[compiler-rt] [builtins] Implement __floattitf() & __floatuntitf()

[mgorny]

Summary

Implement the missing floattitf() and floatuntitf() functions, to
convert 128-bit (unsigned) integers to quad-precision floating-point
types. This is needed e.g. on AArch64 where 'long double' is
a quad-precision type.

The code is based on the existing code for floattixf()
and floatuntixf(), updated to account for different bit field lengths
of quad-precision float. The tests are also copied, with the rounding
tests adjusted for longer significand.

(tested on AArch64)

[weimingz]

The code looks OK except for the formatting. Since the logic is almost the same as floattixf / floatuntixf except that some constants are different, I'm wondering if it makes sense to reuse the existing code? For example, make existing code as floattixf_impl.inc

lib/builtins/floattitf.c

32 ↗

(On Diff #81876)

Please put "{" on the same line like __floattitf(ti_int a) {
Same for other occurrences below

[mgorny]

I think I've shifted all of {/} to be in line with keywords.

As for the splitting the code, you're probably right. However, I'd rather do that in a separate patch (I'll add it to my TODO, I promise).

I think that most of those files reuse a similar code -- I think it's for the case when the integer has greater range than the float. It would certainly also make sense to try to commonize the code used for signed and unsigned. However, I'm a little worried that it might make the result unreadable.

[weimingz]

Thanks for the formatting. Regarding the refactoring, Regarding factoring out common code, I think it's OK to do it in a follow-up patch. For readability, it should be OK. For example, fp_add_impl.inc is used by float, double and long double.

[weimingz]

The patch looks good to me.

[rengolin]

My floating-point-fu is lacking, so I'll trust you guys that this makes sense. :)

I didn't see anything egregious, so LGTM with the nit inline.

cheers,
--renato

lib/builtins/floatuntitf.c

73 ↗

(On Diff #82235)

nit: an hexadecimal pattern here would be clearer. same above.

[mgorny]

lib/builtins/floatuntitf.c

73 ↗

(On Diff #82235)

Do you mean something like: (foo << 48) & 0xffff....?

[rengolin]

lib/builtins/floatuntitf.c

73 ↗

(On Diff #82235)

No, just the 16383 to 0x3FFF

[scanon]

Please s/mantissa/significand/. I know that "mantissa" is used in a number of places in llvm sources, but it's incorrect terminology. Significand is the IEEE-754 nomenclature, which any new work should follow.

[rengolin]

lib/builtins/floatuntitf.c

73 ↗

(On Diff #82235)

Though, now I think your proposal may be better. :)

Whatever works, I'm not too concerned about that.

[scanon]

lib/builtins/floattitf.c

65 ↗

(On Diff #82235)

Strictly speaking there's no need to adjust a here. If we rounded up into a new binade, then a is necessarily 0b1000...0, and the leading 1 bit will get killed by the mask when we assemble fb.u.high.all regardless of its position. Same comment applies to floatuntitf.

[mgorny]

lib/builtins/floatuntitf.c

73 ↗

(On Diff #82235)

Well, I used the decimal form because that's how the IEEE 754 standard specifies it, so I guessed the correlation would be clearer that way.

[scanon]

lib/builtins/floatuntitf.c

73 ↗

(On Diff #82235)

FWIW, I think both are pretty obvious, but if you want to be totally explicit:

const int bias = 0x3fff;
... (e + bias) ...

[mgorny]

lib/builtins/floattitf.c

65 ↗

(On Diff #82235)

I'm sorry but I don't feel confident changing that. AFAIU if the LDBL_MANT_DIG+1 bit is set, this code shifts it lower, so it won't actually be killed by the mask.

[scanon]

lib/builtins/floattitf.c

65 ↗

(On Diff #82235)

In binary128, as in all IEEE 754 binary interchange format encodings, the leading bit of the significand is implicit. The only way to end up in this code path is 0b111...1 rounding up to 0b100...00, meaning that the significand is 1.0, which is stored as all-zeros (i.e. the leading bit is necessarily masked).

To be more explicit, LDBL_MANT_DIG is 113. If this shift happens, after the shift bit 112 is set, and bits 111:0 are zero. The mask ((a >> 64) & 0x0000ffffffffffffLL) discards bit 112 (= 64 + 48).

[scanon]

I don't particularly care about the shift, since the code is completely equivalent either way, though it would be nice to clean up. However, please replace "mantissa" with "significand" before committing.

[mgorny]

Changed mantissa -> significand.

[mgorny]

lib/builtins/floattitf.c

65 ↗

(On Diff #82235)

Well, I've tried removing this and it causes one of the tests to fail:

error in __floatuntitf(0xFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFFF) = 0X1P+127, expected 0X1P+128

[scanon]

lib/builtins/floattitf.c

65 ↗

(On Diff #82235)

This sounds like you removed the exponent adjustment (++e) as well as (or instead of) the shift (a >>= 1). Without seeing the change, I can't be certain, of course.

[mgorny]

lib/builtins/floattitf.c

65 ↗

(On Diff #82235)

Yes, I did that. Now that you state that plainly, I finally understand what you meant ;-). However, if I were to do that, I would rather to do that separately, either in all relevant files or post unifying the code into one logical .inc. This would keep the history of changes cleaner.

[scanon]

OK, I'm satisfied with that.

compiler-rt/trunk/lib/builtins/CMakeLists.txt

set(aarch64_SOURCES
   fixtfdi.c
   fixtfsi.c
   fixtfti.c
   fixunstfdi.c
   fixunstfsi.c
   fixunstfti.c
   floatditf.c
   floatsitf.c
+  floattitf.c
   floatunditf.c
   floatunsitf.c
+  floatuntitf.c
   multc3.c
   trunctfdf2.c
   trunctfsf2.c
   ${GENERIC_SOURCES})
 
 set(armhf_SOURCES ${arm_SOURCES})
 set(armv7_SOURCES ${arm_SOURCES})
 set(armv7s_SOURCES ${arm_SOURCES})

compiler-rt/trunk/lib/builtins/floattitf.c

+//===-- lib/floattitf.c - int128 -> quad-precision conversion -----*- C -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements ti_int to quad-precision conversion for the
+// compiler-rt library in the IEEE-754 default round-to-nearest, ties-to-even
+// mode.
+//
+//===----------------------------------------------------------------------===//
+
+#define QUAD_PRECISION
+#include "fp_lib.h"
+#include "int_lib.h"
+
+/* Returns: convert a ti_int to a fp_t, rounding toward even. */
+
+/* Assumption: fp_t is a IEEE 128 bit floating point type
+ *             ti_int is a 128 bit integral type
+ */
+
+/* seee eeee eeee eeee mmmm mmmm mmmm mmmm | mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm |
+ * mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm | mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm
+ */
+
+#if defined(CRT_HAS_128BIT) && defined(CRT_LDBL_128BIT)
+COMPILER_RT_ABI fp_t
+__floattitf(ti_int a) {
+    if (a == 0)
+        return 0.0;
+    const unsigned N = sizeof(ti_int) * CHAR_BIT;
+    const ti_int s = a >> (N-1);
+    a = (a ^ s) - s;
+    int sd = N - __clzti2(a);  /* number of significant digits */
+    int e = sd - 1;            /* exponent */
+    if (sd > LDBL_MANT_DIG) {
+        /*  start:  0000000000000000000001xxxxxxxxxxxxxxxxxxxxxxPQxxxxxxxxxxxxxxxxxx
+         *  finish: 000000000000000000000000000000000000001xxxxxxxxxxxxxxxxxxxxxxPQR
+         *                                                12345678901234567890123456
+         *  1 = msb 1 bit
+         *  P = bit LDBL_MANT_DIG-1 bits to the right of 1
+         *  Q = bit LDBL_MANT_DIG bits to the right of 1
+         *  R = "or" of all bits to the right of Q
+         */
+        switch (sd) {
+        case LDBL_MANT_DIG + 1:
+            a <<= 1;
+            break;
+        case LDBL_MANT_DIG + 2:
+            break;
+        default:
+            a = ((tu_int)a >> (sd - (LDBL_MANT_DIG+2))) |
+                ((a & ((tu_int)(-1) >> ((N + LDBL_MANT_DIG+2) - sd))) != 0);
+        };
+        /* finish: */
+        a |= (a & 4) != 0;  /* Or P into R */
+        ++a;  /* round - this step may add a significant bit */
+        a >>= 2;  /* dump Q and R */
+        /* a is now rounded to LDBL_MANT_DIG or LDBL_MANT_DIG+1 bits */
+        if (a & ((tu_int)1 << LDBL_MANT_DIG)) {
+            a >>= 1;
+            ++e;
+        }
+        /* a is now rounded to LDBL_MANT_DIG bits */
+    } else {
+        a <<= (LDBL_MANT_DIG - sd);
+        /* a is now rounded to LDBL_MANT_DIG bits */
+    }
+
+    long_double_bits fb;
+    fb.u.high.all = (s & 0x8000000000000000LL)           /* sign */
+                  | (du_int)(e + 16383) << 48            /* exponent */
+                  | ((a >> 64) & 0x0000ffffffffffffLL);  /* significand */
+    fb.u.low.all = (du_int)(a);
+    return fb.f;
+}
+
+#endif

compiler-rt/trunk/lib/builtins/floatuntitf.c

+//===-- lib/floatuntitf.c - uint128 -> quad-precision conversion --*- C -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file implements tu_int to quad-precision conversion for the
+// compiler-rt library in the IEEE-754 default round-to-nearest, ties-to-even
+// mode.
+//
+//===----------------------------------------------------------------------===//
+
+#define QUAD_PRECISION
+#include "fp_lib.h"
+#include "int_lib.h"
+
+/* Returns: convert a tu_int to a fp_t, rounding toward even. */
+
+/* Assumption: fp_t is a IEEE 128 bit floating point type
+ *             tu_int is a 128 bit integral type
+ */
+
+/* seee eeee eeee eeee mmmm mmmm mmmm mmmm | mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm |
+ * mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm | mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm
+ */
+
+#if defined(CRT_HAS_128BIT) && defined(CRT_LDBL_128BIT)
+COMPILER_RT_ABI fp_t
+__floatuntitf(tu_int a) {
+    if (a == 0)
+        return 0.0;
+    const unsigned N = sizeof(tu_int) * CHAR_BIT;
+    int sd = N - __clzti2(a);  /* number of significant digits */
+    int e = sd - 1;            /* exponent */
+    if (sd > LDBL_MANT_DIG) {
+        /*  start:  0000000000000000000001xxxxxxxxxxxxxxxxxxxxxxPQxxxxxxxxxxxxxxxxxx
+         *  finish: 000000000000000000000000000000000000001xxxxxxxxxxxxxxxxxxxxxxPQR
+         *                                                12345678901234567890123456
+         *  1 = msb 1 bit
+         *  P = bit LDBL_MANT_DIG-1 bits to the right of 1
+         *  Q = bit LDBL_MANT_DIG bits to the right of 1
+         *  R = "or" of all bits to the right of Q
+         */
+        switch (sd) {
+        case LDBL_MANT_DIG + 1:
+            a <<= 1;
+            break;
+        case LDBL_MANT_DIG + 2:
+            break;
+        default:
+            a = (a >> (sd - (LDBL_MANT_DIG+2))) |
+                ((a & ((tu_int)(-1) >> ((N + LDBL_MANT_DIG+2) - sd))) != 0);
+        };
+        /* finish: */
+        a |= (a & 4) != 0;  /* Or P into R */
+        ++a;  /* round - this step may add a significant bit */
+        a >>= 2;  /* dump Q and R */
+        /* a is now rounded to LDBL_MANT_DIG or LDBL_MANT_DIG+1 bits */
+        if (a & ((tu_int)1 << LDBL_MANT_DIG)) {
+            a >>= 1;
+            ++e;
+        }
+        /* a is now rounded to LDBL_MANT_DIG bits */
+    } else {
+        a <<= (LDBL_MANT_DIG - sd);
+        /* a is now rounded to LDBL_MANT_DIG bits */
+    }
+
+    long_double_bits fb;
+    fb.u.high.all = (du_int)(e + 16383) << 48            /* exponent */
+                  | ((a >> 64) & 0x0000ffffffffffffLL);  /* significand */
+    fb.u.low.all = (du_int)(a);
+    return fb.f;
+}
+
+#endif

compiler-rt/trunk/test/builtins/Unit/floattitf_test.c

+//===-- floattitf.c - Test __floattitf ------------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file tests __floattitf for the compiler_rt library.
+//
+//===----------------------------------------------------------------------===//
+
+#define QUAD_PRECISION
+#include "fp_lib.h"
+#include "int_lib.h"
+#include <float.h>
+#include <stdio.h>
+
+#if defined(CRT_HAS_128BIT) && defined(CRT_LDBL_128BIT)
+
+/* Returns: convert a ti_int to a fp_t, rounding toward even. */
+
+/* Assumption: fp_t is a IEEE 128 bit floating point type
+ *             ti_int is a 128 bit integral type
+ */
+
+/* seee eeee eeee eeee mmmm mmmm mmmm mmmm | mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm |
+ * mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm | mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm
+ */
+
+COMPILER_RT_ABI fp_t __floattitf(ti_int a);
+
+int test__floattitf(ti_int a, fp_t expected) {
+    fp_t x = __floattitf(a);
+    if (x != expected) {
+        twords at;
+        at.all = a;
+        printf("error in __floattitf(0x%.16llX%.16llX) = %LA, expected %LA\n",
+               at.s.high, at.s.low, x, expected);
+    }
+    return x != expected;
+}
+
+char assumption_1[sizeof(ti_int) == 2*sizeof(di_int)] = {0};
+char assumption_2[sizeof(ti_int)*CHAR_BIT == 128] = {0};
+char assumption_3[sizeof(fp_t)*CHAR_BIT == 128] = {0};
+
+#endif
+
+int main() {
+#if defined(CRT_HAS_128BIT) && defined(CRT_LDBL_128BIT)
+    if (test__floattitf(0, 0.0))
+        return 1;
+
+    if (test__floattitf(1, 1.0))
+        return 1;
+    if (test__floattitf(2, 2.0))
+        return 1;
+    if (test__floattitf(20, 20.0))
+        return 1;
+    if (test__floattitf(-1, -1.0))
+        return 1;
+    if (test__floattitf(-2, -2.0))
+        return 1;
+    if (test__floattitf(-20, -20.0))
+        return 1;
+
+    if (test__floattitf(0x7FFFFF8000000000LL, 0x1.FFFFFEp+62))
+        return 1;
+    if (test__floattitf(0x7FFFFFFFFFFFF800LL, 0x1.FFFFFFFFFFFFEp+62))
+        return 1;
+    if (test__floattitf(0x7FFFFF0000000000LL, 0x1.FFFFFCp+62))
+        return 1;
+    if (test__floattitf(0x7FFFFFFFFFFFF000LL, 0x1.FFFFFFFFFFFFCp+62))
+        return 1;
+
+    if (test__floattitf(make_ti(0x8000008000000000LL, 0), -0x1.FFFFFEp+126))
+        return 1;
+    if (test__floattitf(make_ti(0x8000000000000800LL, 0), -0x1.FFFFFFFFFFFFEp+126))
+        return 1;
+    if (test__floattitf(make_ti(0x8000010000000000LL, 0), -0x1.FFFFFCp+126))
+        return 1;
+    if (test__floattitf(make_ti(0x8000000000001000LL, 0), -0x1.FFFFFFFFFFFFCp+126))
+        return 1;
+
+    if (test__floattitf(make_ti(0x8000000000000000LL, 0), -0x1.000000p+127))
+        return 1;
+    if (test__floattitf(make_ti(0x8000000000000001LL, 0), -0x1.FFFFFFFFFFFFFFFCp+126L))
+        return 1;
+
+    if (test__floattitf(0x0007FB72E8000000LL, 0x1.FEDCBAp+50))
+        return 1;
+
+    if (test__floattitf(0x0007FB72EA000000LL, 0x1.FEDCBA8p+50))
+        return 1;
+    if (test__floattitf(0x0007FB72EB000000LL, 0x1.FEDCBACp+50))
+        return 1;
+    if (test__floattitf(0x0007FB72EBFFFFFFLL, 0x1.FEDCBAFFFFFFCp+50))
+        return 1;
+    if (test__floattitf(0x0007FB72EC000000LL, 0x1.FEDCBBp+50))
+        return 1;
+    if (test__floattitf(0x0007FB72E8000001LL, 0x1.FEDCBA0000004p+50))
+        return 1;
+
+    if (test__floattitf(0x0007FB72E6000000LL, 0x1.FEDCB98p+50))
+        return 1;
+    if (test__floattitf(0x0007FB72E7000000LL, 0x1.FEDCB9Cp+50))
+        return 1;
+    if (test__floattitf(0x0007FB72E7FFFFFFLL, 0x1.FEDCB9FFFFFFCp+50))
+        return 1;
+    if (test__floattitf(0x0007FB72E4000001LL, 0x1.FEDCB90000004p+50))
+        return 1;
+    if (test__floattitf(0x0007FB72E4000000LL, 0x1.FEDCB9p+50))
+        return 1;
+
+    if (test__floattitf(0x023479FD0E092DC0LL, 0x1.1A3CFE870496Ep+57))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DA1LL, 0x1.1A3CFE870496D08p+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DB0LL, 0x1.1A3CFE870496D8p+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DB8LL, 0x1.1A3CFE870496DCp+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DB6LL, 0x1.1A3CFE870496DBp+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DBFLL, 0x1.1A3CFE870496DF8p+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DC1LL, 0x1.1A3CFE870496E08p+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DC7LL, 0x1.1A3CFE870496E38p+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DC8LL, 0x1.1A3CFE870496E4p+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DCFLL, 0x1.1A3CFE870496E78p+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DD0LL, 0x1.1A3CFE870496E8p+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DD1LL, 0x1.1A3CFE870496E88p+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DD8LL, 0x1.1A3CFE870496ECp+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DDFLL, 0x1.1A3CFE870496EF8p+57L))
+        return 1;
+    if (test__floattitf(0x023479FD0E092DE0LL, 0x1.1A3CFE870496Fp+57))
+        return 1;
+
+    if (test__floattitf(make_ti(0x023479FD0E092DC0LL, 0), 0x1.1A3CFE870496Ep+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DA1LL, 1), 0x1.1A3CFE870496D08p+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DB0LL, 2), 0x1.1A3CFE870496D8p+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DB8LL, 3), 0x1.1A3CFE870496DCp+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DB6LL, 4), 0x1.1A3CFE870496DBp+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DBFLL, 5), 0x1.1A3CFE870496DF8p+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DC1LL, 6), 0x1.1A3CFE870496E08p+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DC7LL, 7), 0x1.1A3CFE870496E38p+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DC8LL, 8), 0x1.1A3CFE870496E4p+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DCFLL, 9), 0x1.1A3CFE870496E78p+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DD0LL, 0), 0x1.1A3CFE870496E8p+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DD1LL, 11), 0x1.1A3CFE870496E88p+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DD8LL, 12), 0x1.1A3CFE870496ECp+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DDFLL, 13), 0x1.1A3CFE870496EF8p+121L))
+        return 1;
+    if (test__floattitf(make_ti(0x023479FD0E092DE0LL, 14), 0x1.1A3CFE870496Fp+121L))
+        return 1;
+
+    if (test__floattitf(make_ti(0, 0xFFFFFFFFFFFFFFFFLL), 0x1.FFFFFFFFFFFFFFFEp+63L))
+        return 1;
+
+    if (test__floattitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC2801LL),
+                        0x1.23456789ABCDEF0123456789ABC3p+124L))
+        return 1;
+    if (test__floattitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC3000LL),
+                        0x1.23456789ABCDEF0123456789ABC3p+124L))
+        return 1;
+    if (test__floattitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC37FFLL),
+                        0x1.23456789ABCDEF0123456789ABC3p+124L))
+        return 1;
+    if (test__floattitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC3800LL),
+                        0x1.23456789ABCDEF0123456789ABC4p+124L))
+        return 1;
+    if (test__floattitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC4000LL),
+                        0x1.23456789ABCDEF0123456789ABC4p+124L))
+        return 1;
+    if (test__floattitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC47FFLL),
+                        0x1.23456789ABCDEF0123456789ABC4p+124L))
+        return 1;
+    if (test__floattitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC4800LL),
+                        0x1.23456789ABCDEF0123456789ABC4p+124L))
+        return 1;
+    if (test__floattitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC4801LL),
+                        0x1.23456789ABCDEF0123456789ABC5p+124L))
+        return 1;
+    if (test__floattitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC57FFLL),
+                        0x1.23456789ABCDEF0123456789ABC5p+124L))
+        return 1;
+#else
+    printf("skipped\n");
+#endif
+   return 0;
+}

compiler-rt/trunk/test/builtins/Unit/floatuntitf_test.c

+//===-- floatuntitf.c - Test __floatuntitf --------------------------------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is dual licensed under the MIT and the University of Illinois Open
+// Source Licenses. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file tests __floatuntitf for the compiler_rt library.
+//
+//===----------------------------------------------------------------------===//
+
+#define QUAD_PRECISION
+#include "fp_lib.h"
+#include "int_lib.h"
+#include <float.h>
+#include <stdio.h>
+
+#if defined(CRT_HAS_128BIT) && defined(CRT_LDBL_128BIT)
+
+/* Returns: convert a tu_int to a fp_t, rounding toward even. */
+
+/* Assumption: fp_t is a IEEE 128 bit floating point type
+ *             tu_int is a 128 bit integral type
+ */
+
+/* seee eeee eeee eeee mmmm mmmm mmmm mmmm | mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm |
+ * mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm | mmmm mmmm mmmm mmmm mmmm mmmm mmmm mmmm
+ */
+
+COMPILER_RT_ABI fp_t __floatuntitf(tu_int a);
+
+int test__floatuntitf(tu_int a, fp_t expected) {
+    fp_t x = __floatuntitf(a);
+    if (x != expected) {
+        utwords at;
+        at.all = a;
+        printf("error in __floatuntitf(0x%.16llX%.16llX) = %LA, expected %LA\n",
+               at.s.high, at.s.low, x, expected);
+    }
+    return x != expected;
+}
+
+char assumption_1[sizeof(tu_int) == 2*sizeof(du_int)] = {0};
+char assumption_2[sizeof(tu_int)*CHAR_BIT == 128] = {0};
+char assumption_3[sizeof(fp_t)*CHAR_BIT == 128] = {0};
+
+#endif
+
+int main() {
+#if defined(CRT_HAS_128BIT) && defined(CRT_LDBL_128BIT)
+    if (test__floatuntitf(0, 0.0))
+        return 1;
+
+    if (test__floatuntitf(1, 1.0))
+        return 1;
+    if (test__floatuntitf(2, 2.0))
+        return 1;
+    if (test__floatuntitf(20, 20.0))
+        return 1;
+
+    if (test__floatuntitf(0x7FFFFF8000000000ULL, 0x1.FFFFFEp+62))
+        return 1;
+    if (test__floatuntitf(0x7FFFFFFFFFFFF800ULL, 0x1.FFFFFFFFFFFFEp+62))
+        return 1;
+    if (test__floatuntitf(0x7FFFFF0000000000ULL, 0x1.FFFFFCp+62))
+        return 1;
+    if (test__floatuntitf(0x7FFFFFFFFFFFF000ULL, 0x1.FFFFFFFFFFFFCp+62))
+        return 1;
+    if (test__floatuntitf(0x7FFFFFFFFFFFFFFFULL, 0xF.FFFFFFFFFFFFFFEp+59L))
+        return 1;
+    if (test__floatuntitf(0xFFFFFFFFFFFFFFFEULL, 0xF.FFFFFFFFFFFFFFEp+60L))
+        return 1;
+    if (test__floatuntitf(0xFFFFFFFFFFFFFFFFULL, 0xF.FFFFFFFFFFFFFFFp+60L))
+        return 1;
+
+    if (test__floatuntitf(0x8000008000000000ULL, 0x8.000008p+60))
+        return 1;
+    if (test__floatuntitf(0x8000000000000800ULL, 0x8.0000000000008p+60))
+        return 1;
+    if (test__floatuntitf(0x8000010000000000ULL, 0x8.00001p+60))
+        return 1;
+    if (test__floatuntitf(0x8000000000001000ULL, 0x8.000000000001p+60))
+        return 1;
+
+    if (test__floatuntitf(0x8000000000000000ULL, 0x8p+60))
+        return 1;
+    if (test__floatuntitf(0x8000000000000001ULL, 0x8.000000000000001p+60L))
+        return 1;
+
+    if (test__floatuntitf(0x0007FB72E8000000LL, 0x1.FEDCBAp+50))
+        return 1;
+
+    if (test__floatuntitf(0x0007FB72EA000000LL, 0x1.FEDCBA8p+50))
+        return 1;
+    if (test__floatuntitf(0x0007FB72EB000000LL, 0x1.FEDCBACp+50))
+        return 1;
+    if (test__floatuntitf(0x0007FB72EBFFFFFFLL, 0x1.FEDCBAFFFFFFCp+50))
+        return 1;
+    if (test__floatuntitf(0x0007FB72EC000000LL, 0x1.FEDCBBp+50))
+        return 1;
+    if (test__floatuntitf(0x0007FB72E8000001LL, 0x1.FEDCBA0000004p+50))
+        return 1;
+
+    if (test__floatuntitf(0x0007FB72E6000000LL, 0x1.FEDCB98p+50))
+        return 1;
+    if (test__floatuntitf(0x0007FB72E7000000LL, 0x1.FEDCB9Cp+50))
+        return 1;
+    if (test__floatuntitf(0x0007FB72E7FFFFFFLL, 0x1.FEDCB9FFFFFFCp+50))
+        return 1;
+    if (test__floatuntitf(0x0007FB72E4000001LL, 0x1.FEDCB90000004p+50))
+        return 1;
+    if (test__floatuntitf(0x0007FB72E4000000LL, 0x1.FEDCB9p+50))
+        return 1;
+
+    if (test__floatuntitf(0x023479FD0E092DC0LL, 0x1.1A3CFE870496Ep+57))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DA1LL, 0x1.1A3CFE870496D08p+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DB0LL, 0x1.1A3CFE870496D8p+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DB8LL, 0x1.1A3CFE870496DCp+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DB6LL, 0x1.1A3CFE870496DBp+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DBFLL, 0x1.1A3CFE870496DF8p+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DC1LL, 0x1.1A3CFE870496E08p+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DC7LL, 0x1.1A3CFE870496E38p+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DC8LL, 0x1.1A3CFE870496E4p+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DCFLL, 0x1.1A3CFE870496E78p+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DD0LL, 0x1.1A3CFE870496E8p+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DD1LL, 0x1.1A3CFE870496E88p+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DD8LL, 0x1.1A3CFE870496ECp+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DDFLL, 0x1.1A3CFE870496EF8p+57L))
+        return 1;
+    if (test__floatuntitf(0x023479FD0E092DE0LL, 0x1.1A3CFE870496Fp+57))
+        return 1;
+
+    if (test__floatuntitf(make_ti(0x023479FD0E092DC0LL, 0), 0x1.1A3CFE870496Ep+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DA1LL, 1), 0x1.1A3CFE870496D08p+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DB0LL, 2), 0x1.1A3CFE870496D8p+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DB8LL, 3), 0x1.1A3CFE870496DCp+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DB6LL, 4), 0x1.1A3CFE870496DBp+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DBFLL, 5), 0x1.1A3CFE870496DF8p+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DC1LL, 6), 0x1.1A3CFE870496E08p+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DC7LL, 7), 0x1.1A3CFE870496E38p+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DC8LL, 8), 0x1.1A3CFE870496E4p+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DCFLL, 9), 0x1.1A3CFE870496E78p+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DD0LL, 0), 0x1.1A3CFE870496E8p+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DD1LL, 11), 0x1.1A3CFE870496E88p+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DD8LL, 12), 0x1.1A3CFE870496ECp+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DDFLL, 13), 0x1.1A3CFE870496EF8p+121L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x023479FD0E092DE0LL, 14), 0x1.1A3CFE870496Fp+121L))
+        return 1;
+
+    if (test__floatuntitf(make_ti(0, 0xFFFFFFFFFFFFFFFFLL), 0x1.FFFFFFFFFFFFFFFEp+63L))
+        return 1;
+
+    if (test__floatuntitf(make_ti(0xFFFFFFFFFFFFFFFFLL, 0x0000000000000000LL),
+                          0x1.FFFFFFFFFFFFFFFEp+127L))
+        return 1;
+    if (test__floatuntitf(make_ti(0xFFFFFFFFFFFFFFFFLL, 0xFFFFFFFFFFFFFFFFLL),
+                          0x1.0000000000000000p+128L))
+        return 1;
+
+    if (test__floatuntitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC2801LL),
+                        0x1.23456789ABCDEF0123456789ABC3p+124L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC3000LL),
+                        0x1.23456789ABCDEF0123456789ABC3p+124L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC37FFLL),
+                        0x1.23456789ABCDEF0123456789ABC3p+124L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC3800LL),
+                        0x1.23456789ABCDEF0123456789ABC4p+124L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC4000LL),
+                        0x1.23456789ABCDEF0123456789ABC4p+124L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC47FFLL),
+                        0x1.23456789ABCDEF0123456789ABC4p+124L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC4800LL),
+                        0x1.23456789ABCDEF0123456789ABC4p+124L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC4801LL),
+                        0x1.23456789ABCDEF0123456789ABC5p+124L))
+        return 1;
+    if (test__floatuntitf(make_ti(0x123456789ABCDEF0LL, 0x123456789ABC57FFLL),
+                        0x1.23456789ABCDEF0123456789ABC5p+124L))
+        return 1;
+#else
+    printf("skipped\n");
+#endif
+   return 0;
+}