Bug 5565 – [64-bit] Wrong Floating Point Results, Related to Mixing With size_t

The following code produces erratic results. I don't know where exactly the culprit is, but I'm hoping I've reduced it enough that someone with a decent mental model of how the compiler works will be able to finish the job. import std.stdio; double foo(int[] data, double q = 0.25) { immutable double N = data.length; // This actually affects the result, making it more severe and obvious // in addition to showing that the value is correct. stderr.writeln(N); // Prints 8. Correct. // lowEnd should be floor(7 * 0.25) == 1. immutable lowEnd = cast(size_t) ((N - 1) * q); // highEnd should be floor(7 * 0.75) - 1 = 4. immutable highEnd = cast(size_t) ((N - 1) * (1.0 - q) - lowEnd); // highFract should be 7 * 0.75 - 4 - 1 = 0.25 immutable highFract = (N - 1) * (1.0 - q) - lowEnd - highEnd; stderr.writeln(lowEnd, '\t', highEnd, '\t', highFract); return highFract; } void main() { writeln(foo([1,2,3,4,5,6,7,8])); } Prints: 8 0 4545916 -4.54591e+06 -4.54591e+06

Forgot to mention: This test case doesn't reproduce the bug if -O is enabled, though the test case I reduced it from did, I think. (I'm not 100% sure that I wasn't just seeing two different bugs simultaneously.)

https://github.com/D-Programming-Language/dmd/commit/95b70d83ac51496eb171c1b408b5781eeee1a29d https://github.com/D-Programming-Language/dmd/commit/fe147641a0334bca8d5ac3def7a0aba8529a0a66

I think this was actually two bugs that I thought had a single root cause. You only got one of them. Now, the result from running this code with -m64 and without -O on linux is: 8 1 4545912 -4.54591e+06 -4.54591e+06 On the line with the three columns, the first one is now correct. The other two are still wrong, unless -O is enabled.

https://github.com/D-Programming-Language/dmd/commit/2421fd325b8e7528f5bafd2c4a4d6fd5008f3ab1 https://github.com/D-Programming-Language/dmd/commit/c51de3ab026f811744a4dc81b6f50c2525fc07a5