Tracking issue: 32bit x86 targets lose float NaN payload in return values

[RalfJung]

On x86 (32bit) targets, returning a floating-point number by-value can change the bits of their NaN payloads. This is caused by using x87 registers to pass return values, as mandated by the ABI for this target.

This is a known and long-standing problem, and very hard to fix. The purpose of this issue mostly is to document its existence and to give it a URL that can be referenced.

A complete fix is not possible since the ABI for extern "C" functions is not up for change. However we could change the default "Rust" ABI to not use the x87 registers ever if SSE is available. That would make the problem much less likely to appear in practice.

Related issues:

• x86-32: floating-point return values undergo implicit format conversion llvm/llvm-project#66803: the same bug on the LLVM side.
• Tracking issue: 32bit x86 targets without SSE2 have unsound floating point behavior #114479: that's a different problem, also affecting x86-32 but concerning the actual FP computations, not the NaN bits of return values.

Prior issues:

• i686 floating point behavior does not agree with unit tests in debug mode #73288
• Some num tests fail for {i586, i686, x86_64}-unknown-linux-gnu #46948

[Diggsey]

It looks like the 80-bit float load/store operations don't modify the bits, so in principle one could fix this by performing the conversion from 32-bit or 64-bit float to 80-bit float (and back) in software, although I expect that would be prohibitively expensive.

[RalfJung]

Could we do a quick check whether the float is a NaN or not, and only do the software conversion for NaNs? That should be a lot faster I hope.

[Diggsey]

Came up with this code for software conversion:

typedef union {
    long double ld;
    struct {
        unsigned long long ull;
        unsigned short us;
    };
} ld_t;

typedef union {
    double d;
    unsigned long long ull;
} d_t;

long double identity(double x) {
    d_t y;
    y.d = x;
    ld_t z;
    z.ull = (y.ull << 11) & 0x7FFFFFFFFFFFFFFFull | 0x8000000000000000ull;
    unsigned short e = (y.ull >> 52) & 0x7FF;
    z.us = ((e == 0x7FF) ? 0x7FFF : e + 0x3C00) | ((y.ull >> 48) & 0x8000);
    return z.ld;
}

identity:
        sub     esp, 28
        movq    xmm1, QWORD PTR [esp+32]
        movdqa  xmm0, xmm1
        psllq   xmm1, 11
        psrlq   xmm0, 32
        movd    eax, xmm0
        movdqa  xmm0, XMMWORD PTR .LC0
        mov     edx, eax
        shr     edx, 20
        por     xmm1, xmm0
        and     dx, 2047
        movq    QWORD PTR [esp], xmm1
        cmp     dx, 2047
        lea     ecx, [edx+15360]
        mov     edx, 32767
        cmovne  edx, ecx
        shr     eax, 16
        and     ax, -32768
        or      eax, edx
        mov     WORD PTR [esp+8], ax
        fld     TBYTE PTR [esp]
        add     esp, 28
        ret
.LC0:
        .long   0
        .long   -2147483648
        .long   0
        .long   0

Or, special casing NaNs:

#include "math.h"

typedef union {
    long double ld;
    struct {
        unsigned long long ull;
        unsigned short us;
    };
} ld_t;

typedef union {
    double d;
    unsigned long long ull;
} d_t;

long double identity(double x) {
    if (!isnan(x)) {
        return x;
    }
    d_t y;
    y.d = x;
    ld_t z;
    z.ull = (y.ull << 11) & 0x7FFFFFFFFFFFFFFFull | 0x8000000000000000ull;
    z.us = (y.ull >> 48) | 0x7FFF;
    return z.ld;
}

identity:
        sub     esp, 44
        fld     QWORD PTR [esp+48]
        fst     QWORD PTR [esp+8]
        fucomi  st, st(0)
        jp      .L2
        add     esp, 44
        ret
.L2:
        fstp    st(0)
        mov     eax, DWORD PTR [esp+8]
        mov     edx, DWORD PTR [esp+12]
        movd    xmm0, eax
        movd    xmm1, edx
        mov     eax, edx
        punpckldq       xmm0, xmm1
        shr     eax, 16
        movdqa  xmm1, XMMWORD PTR .LC0
        psllq   xmm0, 11
        or      ax, 32767
        por     xmm0, xmm1
        mov     WORD PTR [esp+24], ax
        movq    QWORD PTR [esp+16], xmm0
        fld     TBYTE PTR [esp+16]
        add     esp, 44
        ret
.LC0:
        .long   0
        .long   -2147483648
        .long   0
        .long   0```

[RalfJung]

Nice. :)

So, I wonder -- what would it take to actually use this when returning floats or doubles? Do we have to patch LLVM or can the codegen backend do this?

[RalfJung]

Btw, can someone explain why this only affects return values? How are float arguments passed?

[Diggsey]

So, I wonder -- what would it take to actually use this when returning floats or doubles? Do we have to patch LLVM or can the codegen backend do this?

No idea :)

Btw, can someone explain why this only affects return values? How are float arguments passed?

Arguments on x86 (with the C calling convention) are always passed on the stack regardless of type. So the problem still exists as soon as you fld the value, but it's not inherent to the calling convention.

[comex]

By the same token, for non-extern "C" functions, floating point returns could be changed to go on the stack, even without SSE.

[RalfJung]

Here there is a claim that this issue only affects signaling NaN return values. Is that true? The example given here does seem to be a signaling NaN.