Bug 282705

Comment 1 Ed Maste 2024-11-12 13:29:07 UTC

In lib/libcompiler_rt/Makefile:

# These are already shipped by libc.a on some architectures.
.if ${MACHINE_CPUARCH} != "arm" && ${MACHINE_CPUARCH} != "riscv"
...
SRCF+=          truncdfsf2

In lib/libc/softfloat/Symbol.map:

FBSDprivate_1.0 {
...
        __truncdfsf2;

I'm not sure where the implementation supposedly provided by libc comes/came from.

Comment 2 Dimitry Andric 2024-11-15 12:48:30 UTC

(In reply to Ed Maste from comment #1)
These are all in lib/libc/softfloat. However, the main libc Makefile has:

.if (${LIBC_ARCH} == "arm" && (defined(CPUTYPE) && ${CPUTYPE:M*soft*}))
.include "${LIBC_SRCTOP}/softfloat/Makefile.inc"
.endif

so it won't apply to arm CPUs with hardware floating point support. As far as I remember, the "# These are already shipped by libc.a on some architectures." comment has always been in the compiler-rt Makefile.inc.

In fact, they were there when Ed Schouten imported the libcompiler-rt build infra in <https://cgit.freebsd.org/src/commit/?id=a3cf0ef5a295c>.

It might be good to check with somebody who knows 32-bit arm, if we can change the check in Makefile.inc to be similar to the one in libc's Makefile, so these functions are included in compiler-rt from that point onward.

Comment 3 Dimitry Andric 2024-11-15 18:00:57 UTC

Looks like the rabbit hole goes a little deeper. We already compile VFP versions of these, via the Makefile.inc fragment:

.for file in ${SRCF}
.if ${MACHINE_CPUARCH} == "arm" && (!defined(CPUTYPE) || ${CPUTYPE:M*soft*} == "") \
    && exists(${CRTSRC}/${CRTARCH}/${file}vfp.S)
SRCS+=          ${file}vfp.S
. elif exists(${CRTSRC}/${CRTARCH}/${file}.S)
SRCS+=          ${file}.S
. else
SRCS+=          ${file}.c
. endif
.endfor

In contrib/llvm-project/compiler-rt/lib/builtins/arm there are:

adddf3vfp.S
addsf3vfp.S
divdf3vfp.S
divsf3vfp.S
extendsfdf2vfp.S
fixdfsivfp.S
fixsfsivfp.S
floatsidfvfp.S
floatsisfvfp.S
muldf3vfp.S
mulsf3vfp.S
subdf3vfp.S
subsf3vfp.S
truncdfsf2vfp.S

and a bunch more. However, llvm does not emit these on FreeBSD. In contrib/llvm-project/llvm/lib/Target/ARM/ARMISelLowering.cpp there is:

ARMTargetLowering::ARMTargetLowering(const TargetMachine &TM,
                                     const ARMSubtarget &STI)
...
  if (Subtarget->isTargetMachO()) {
    // Uses VFP for Thumb libfuncs if available.
    if (Subtarget->isThumb() && Subtarget->hasVFP2Base() &&
        Subtarget->hasARMOps() && !Subtarget->useSoftFloat()) {
      static const struct {
        const RTLIB::Libcall Op;
        const char * const Name;
        const ISD::CondCode Cond;
      } LibraryCalls[] = {
        // Single-precision floating-point arithmetic.
        { RTLIB::ADD_F32, "__addsf3vfp", ISD::SETCC_INVALID },
        { RTLIB::SUB_F32, "__subsf3vfp", ISD::SETCC_INVALID },
        { RTLIB::MUL_F32, "__mulsf3vfp", ISD::SETCC_INVALID },
        { RTLIB::DIV_F32, "__divsf3vfp", ISD::SETCC_INVALID },
...

I am unsure whether the VFP routines are ABI compatible with the non-VFP ones. If they are, we could alias them somehow. If they are not, we should stop compiling the VFP versions, and compile the regular versions.

Is the code using the wrong target, soft floating point (armv7-freebsd) instead of hardware FP (armv7-freebsd-gnueabihf)?  See, for example, bug 272806 and https://lists.freebsd.org/archives/freebsd-arm/2023-August/002933.html and replies.  FreeBSD-CURRENT no longer supports ARM versions older than armv7.