After writing some benchmarks for ebml::reader::vuint_at() I noticed that LLVM doesn't seem to inline the from_be32 function even though it only does a call to the bswap32 intrinsic in the x86_64 case. Marking the functions with #[inline(always)] fixes that and seems to me a reasonable thing to do. I got the following measurements in my vuint_at() benchmarks:
- Before
test ebml::bench::vuint_at_A_aligned ... bench: 1075 ns/iter (+/- 58)
test ebml::bench::vuint_at_A_unaligned ... bench: 1073 ns/iter (+/- 5)
test ebml::bench::vuint_at_D_aligned ... bench: 1150 ns/iter (+/- 5)
test ebml::bench::vuint_at_D_unaligned ... bench: 1151 ns/iter (+/- 6)
- Inline from_be32
test ebml::bench::vuint_at_A_aligned ... bench: 769 ns/iter (+/- 9)
test ebml::bench::vuint_at_A_unaligned ... bench: 795 ns/iter (+/- 6)
test ebml::bench::vuint_at_D_aligned ... bench: 758 ns/iter (+/- 8)
test ebml::bench::vuint_at_D_unaligned ... bench: 759 ns/iter (+/- 8)
- Using vuint_at_slow()
test ebml::bench::vuint_at_A_aligned ... bench: 646 ns/iter (+/- 7)
test ebml::bench::vuint_at_A_unaligned ... bench: 645 ns/iter (+/- 3)
test ebml::bench::vuint_at_D_aligned ... bench: 907 ns/iter (+/- 4)
test ebml::bench::vuint_at_D_unaligned ... bench: 1085 ns/iter (+/- 16)
As expected inlining from_be32() gave a considerable speedup.
I also tried how the "slow" version fared against the optimized version and noticed that it's
actually a bit faster for small A class integers (using only two bytes) but slower for big D class integers (using four bytes)
If there is a lot of data in thread-local storage some implementations
of pthreads (e.g. glibc) fail if you don't request a stack large enough
-- by adjusting for the minimum size we guarantee that our stacks are
always large enough. Issue #6233.
In a multi-host build the mklldeps.py tool is getting called before
all the llvm-configs are built. I am not actually sure the cause. I
had convinced myself that DEF_LLVM_RULES needed to be called before
the llvmdeps.rs rule, but now looking at it again I can't see why.
