[llvm] r359426 - [DebugInfo] Terminate more location-list ranges at the end of blocks

[Jeremy Morse via llvm-commits]

Hi,

I'm about to reapply this patch and commit D61940 to fix the
performance regression. I've been doing some due diligence, here are
times for two asan-enabled LLVM build steps with a) trunk @ r362923,
b) trunk with the revert in r360301 reverted, then c) D61940 applied

ninja ./lib/Target/X86/CMakeFiles/LLVMX86CodeGen.dir/X86ISelLowering.cpp.o
a) 81.83s user 0.44s system 99% cpu 1:22.34 total
b) 100.49s user 0.56s system 99% cpu 1:41.07 total
c) 82.92s user 0.46s system 99% cpu 1:23.39 total

ninja ./bin/clang-9 -j 12
a) 24551.45s user 820.57s system 1173% cpu 36:01.63 total
b) 26575.33s user 555.85s system 1182% cpu 38:15.00 total
c) 24258.08s user 538.32s system 1176% cpu 35:07.13 total

X86ISelLowering is one of the largest cpp files in llvm, so I figured
it's a good test for checking the performance regression -- it appears
as an extra 20s compile time, which applying D61940 eliminates.

Building all of clang also has an extra dose of compile time with the
performance regression that D61940 eliminates. Note the anomalous
extra 300s system time in the first reading, I don't think that
damages the overall story though.

--
Thanks,
Jeremy