Use simpler ops for binary instructions (fixes GCC issue)#513
Conversation
Codecov Report
@@ Coverage Diff @@
## master #513 +/- ##
=======================================
Coverage 99.69% 99.69%
=======================================
Files 54 54
Lines 17183 17198 +15
=======================================
+ Hits 17131 17146 +15
Misses 52 52 |
| template <typename T> | ||
| inline constexpr T add(T a, T b) noexcept | ||
| { | ||
| return a + b; |
There was a problem hiding this comment.
I'd prefer we add the asserts for is_integral||is_float to avoid any random overloading.
There was a problem hiding this comment.
This is big overkill and still is not complete (as bool, char, long double would be a match). I can't imagine accidentally adding two things.
There was a problem hiding this comment.
Realistically what could happen is that Value is overloaded and this matches for Value+Value. Why would it be an overkill?
There was a problem hiding this comment.
This function is perfectly valid for any type that is copyable and has + operator. The Value does not support + operator as of #464. Also all usages of it explicitly specify the type.
| } | ||
|
|
||
| template <typename T> | ||
| inline constexpr T rem(T a, T b) noexcept |
So it was a problem only for Do we replace |
The problem seems to be only for add. But I replaced also other functions in the same group (arithmetic binops). |
|
I'm getting significant boost with it (with clang) |
|
Oh, nice. I will check my machine soon. |
This replaces usage of std::plus{} and others with simpler in-house
implementation. The difference is that the std:: versions are passing
arguments by reference. In GCC builds without optimization this may
cause the arguments to be swapped in commutative instructions. This is
at least annoyance for floating-point add which may produce different
NaN results if both inputs are NaNs. This is not a problem for Wasm
conformance, but prevents running exhaustive tests against software
floats implementations.
|
I checked GCC10 and Clang10 and both generate unchanged assembly. |
3 participants
This replaces usage of std::plus{} and others with simpler in-house
implementation. The difference is that the std:: versions are passing
arguments by reference. In GCC builds without optimization this may
cause the arguments to be swapped in commutative instructions. This is
at least annoyance for floating-point add which may produce different
NaN results if both inputs are NaNs. This is not a problem for Wasm
conformance, but prevents running exhaustive tests against software
floats implementations.