[PATCH 0/1] MR8242: libs/musl: Force built-in rint for i386.

[Daniel Lehman (＠dlehman25)]

https://gitlab.winehq.org/wine/wine/-/merge_requests/7748#note_105461

huge thanks to @bernhardu for narrowing it down.  does this work for you?

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/8242

[PATCH 1/1] libs/musl: Force built-in rint for i386.

[Daniel Lehman]

From: Daniel Lehman <dlehman25(a)gmail.com>

---
 libs/musl/src/math/rint.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/libs/musl/src/math/rint.c b/libs/musl/src/math/rint.c
index ad35201b266..4c5c011fd17 100644
--- a/libs/musl/src/math/rint.c
+++ b/libs/musl/src/math/rint.c
@@ -12,6 +12,9 @@ static const double_t toint = 1/EPS;
 
 double __cdecl rint(double x)
 {
+#if defined(__GNUC__) && defined(__i386__)
+	return __builtin_rint(x);
+#endif
 	union {double f; uint64_t i;} u = {x};
 	int e = u.i>>52 & 0x7ff;
 	int s = u.i>>63;
-- 
GitLab

https://gitlab.winehq.org/wine/wine/-/merge_requests/8242

[Bernhard Übelacker]

> huge thanks to @bernhardu for narrowing it down. does this work for you?

Hello, I checked to receive this failure with my development tree **without** your patch (gcc-mingw-oldwow64 32-bit)
```
wine dlls/d3dcompiler_42/tests/i386-windows/d3dcompiler_42_test.exe hlsl_d3d9
...
hlsl_d3d9.c:713: Test failed: Test 24: Got {0.00000000e+000, 5.00000000e-001, 0.00000000e+000, 0.00000000e+000}, expected {-1.54972076e-006, 4.99991417e-001, 0.00000000e+000, 0.00000000e+000}.
```
And after applying your patch the failure vanished. So yes, **with** your patch it works for me, thanks.

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/8242#note_105799

[Piotr Caban (＠piotr)]

It will not work with other compilers as well. I would try doing something like:
```c
if ((_control87(0, 0) & _MCW_PC) == _PC_24) toint = 1 / FLT_EPSILON;
else toint = 1 / DBL_EPSILON;
```
It's not tested. If it doesn't work with 64-bit precision we will probably need  to implement it without floating point operations (or set/restore precision around the function).

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/8242#note_105900

[Daniel Lehman (＠dlehman25)]

> It will not work with other compilers as well

not sure what you mean.  the failure is specific to 32-bit gcc - should the fix be free of #ifdef's?
> I would try doing something like... It's not tested

will give that a try
> or set/restore precision around the function

had tried something like that, based on MSVCRT_rint, that also works:
https://gitlab.winehq.org/dlehman25/wine/-/commits/rint-cp
would that be ok? (full build fails, but works if i rebuild musl + msvcrt + d3dcompiler_4*)

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/8242#note_105987

[Piotr Caban (＠piotr)]

On Tue Jun 10 12:19:17 2025 +0000, Daniel Lehman wrote:
> > It will not work with other compilers as well
> > not sure what you mean.  the failure is specific to 32-bit gcc - should
> > the fix be free of #ifdef's?
> > I would try doing something like... It's not tested
> > will give that a try
> > or set/restore precision around the function
> > had tried something like that, based on MSVCRT_rint, that also works:
> > https://gitlab.winehq.org/dlehman25/wine/-/commits/rint-cp
> > would that be ok? (full build fails, but works if i rebuild musl +
> > msvcrt + d3dcompiler_4*)
> > The code is not working if floating point operations are done with non 53-bit precision. Any compiler that uses x87 instructions will be affected.

-- 
https://gitlab.winehq.org/wine/wine/-/merge_requests/8242#note_106033