[reach2sayan]

This continues #61, by @SonnyChecksIn. In his review there, @Lastique pointed out that the merge GCC did on sqrt "can happen with any operation, not just the square root. And on any target, not just x86".

#61 fixed sqrt; this PR does the same for the rest of the rounded primitives, using detail::reload_operand introduced in #61

The problem

An interval function calls the _down and _up primitives on the same operands, with a rounding-mode change in between. The compiler does not treat the mode change as a dependency, so it can compute x + y once and use it for both ends. The interval is then too narrow and can miss the true value.

test/rounding_merge.cpp (new) checks add, sub, mul, div, int and conv for float, double and long double, under rounded_arith_std and rounded_arith_opp.
On develop, with -frounding-math:

Compiler	Merged on develop
g++ 13, x86-64, -O2/-O3	std + - * / (all types), std conv
g++ 13, -m32, -O2/-O3	also std int, and opp (default policy) long double + - * /
clang 18, x86-64, -O2/-O3	std long double + - * /
clang 18, -m32, -O2/-O3	also std int
clang 12, s390x, -O2/-O3	opp int_up returns the floor
any, -O0	none

Changes

• rounded_arith_std: + - * /, conv and int read one operand through detail::reload_operand, as sqrt already does after Stop g++ merging the two square roots of a point interval #61. int_down and int_up also go through force_rounding, like every other primitive.
• x87 long double (GCC and compatibles): force_rounding was an identity, so nothing kept the result before the next mode change. It is now an empty asm volatile with a "+t" operand and a "memory" clobber.
• test/sqrt_rounding.cpp passes the same operand to both roots, so it also catches a merge. It now runs long double everywhere except IBM double-double, which ignores the rounding mode.

CI

The multiarch setup step no longer installs qemu-user-static. On ubuntu-latest, now Ubuntu 26.04, it is a virtual package and apt fails, so the four QEMU jobs could not start. The host package was never used anyway. It registers binfmt_misc handlers that point at the host's QEMU binaries, but the next step, bdde-multiarch, runs tonistiigi/binfmt --uninstall "qemu-*", which removes them, and then --install all, which registers the QEMU binaries from that image instead.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 82.79%. Comparing base (27a6fdf) to head (a4d2a46).

Additional details and impacted files

@@             Coverage Diff             @@
##           develop      #65      +/-   ##
===========================================
+ Coverage    82.60%   82.79%   +0.18%     
===========================================
  Files           23       23              
  Lines         1006     1017      +11     
  Branches       410      412       +2     
===========================================
+ Hits           831      842      +11     
  Misses         167      167              
  Partials         8        8              

Files with missing lines	Coverage Δ
...t/numeric/interval/detail/x86_rounding_control.hpp	100.00% <100.00%> (ø)
...umeric/interval/detail/x86gcc_rounding_control.hpp	100.00% <100.00%> (ø)
include/boost/numeric/interval/rounded_arith.hpp	100.00% <100.00%> (ø)

... and 2 files with indirect coverage changes

---

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 27a6fdf...a4d2a46. Read the comment docs.

🚀 New features to boost your workflow:

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