Module esp32_hal::mem [−][src]
Expand description
Custom versions of routines used by LLVM (like memcpy, memset, etc.)
These are normally part of the compiler-builtins crate. However the default routines do not use word sized aligned instructions, which is slow and moreover leads to crashes when using memories/processors which only allows aligned accesses.
Implementation is optimized for large blocks of data. Assumption is that for small data, they are inlined by the compiler. Some optimization done for often used small sizes as otherwise significant slowdown in debug mode.
Implementation is optimized when dst/s1 and src/s2 have the same alignment. If alignment of s1 and s2 is unequal, then either s1 or s2 accesses are not aligned resulting in slower performance. (If s1 or s2 is aligned, then those accesses are aligned.)
Further optimization is possible by having a dedicated code path for unaligned accesses, which uses 2*PTR_SIZE to PTR_SIZE shift operation (i.e. llvm.fshr); but implementation of this intrinsic is not yet optimized and currently leads to worst results.
Also loop unrolling in the memcpy_reverse function is not fully optimal due to limited current llvm optimization: uses add with negative offset + store, instead of store with positive offset; so 3 instructions per loop instead of 2
A further future optimization possibility is using zero overhead loop, but again currently not yet supported by llvm for xtensa.
For large aligned memset and memcpy reaches ~88% of maximum memory bandwidth; for memcpy_reverse ~60%.
Functions
Compare n-bytes of data from s1 and s2 and returns 0 for s1==s2 and !=0 otherwise
Compare n-bytes of data from s1 and s2 and returns <0 for s1<s2, 0 for s1=s2 and >0 for s1>s2
Copies n-bytes of data from src to dst
Copies n-bytes of data from src to dst
Copies n-bytes of data from src to dst and properly handles overlapping data
Fills n-bytes with byte sized value