| Ticket UUID: | 2dff98e0791720bdde656d47762553c1089684a4 | ||
| Title: | format %A garbles negative, padded and infinite values | ||
| Type: | Bug | Created on: | 2026-10-16 22:13:57 |
| Submitter: | serhiy.storchaka | Assigned to: | jan.nijtmans |
| Subsystem: | 16. Commands A-H | Severity: | Minor |
| Priority: | 5 Medium | Last modified: | 2026-10-18 21:07:22 |
| Status: | Closed | Closed by: | jan.nijtmans |
| Resolution: | Fixed | Closed on: | 2026-10-18 21:07:22 |
| Version: | 9.0.5, 9.1.0 | ||
| Description: | ||||
|
[format %A] overwrites the second character of the result with "x", assuming that it starts with "0X": format %A -1.5 ;# -xX1.8p+0 format %+A 1.0 ;# +xX1p+0 format %10A 1.5 ;# x0X1.8p+0 format %A Inf ;# IxF Only plain positive values come out right (0x1.8p+0). The code came with %a/%A in [5cfe15b5d6]. | ||||
| User Comments: | ||||
serhiy.storchaka added on 2026-10-16 22:18:07:
Proposed fix in branch format-A-prefix: lowercase the "0X" prefix and the "P" exponent marker where they are, which is after the padding and the sign, and leave INF alone. Test format-19.5. jan.nijtmans added on 2026-10-17 06:54:38:
+1 Please go ahead serhiy.storchaka added on 2026-10-17 07:25:45:
Merged into main in [9336b96c43] and into 9.0 in [06f25062c2]. jan.nijtmans added on 2026-10-18 08:53:28:
Just one thing left to fix. With msvcrt (not ucrt): ==== format-19.5 %A with sign, padding and Inf, bug [2dff98e079] FAILED ==== Contents of test case:# explicit precision: without it MSVC prints all 13 hex digits list [format %.1A 1.5] [format %.1A -1.5] [format %+.0A 1.0] [format "% .0A" 1.0] [format %10.1A 1.5] [format %-10.0A| -2.0] [format %A Inf] [format %A -Inf] Matching with -glob would be easiest: If it contains 'INF' that should be OK jan.nijtmans added on 2026-10-18 08:59:20:
My proposal [6715b0e7eddf95f|here] (unless someone has a better idea) jan.nijtmans added on 2026-10-18 08:59:32:
My proposal [6715b0e7eddf95f|here] (unless someone has a better idea) jan.nijtmans added on 2026-10-18 09:00:57:
My proposal [96715b0e7e|here] (unless someone has a better idea) serhiy.storchaka added on 2026-10-18 09:21:47:
Another option, done in [f5bb9702a9]: format infinities ourselves instead of passing them to the C library, as Python does. msvcrt gets them wrong for all float conversions, not only %A (1.#INF00e+000, 1.#INF even for %G). Now they are inf or INF (for %A, %E, %G) with the sign flags, padded with spaces, as C99 specifies; the output on glibc and ucrt does not change. format-19.5 is exact again, new test format-19.6. Tested with a mingw msvcrt build on Windows 11: format.test passes, and fails without this change. NaN cannot get there, it is rejected earlier. jan.nijtmans added on 2026-10-18 21:07:22:
Agreed! Follow-up patch now merged too. Thanks! | ||||