multf3 has incorrect rounding of subnormals

[tgross35]

Testing with the following program:

#define __STDC_WANT_IEC_60559_TYPES_EXT__

#include <stdio.h>
#include <stdlib.h>
#include <inttypes.h>

#if defined(__clang__) && (defined(__i386) || defined(__x86_64))
#define _Float128 __float128
#endif

typedef struct {
    uint64_t lower, upper;
} u128;


void f128_print(_Float128 val) {
    u128 ival = *((u128 *)(&val));

    #ifndef __clang__
    char buf[1024];
    strfromf128(buf, sizeof(buf), "%.32g", val);
    printf("%#018" PRIx64 "%016" PRIx64 " %s\n", ival.upper, ival.lower, buf);
    #else
    printf("%#018" PRIx64 "%016" PRIx64 " %lf\n", ival.upper, ival.lower, (double)val);
    #endif
}

_Float128 new_f128(uint64_t upper, uint64_t lower) {
    u128 val;
    val.lower = lower;
    val.upper = upper;
    return *((_Float128 *)(&val));
}

int main() {
    _Float128 a = new_f128(0x00007fffffffffff, 0xffffffffffffffff);
    _Float128 b = new_f128(0x40007fffffffffff, 0xffffffffffffffff);
    f128_print(a);
    f128_print(b);
    _Float128 c = a * b;
    f128_print(c);

    return 0;
}

I am testing on aarch64. Using the built-in version of Clang that links against system libraries gets the multf3 symbol from either /usr/lib/gcc/aarch64-linux-gnu/11/libgcc.a or /usr/lib/llvm-14/lib/clang/14.0.0/lib/linux/libclang_rt.builtins-aarch64.a. This is the output, which is correct:

$ clang f128_demo.c -o f128_demo
$ ./f128_demo
0x00007fffffffffffffffffffffffffff 0.000000
0x40007fffffffffffffffffffffffffff 3.000000
0x00017ffffffffffffffffffffffffffc 0.000000

However, building with the current multf3.c produces an incorrect value (note the ending b):

$ clang f128_demo.c llvm-project/compiler-rt/lib/builtins/multf3.c -o f128_demo_current_rt
$ ./f128_demo_current_rt
0x00007fffffffffffffffffffffffffff 0.000000
0x40007fffffffffffffffffffffffffff 3.000000
0x00017ffffffffffffffffffffffffffb 0.000000

This was just tested in a ubuntu docker container which has Clang 14.0.0 and GCC 11.4.0.

I am not sure why this appears to be aarch64 only. reproduced on x86. See some more background at rust-lang/compiler-builtins#607.

[llvmbot]

@llvm/issue-subscribers-backend-aarch64

Author: Trevor Gross (tgross35)

Testing with the following program:

#define __STDC_WANT_IEC_60559_TYPES_EXT__

#include <stdio.h>
#include <stdlib.h>
#include <inttypes.h>

#if defined(__clang__) && (defined(__i386) || defined(__x86_64))
#define _Float128 __float128
#endif

typedef struct {
    uint64_t lower, upper;
} u128;


void f128_print(_Float128 val) {
    u128 ival = *((u128 *)(&val));

    #ifndef __clang__
    char buf[1024];
    strfromf128(buf, sizeof(buf), "%.32g", val);
    printf("%#<!-- -->018" PRIx64 "%016" PRIx64 " %s\n", ival.upper, ival.lower, buf);
    #else
    printf("%#<!-- -->018" PRIx64 "%016" PRIx64 " %lf\n", ival.upper, ival.lower, (double)val);
    #endif
}

_Float128 new_f128(uint64_t upper, uint64_t lower) {
    u128 val;
    val.lower = lower;
    val.upper = upper;
    return *((_Float128 *)(&amp;val));
}

int main() {
    _Float128 a = new_f128(0x00007fffffffffff, 0xffffffffffffffff);
    _Float128 b = new_f128(0x40007fffffffffff, 0xffffffffffffffff);
    f128_print(a);
    f128_print(b);
    _Float128 c = a * b;
    f128_print(c);

    return 0;
}

Using the builtin version of Clang that links against system libraries gets the multf3 symbol from either /usr/lib/gcc/aarch64-linux-gnu/11/libgcc.a or /usr/lib/llvm-14/lib/clang/14.0.0/lib/linux/libclang_rt.builtins-aarch64.a. This is the output, which is correct:

$ clang f128_demo.c -o f128_demo
$ ./f128_demo
0x00007fffffffffffffffffffffffffff 0.000000
0x40007fffffffffffffffffffffffffff 3.000000
0x00017ffffffffffffffffffffffffffc 0.000000

However, building with the current multf3.c produces an incorrect value (note the ending b):

$ clang f128_demo.c llvm-project/compiler-rt/lib/builtins/multf3.c -o f128_demo_current_rt
$ ./f128_demo_current_rt
0x00007fffffffffffffffffffffffffff 0.000000
0x40007fffffffffffffffffffffffffff 3.000000
0x00017ffffffffffffffffffffffffffb 0.000000

This was just tested in a ubuntu docker container which has clang 14.0.0 and gcc 11.4.0

[beetrees]

I've reproduced this bug on x86_64-unknown-linux-gnu.

[tgross35]

@EugeneZelenko since this is no longer aarch64-only, the labels could probably be updated

[beetrees]

The bug is occurring within the wideMultiply (no rounding is involved). Specifically: 0x0001ffff_ffffffff_ffffffff_fffffffc * 0xbfffffff_ffffffff_ffffffff_ffff8000 should result in a productHi of 0x00017fff_ffffffff_ffffffff_fffffffc but instead results in a productHi of 0x00017fff_ffffffff_ffffffff_fffffffb. A quick search of the issue tracker suggests it might be related to #90009.