Performance profiling on Linux AMD64 showed this to be a significant bottleneck. The non-inlined functions are expensive due to XMM registers spilling onto the stack.
* fix integer overflow in mapgen
Some calculations involving the magic seed had overflow because the result of an intermediate arithmetic step could not fit in an s32. By making the magic seed unsigned, the other operand in the equation will be cast to unsigned, and possibly other operands or intermediate operands. This will result in unexpected behavior if an operand is negative, which is technically possible, but logically should not happen.
* comment noise2d bitshift
While working through the code I was momentarily concerned that the right bitshift in noise2d could fill ones in some cases. It turns out that with signed integers, this is indeed true, but this one is shifting an unsigned integer, so the behavior is as expected. I put a comment here to clarify this, in case someone else wonders the same thing down the line.
* noise2d and noise3d unittests
I have added 3 tests each for noise2d and noise3d, testing all zero inputs, a very large seed (case which caused UB in the old implementation) and some fun primes I picked for no particular reason. This should be sufficient to demonstrate that the behavior of the new implementation has not changed. I used uniform initialization because it is a good feature of C++11. Please do not explode.
* uncomment the noise2d bitshift
This reverts commit 583b77ee9f1ad6bb77340ebb5ba51eb9a88ff51c. It's a
well-defined language semantic; it doesn't need to be commented.
* code cleanliness
* MSVC: Fix '/std:c++11' is not a valid compiler option
* MSVC/MINGW: Define 'WIN32_LEAN_AND_MEAN' for the whole project
In some obscure cases 'Windows.h" got includet before that definition, which leaded to compilation warnings+errors
* MSVC: '/arch:SSE' is only available for x86
* MSVC: Fix float conversation
* MSVC/MINGW: use winthreads on Windows
* MSVC: 'USE_CMAKE_CONFIG' might be already definied by CMake build system
* MSVC: Use all available cpu cores for compiling
* Add missing include ctime and use std::time_t
When compiled with optimizations, the most recent versions of clang seem
to 'optimize' out a crucial "and %reg, 0x7FFFFFFF" instruction in noise2d(),
probably because it somehow assumed the variable n would never become greater
than that amount.
Indeed, signed integer underflow is undefined behavior in C and C++, so while
this optimization is "correct" in that sense, it breaks lots of existing code.
Solved by changing n to an unsigned type, making behavior well-defined.
Before, this lua code led to a crash:
local pcg = PcgRandom(42)
local value = pcg:next()
This was because if you called s32 PcgRandom::range(min, max) with the
minimum and maximum possible values for s32 integers (which the lua
binding code did), u32 PcgRandom::range(bound) got called with 0 as the
bound. The bound however is one above the maximum value, so 0 is a "special"
value to pass to this function. This commit fixes the lua crash by
assigning the RNG's full range to the bound 0, which is also fits to the
"maximum is bound - 1" principle, as (u32)-1 is the maximum value in the
u32 range.
Add 'absolute value' option to noise map functions
Extend persistence modulation to 3D noise
Extend 'eased' option to noise2d_perlin* functions
Some noise.cpp formatting fixups
