[compiler-rt] __truncXfYf2 functions ignore the current rounding mode and always round to nearest, ties to even

[overmighty]

https://godbolt.org/z/dchq7rxe7

C code:

#include <fenv.h>
#include <stdio.h>

_Float16 bar(float x) {
  return (_Float16)x;
}

int main() {
  volatile float x = 0x1.0p-25f;
  fesetround(FE_UPWARD);
  printf("%a\n", (double)bar(x));
  fesetround(FE_DOWNWARD);
  printf("%a\n", (double)bar(x));
}

Output with libgcc:

0x1p-24
0x0p+0

Output with compiler-rt:

0x0p+0
0x0p+0

Relevant compiler-rt source:

llvm-project/compiler-rt/lib/builtins/truncsfhf2.c

Lines 15 to 17 in 046a177

	COMPILER_RT_ABI NOINLINE dst_t __truncsfhf2(float a) {
	return __truncXfYf2__(a);
	}

llvm-project/compiler-rt/lib/builtins/fp_trunc_impl.inc

Lines 9 to 12 in 046a177

	// This file implements a fairly generic conversion from a wider to a narrower
	// IEEE-754 floating-point type in the default (round to nearest, ties to even)
	// rounding mode. The constants and types defined following the includes below
	// parameterize the conversion.

cc @lntue

[llvmbot]

@llvm/issue-subscribers-libc

Author: OverMighty (overmighty)

https://godbolt.org/z/dchq7rxe7

C code:

#include <fenv.h>
#include <stdio.h>

_Float16 bar(float x) {
  return (_Float16)x;
}

int main() {
  volatile float x = 0x1.0p-25f;
  fesetround(FE_UPWARD);
  printf("%a\n", (double)bar(x));
  fesetround(FE_DOWNWARD);
  printf("%a\n", (double)bar(x));
}

Output with libgcc:

0x1p-24
0x0p+0

Output with compiler-rt:

0x0p+0
0x0p+0

Relevant compiler-rt source:

llvm-project/compiler-rt/lib/builtins/truncsfhf2.c

Lines 15 to 17 in 046a177

	COMPILER_RT_ABI NOINLINE dst_t __truncsfhf2(float a) {
	return __truncXfYf2__(a);
	}

llvm-project/compiler-rt/lib/builtins/fp_trunc_impl.inc

Lines 9 to 12 in 046a177

	// This file implements a fairly generic conversion from a wider to a narrower
	// IEEE-754 floating-point type in the default (round to nearest, ties to even)
	// rounding mode. The constants and types defined following the includes below
	// parameterize the conversion.

cc @lntue

[overmighty]

Added libc label as this impacts development of LLVM libc math functions.

[jhuber6]

Isn't it undefined to change the fpenv without setting #pragma STDC FENV_ACCESS?

[SEt-t]

The issue is still present with #pragma STDC FENV_ACCESS ON as library function just doesn't check current rounding mode at all.

A relatively simple workaround if you have hardware single->half conversion is to do direct double->half when rounding to even and otherwise do double->single->half.