Optimize fabs, fneg, fcopysign #560

Yellow-King21 · 2020-09-29T20:21:28Z

Closes #540.

I created optimized first single case of issue #540 to be sure that i understood a problem.

lib/fizzy/float_handling.hpp

lib/fizzy/execute.cpp

axic · 2020-10-02T12:47:01Z

Also please squash the commits into a single one.

lib/fizzy/execute.cpp

codecov · 2020-10-03T13:38:57Z

Codecov Report

Merging #560 into master will increase coverage by 0.00%.
The diff coverage is 100.00%.

@@           Coverage Diff           @@
##           master     #560   +/-   ##
=======================================
  Coverage   98.23%   98.24%           
=======================================
  Files          62       62           
  Lines        9023     9039   +16     
=======================================
+ Hits         8864     8880   +16     
  Misses        159      159

lib/fizzy/execute.cpp

chfast · 2020-10-07T08:26:19Z

lib/fizzy/execute.cpp

@@ -356,6 +362,62 @@ T fnearest(T value) noexcept
        return t;
 }

+
+template <typename T>
+T fabs(T) noexcept = delete;


@gumb0, @axic I'm currently in opinion that abs() is better name as we don't have iadd() and fadd(), just add().

I was actually going to suggest to keep the f prefix and also use fneg, given these functions are specifically working on floating points. Since there's no abs on integers in Wasm, it doesn't come up as an issue. Should there be one in the future, happy to rename it to abs.

Same for fcopysign then?

https://en.cppreference.com/w/cpp/numeric/math/copysign seems to be floating point only, so fine to keep it as copysign, but okay either way.

Keeping fcopysign then.

lib/fizzy/execute.cpp

axic · 2020-10-08T17:16:29Z

lib/fizzy/execute.cpp

@@ -19,6 +20,11 @@ namespace
 // code_offset + imm_offset + stack_height
 constexpr auto BranchImmediateSize = 3 * sizeof(uint32_t);

+constexpr uint32_t F32AbsMask = 0x7fffffff;


Could go crazy and use std::numeric_limits<uint32_t>::max() >> 1, but the constant is nicer.

Use unsigned integers and bit manipulations to implement three floating-point operators: fabs, fneg, and fcopysign. This is slightly better than using compiler's builtins because compiler is able to inline them earlier and avoid using SSE registers required by calling convention. Co-authored-by: Paweł Bylica <[email protected]>

chfast requested changes Sep 30, 2020

View reviewed changes

lib/fizzy/float_handling.hpp Outdated Show resolved Hide resolved

lib/fizzy/execute.cpp Outdated Show resolved Hide resolved

chfast requested changes Oct 1, 2020

View reviewed changes