A FreeBSD toolchain doesn't pass tests

[kkebo]

Summary

On FreeBSD, swift-numerics tests fail. Only IntegerExponentTests.testFloat80 tests fail.

Swift on FreeBSD toolchain is available here.

Environment

• swift-numerics 178f513

$ swift --version
Swift version 6.3-dev (LLVM b58b2a34d509492, Swift cf535d8b998d09b)
Target: x86_64-unknown-freebsd14.3
Build config: +assertions
$ freebsd-version
15.0-RELEASE

Logs

$ swift test
...
Test Case 'IntegerExponentTests.testFloat80' started at 2025-12-14 20:45:35.476
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:48: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:50: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:56: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:58: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:48: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:50: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:56: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:58: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:48: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:50: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:56: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:58: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:48: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:50: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:56: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:58: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:48: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:50: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:56: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:58: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:48: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:50: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:56: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:58: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:48: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:50: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:56: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:58: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:48: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:50: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:56: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:58: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:48: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:50: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:56: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:58: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:48: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:50: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -0.0 but saw 0.0.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:56: error: IntegerExponentTests.testFloat80 : failed -
Sign was wrong: expected -inf but saw inf.
/home/kebo/swift-numerics/Tests/RealTests/IntegerExponentTests.swift:58: error: IntegerExponentTests.testFloat80 : failed -
Test Case 'IntegerExponentTests.testFloat80' failed (0.0 seconds)
Test Suite 'IntegerExponentTests' failed at 2025-12-14 20:45:35.476
         Executed 4 tests, with 40 failures (0 unexpected) in 0.001 (0.001) seconds
...
Test Suite 'debug.xctest' failed at 2025-12-14 20:45:57.053
         Executed 75 tests, with 40 failures (0 unexpected) in 21.767 (21.767) seconds
Test Suite 'All tests' failed at 2025-12-14 20:45:57.053
         Executed 75 tests, with 40 failures (0 unexpected) in 21.767 (21.767) seconds
◇ Test run started.
↳ Testing Library Version: 6.3-dev (98430f2d6cb8d81)
↳ Target Platform: x86_64-unknown-freebsd
✔ Test run with 0 tests in 0 suites passed after 0.001 seconds.

[stephentyrone]

This sounds like a bug in either FreeBSD's powl implementation or the C runtime int64 -> long double conversion.

[kkebo]

@stephentyrone I think it's neither.

a bug in the C runtime int64 -> long double conversion

You can easily verify that the conversion is expected by the following code.

test.c:

#include <stdint.h>
#include <assert.h>
#include <math.h>
#include <stdio.h>

int main() {
    long double y = (long double)(int64_t)100000000000001;
    assert(floorl(y) == y);
    puts("ok");
    return 0;
}

Results:

$ clang -lm test.c
$ ./a.out
ok

a bug in FreeBSD's powl implementation

Indeed, FreeBSD's powl returns a positive ∞ when the base is negative and the exponent is a large odd integer.

test.c:

#include <stdint.h>
#include <assert.h>
#include <math.h>
#include <stdio.h>

int main() {
    long double y = (long double)(int64_t)100000000000001;
    assert(floorl(y) == y);
    printf("%.Le\n", powl(-3.0L, 100000000000001LL));
    printf("%.Le\n", powl(-3.0L, 101LL));
    printf("%.Le\n", powl(-INFINITY, 100000000000001LL));
    printf("%.Le\n", powl(-INFINITY, 101LL));
    return 0;
}

Results:

$ clang -lm test.c
$ ./a.out
inf
-2e+48
-inf
-inf

However, there is no requirement that the sign of the base should be preserved on overflow even if the base is negative and the exponent is a large odd integer in the C standards. So this implementation is legal and not a bug.

In my example (powl(-3.0L, 100000000000001LL)), this will be triggered.

https://github.com/freebsd/freebsd-src/blob/823f158a01330484f83f69c31d695035c607f8ee/lib/msun/ld80/e_powl.c#L457

/* Test the power of 2 for overflow */
if( w > MEXP )
	return (huge * huge);		/* overflow */

[stephentyrone]

C23 §7.12.1 p5:

A floating result overflows if a finite result value with ordinary accuracy would have magnitude
(absolute value) too large for the representation with full precision in the specified type. A result
that is exactly an infinity does not overflow. If a floating result overflows and default rounding
is in effect, then the function returns the value of the macro HUGE_VAL, HUGE_VALF, or HUGE_VALL
according to the return type, with the same sign as the correct value of the function.

C doesn't do a great job in general of precisely specifying the behavior of <math.h> operations, but the sign of overflow results is one of the few things that it actually does specify.

That said, if you'd like to contribute a patch to work around FreeBSD's bug, I'll be happy to take it. I can make some time to help with writing it if needed.

[kkebo]

Thank you for clarifying that. I read that page too and confirmed that it was likely FreeBSD's bug. I couldn't find the similar bug reports on the FreeBSD Bugzilla, so I'll try to report it to the FreeBSD side.

[kkebo]

@stephentyrone I have an idea for a patch. If you don't mind, I'd like to hear your opinion.

1. Split testIntegerExponentCommon() into testIntegerExponentCommon() and testIntegerExponentOverflowAndUnderflow().
2. Skip testIntegerExponentOverflowAndUnderflow() in testFloat80() if the OS is FreeBSD.

diff --git a/Tests/RealTests/IntegerExponentTests.swift b/Tests/RealTests/IntegerExponentTests.swift
index d41113e..51484ca 100644
--- a/Tests/RealTests/IntegerExponentTests.swift
+++ b/Tests/RealTests/IntegerExponentTests.swift
@@ -31,6 +31,11 @@ internal extension Real where Self: FixedWidthFloatingPoint {
     XCTAssertEqual(Self.pow(-1,  Int.max), -1)
     XCTAssertEqual(Self.pow(-1, -Int.max), -1)
     XCTAssertEqual(Self.pow(-1,  Int.min),  1)
+  }
+
+  static func testIntegerExponentOverflowAndUnderflow() {
+    // TODO: replace with seedable generator, print seed.
+    var g = SystemRandomNumberGenerator()
     // Generate some random test values that definitely overflow or
     // underflow; we want to be sure that we get the right ±0 or ±�
     // result.
@@ -82,6 +87,7 @@ internal extension Real where Self: FixedWidthFloatingPoint {
 extension Float16 {
   static func testIntegerExponent() {
     testIntegerExponentCommon()
+    testIntegerExponentOverflowAndUnderflow()
     testIntegerExponentDoubleAndSmaller()
     let u = Float16(1).nextUp
     let d = Float16(1).nextDown
@@ -109,6 +115,7 @@ extension Float16 {
 extension Float {
   static func testIntegerExponent() {
     testIntegerExponentCommon()
+    testIntegerExponentOverflowAndUnderflow()
     testIntegerExponentDoubleAndSmaller()
     let u = Float(1).nextUp
     let d = Float(1).nextDown
@@ -135,6 +142,7 @@ extension Float {
 extension Double {
   static func testIntegerExponent() {
     testIntegerExponentCommon()
+    testIntegerExponentOverflowAndUnderflow()
     // Following tests only make sense (and are only necessary) on 64b platforms.
 #if arch(arm64) || arch(x86_64)
     testIntegerExponentDoubleAndSmaller()
@@ -193,6 +201,10 @@ final class IntegerExponentTests: XCTestCase {
   #if (arch(i386) || arch(x86_64)) && !os(Windows) && !os(Android)
   func testFloat80() {
     Float80.testIntegerExponentCommon()
+    // FreeBSD's powl(3) has a bug where a minus sign is not perserved on overflow/underflow.
+    #if !os(FreeBSD)
+    Float80.testIntegerExponentOverflowAndUnderflow()
+    #endif
   }
   #endif
 }