Bug 253847
| Summary: | Printf does not work correctly for denormal doubles on non-x86 | ||
|---|---|---|---|
| Product: | Base System | Reporter: | Alex Richardson <arichardson> |
| Component: | standards | Assignee: | freebsd-standards (Nobody) <standards> |
| Status: | Closed FIXED | ||
| Severity: | Affects Only Me | ||
| Priority: | --- | ||
| Version: | CURRENT | ||
| Hardware: | Any | ||
| OS: | Any | ||
A commit in branch main references this bug: URL: https://cgit.FreeBSD.org/src/commit/?id=0e4ff0acbe80c547988bede738af2e227c7eb47c commit 0e4ff0acbe80c547988bede738af2e227c7eb47c Author: Alex Richardson <arichardson@FreeBSD.org> AuthorDate: 2021-03-01 14:27:30 +0000 Commit: Alex Richardson <arichardson@FreeBSD.org> CommitDate: 2021-03-01 14:27:30 +0000 AArch64: Don't set flush-subnormals-to-zero flag on startup This flag has been set on startup since 65618fdda0f272a823e6701966421bdca0efa301. However, This causes some of the math-related tests to fail as they report zero instead of a tiny number. This fixes at least /usr/tests/lib/msun/ldexp_test and possibly others. Additionally, setting this flag prevents printf() from printing subnormal numbers in decimal form. See also https://www.openwall.com/lists/musl/2021/02/26/1 PR: 253847 Reviewed By: mmel Differential Revision: https://reviews.freebsd.org/D28938 lib/libc/tests/stdio/printfloat_test.c | 35 +++++++++++++++++++++++++++++++++- sys/arm64/arm64/vm_machdep.c | 2 +- 2 files changed, 35 insertions(+), 2 deletions(-) A commit in branch stable/13 references this bug: URL: https://cgit.FreeBSD.org/src/commit/?id=d37fb0e37814db8f76462b3b9f1fb0e6dfca6324 commit d37fb0e37814db8f76462b3b9f1fb0e6dfca6324 Author: Alex Richardson <arichardson@FreeBSD.org> AuthorDate: 2021-03-01 14:27:30 +0000 Commit: Alex Richardson <arichardson@FreeBSD.org> CommitDate: 2021-03-17 09:45:56 +0000 AArch64: Don't set flush-subnormals-to-zero flag on startup This flag has been set on startup since 65618fdda0f272a823e6701966421bdca0efa301. However, This causes some of the math-related tests to fail as they report zero instead of a tiny number. This fixes at least /usr/tests/lib/msun/ldexp_test and possibly others. Additionally, setting this flag prevents printf() from printing subnormal numbers in decimal form. See also https://www.openwall.com/lists/musl/2021/02/26/1 PR: 253847 Reviewed By: mmel Differential Revision: https://reviews.freebsd.org/D28938 (cherry picked from commit 0e4ff0acbe80c547988bede738af2e227c7eb47c) lib/libc/tests/stdio/printfloat_test.c | 35 +++++++++++++++++++++++++++++++++- sys/arm64/arm64/vm_machdep.c | 2 +- 2 files changed, 35 insertions(+), 2 deletions(-) A commit in branch releng/13.0 references this bug: URL: https://cgit.FreeBSD.org/src/commit/?id=f979a90f4209418a26d858ca97b1497985bda0a7 commit f979a90f4209418a26d858ca97b1497985bda0a7 Author: Alex Richardson <arichardson@FreeBSD.org> AuthorDate: 2021-03-01 14:27:30 +0000 Commit: Ed Maste <emaste@FreeBSD.org> CommitDate: 2021-03-24 00:07:34 +0000 AArch64: Don't set flush-subnormals-to-zero flag on startup This flag has been set on startup since 65618fdda0f272a823e6701966421bdca0efa301. However, This causes some of the math-related tests to fail as they report zero instead of a tiny number. This fixes at least /usr/tests/lib/msun/ldexp_test and possibly others. Additionally, setting this flag prevents printf() from printing subnormal numbers in decimal form. See also https://www.openwall.com/lists/musl/2021/02/26/1 PR: 253847 Reviewed By: mmel Approved by: re (gjb) Differential Revision: https://reviews.freebsd.org/D28938 (cherry picked from commit 0e4ff0acbe80c547988bede738af2e227c7eb47c) (cherry picked from commit d37fb0e37814db8f76462b3b9f1fb0e6dfca6324) lib/libc/tests/stdio/printfloat_test.c | 35 +++++++++++++++++++++++++++++++++- sys/arm64/arm64/vm_machdep.c | 2 +- 2 files changed, 35 insertions(+), 2 deletions(-) |
Test code: ``` cat test2.cpp #include <stdio.h> #include <limits> int main() { // volatile double x = 0x1p-1537; volatile double x = std::numeric_limits<double>::denorm_min(); printf("x=%g/%a, as long double=%Lg/%La\n", x, x, (long double)x, (long double)x); } ``` Output on aarch64 x=0/0x1p-1537, as long double=4.94066e-324/0x1p-1074 Output on amd64: x=4.94066e-324/0x1p-1074, as long double=4.94066e-324/0x1p-1074 This causes issues with e.g. the msun ldexp test on AArch64. My guess would be that gdtoa depends on the x86 floating point arithmetic behaviour. I also tried updating gdtoa locally, but other than lots of merge conflicts that did not change anything.