From ab316b91d76c7d48613b001e7a35e166b4fce129 Mon Sep 17 00:00:00 2001 From: Arseny Kapoulkine Date: Mon, 31 Jul 2023 11:33:03 -0700 Subject: [PATCH] Update behavior-math-clamp-nan.md --- rfcs/behavior-math-clamp-nan.md | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/rfcs/behavior-math-clamp-nan.md b/rfcs/behavior-math-clamp-nan.md index 566cf950..6e15a9d3 100644 --- a/rfcs/behavior-math-clamp-nan.md +++ b/rfcs/behavior-math-clamp-nan.md @@ -29,7 +29,7 @@ This can be a problem if: - a, if t is NaN As a consequence of this change, `math.clamp` return value `r` will always satisfy `a <= r <= b`. -Note that if the input or one of the limits is a negative zero we will not guarantee that the sign of the output is preserved; positive and negative zeros are considered equal under this proposal. +Note that if the input or one of the limits is a negative zero we will not guarantee that the sign of the output is preserved when it's equal to zero; positive and negative zeros are considered equal under this proposal. The behavior of NaN clamping outside of Luau varies. Shading languages (HLSL in Direct3D, MSL in Metal) typically guarantee that `clamp(t, a, b)` is equal to a when t is NaN, although OpenGL and Vulkan leave result unspecified; C++ `std::clamp` provides no guarantees and in fact treats NaN as a technically-invalid input; C#, Rust and Julia preserve NaN; Zig returns the upper limit. Ruby raises an error when input is NaN. Swift does not implement clamping for floats. Despite the differences, we believe that for Luau domain and given that Luau only has a single number type, guaranteeing that the result is in the specified range is more valuable than preserving NaN, and returning the lower limit is more consistent with industry practice than returning upper limit.