 13:56:50
Version: 8.6.3
Description:
% binary scan [binary format R Inf] R x
% set x
3.4028234663852886e+38

This is FLT_MAX - see the bit pattern described near the bottom of http://wiki.tcl.tk/41088

The issue seems to be the test at http://core.tcl.tk/tcl/artifact/fee105a45aaf25a3?ln=1915-1917

By my reading, the condition should become:

    if (!isinf(dvalue) && (fabs(dvalue) > (double)FLT_MAX))

In the chat, kbk adds:

If the float64 is less than FLT_MAX + 1/2 ulp, then it's not a 'true' overflow, and it should [binary format] to FLT_MAX. Greater than that, and it should be Inf.

Exact FLT_MAX + 1/2 ulp should round to FLT_MAX (round to even).

I have implemented a fix in http://chiselapp.com/user/aspect/repository/tcl/timeline?r=aspect-binary-float-inf

The first checkin http://chiselapp.com/user/aspect/repository/tcl/ci/f4d7334db9 corrects behaviour for Inf/-Inf only, and adds some tests.

Following checkins follow kbk's advice (I hope!) to correct out-of-range rounding behaviour, but with some questionable aspects:

  * c99 reliance for INFINITY
  * FLT_MAX+1/2ulp is calculated with an expression I don't entirely trust
  * changes results (!!) of some existing tests which expect rounding to FLT_MAX

The included tests should be complete, but those in the 3rd checkin don't conform to conventions of their neighbours.

Another way to create FLT_MAX+1/2ulp for comparison would be as follows, though I can't vouch for its portability.

    /* smallest double that should round to +Inf */
    double flt_min_inf {} {
        uint64_t bits = 0x47effffff0000001;
        return *(double*)&bits;
    }

    /* largest double that should round to FLT_MAX */
    double flt_max_dbl {} {
        uint64_t bits = 0x47effffff0000000;
        return *(double*)&bits;
    }

Let me know if there's anything I can do to make this patch easier to consume.
User Comments:
dgp added on 2015-02-04 15:12:30:
The patch changes the expected result of
existing tests in the test suite, so I'd
prefer to have a review from kbk.

dgp added on 2015-02-04 15:15:50:
The patch changes the expected result of
existing tests in the test suite, so I'd
prefer to have a review from kbk.

dgp added on 2015-02-04 15:21:35:
The patch changes the expected result of
existing tests in the test suite, so I'd
prefer to have a review from kbk.

dgp added on 2015-02-04 15:43:29:
The patch changes the expected result of
existing tests in the test suite, so I'd
prefer to have a review from kbk.

dgp added on 2015-02-04 18:03:16:
The patch changes the expected result of
existing tests in the test suite, so I'd
prefer to have a review from kbk.

dgp added on 2015-02-04 18:10:28:
The patch changes the expected result of
existing tests in the test suite, so I'd
prefer to have a review from kbk.

dgp added on 2015-02-04 18:14:59:
The patch changes the expected result of
existing tests in the test suite, so I'd
prefer to have a review from kbk.

dgp added on 2015-02-04 18:18:08:
The patch changes the expected result of
existing tests in the test suite, so I'd
prefer to have a review from kbk.

anonymous (claiming to be aspect) added on 2015-02-20 23:02:20:
fossil bundle attached to this ticket.

dkf added on 2015-05-15 22:17:16:

Imported as branch bug-85ce4bf928 (see [21866fab39])


aspect added on 2016-11-10 13:01:19:
As another data point, the same is not true of doubles:

% binary scan [binary format q Inf] q x
1
% set x
Inf

jan.nijtmans added on 2022-11-02 15:06:58:

Rebased to core-8-branch: [20959e0da2451a45]


jan.nijtmans added on 2022-11-10 13:56:50:

Fixed [48706a2425734695|here]. Since it depends on C99, the fix will not be in 8.6, but in 8.7


Attachments: