[ibmibmibm]

This PR has 2 commits.

• The cmath page no longer shows an int* quo parameter for fdim, fmax and fmin, and calls the result of fdim the positive difference.
• fdim gives +0 for an infinite x that is not above y, and a NaN for an infinite x with a NaN y.
• fdim gives a quiet NaN for a signaling NaN argument.
• fdim no longer has its own NaN and infinity branch. The compare y >= x is false for a NaN, and x - y gives a quiet NaN and the correct result for an infinite x.
• A new test covers infinite arguments and signaling NaN arguments for the six types.

C17 7.12.12.1 gives x - y if x > y, and +0 if x <= y. IEEE 754-2019 6.2 and 7.2 give a quiet NaN for an operation on a signaling NaN. Intel libbid and glibc give the same results as the fix.

Fixes #1525

Correctness

• The reproducer of fdim gives an incorrect result for an infinite x and for a signaling NaN #1525 gives the correct results for all six types with the fix:

decimal32_t: fdim(-inf, 1) = 0, fdim(inf, inf) = 0, fdim(inf, qNaN) = nan, fdim(snan, 1) = qNaN, fdim(1, snan) = qNaN
decimal64_t: fdim(-inf, 1) = 0, fdim(inf, inf) = 0, fdim(inf, qNaN) = nan, fdim(snan, 1) = qNaN, fdim(1, snan) = qNaN
decimal128_t: fdim(-inf, 1) = 0, fdim(inf, inf) = 0, fdim(inf, qNaN) = nan, fdim(snan, 1) = qNaN, fdim(1, snan) = qNaN
decimal_fast32_t: fdim(-inf, 1) = 0, fdim(inf, inf) = 0, fdim(inf, qNaN) = nan, fdim(snan, 1) = qNaN, fdim(1, snan) = qNaN
decimal_fast64_t: fdim(-inf, 1) = 0, fdim(inf, inf) = 0, fdim(inf, qNaN) = nan, fdim(snan, 1) = qNaN, fdim(1, snan) = qNaN
decimal_fast128_t: fdim(-inf, 1) = 0, fdim(inf, inf) = 0, fdim(inf, qNaN) = nan, fdim(snan, 1) = qNaN, fdim(1, snan) = qNaN

• Compared with Intel libbid __bid32_fdim, __bid64_fdim and __bid128_fdim on 160000 pairs per type in the 5 rounding modes. The pairs include zeros, infinities, quiet and signaling NaNs, subnormals, the largest values, equal values and neighbors. develop gives 4060 (decimal32_t), 2025 (decimal64_t) and 2025 (decimal128_t) incorrect values. The fix gives 0. Two differences remain, and both are permitted:
  • A result can have the same value as libbid with a different exponent. This comes from operator-, not from fdim.
  • If x is a quiet NaN and y is a signaling NaN of the other sign, the fix gives the NaN of y, and libbid gives the NaN of x. IEEE 754-2019 6.2.3 permits either.
• github_issue_1525 fails on develop with 162 errors and passes with the fix.
• github_issue_1525 and test_cmath pass with GCC 16.2 and Clang 23.1, in C++14 and C++20, with -Wall -Wextra -Werror. The new test also passes with BOOST_DECIMAL_FAST_MATH.
• With BOOST_DECIMAL_FAST_MATH, the code does not change: the removed branch was already disabled in that mode.

Speed

Nanoseconds for each call, best of 3 runs of 1024 pairs, -O2, AMD Ryzen 9 3900X, one core. mix has random pairs, gt has x > y and le has x <= y:

type	inputs	GCC develop	GCC fix	Clang develop	Clang fix
decimal32_t	mix	8.7	9.3	10.8	10.8
decimal32_t	gt	13.7	14.6	17.8	17.7
decimal32_t	le	2.9	3.1	4.1	4.2
decimal64_t	mix	11.4	11.3	13.6	13.5
decimal64_t	gt	18.1	17.9	23.1	22.7
decimal64_t	le	4.5	4.4	4.9	4.9
decimal128_t	mix	23.6	24.7	27.9	27.9
decimal128_t	gt	40.8	42.0	46.8	46.5
decimal128_t	le	8.9	9.0	12.1	12.1
decimal_fast32_t	mix	8.7	9.0	8.1	7.6
decimal_fast32_t	gt	13.9	14.8	13.8	12.3
decimal_fast32_t	le	3.6	3.5	4.2	3.4
decimal_fast64_t	mix	10.0	10.1	10.6	9.0
decimal_fast64_t	gt	17.8	18.0	17.4	16.7
decimal_fast64_t	le	3.6	3.2	5.0	3.6
decimal_fast128_t	mix	15.6	16.1	15.6	15.3
decimal_fast128_t	gt	26.9	26.6	31.4	30.5
decimal_fast128_t	le	7.1	7.3	3.3	3.1

GCC is up to 7% slower for decimal32_t and decimal_fast32_t, and Clang is up to 28% faster for the fast types. Intel libbid __bid64_fdim takes 18.8 ns for the mix pairs on the same machine.

Size

.text bytes of one object file with six functions that call fdim, one for each type:

compiler	flags	develop	fix
GCC 16.2	-O2	8297	9329
GCC 16.2	-Os	5349	5716
Clang 23.1	-O2	863	896
Clang 23.1	-Os	261	307

A NaN now goes to operator-, so the code of operator- for values that are not finite stays in fdim.

[cppalliance-bot]

An automated preview of the documentation is available at https://1526.decimal.prtest3.cppalliance.org/libs/decimal/doc/html/index.html

If more commits are pushed to the pull request, the docs will rebuild at the same URL.

2026-10-21 21:25:17 UTC

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 98.5%. Comparing base (59c75da) to head (fcae535).

Additional details and impacted files

@@            Coverage Diff            @@
##           develop   #1526     +/-   ##
=========================================
+ Coverage     98.5%   98.5%   +0.1%     
=========================================
  Files          325     326      +1     
  Lines        27657   27689     +32     
  Branches      2300    2301      +1     
=========================================
+ Hits         27216   27248     +32     
  Misses         441     441             

Files with missing lines	Coverage Δ
include/boost/decimal/detail/cmath/fdim.hpp	100.0% <ø> (ø)
test/github_issue_1525.cpp	100.0% <100.0%> (ø)

---

Continue to review full report in Codecov by Harness.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 59c75da...fcae535. Read the comment docs.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[mborland]

Looks good. Thank you!