diff options
author | Richard Yao <[email protected]> | 2023-05-26 13:03:12 -0400 |
---|---|---|
committer | GitHub <[email protected]> | 2023-05-26 10:03:12 -0700 |
commit | 677c6f8457943fe5b56d7aa8807010a104563e4a (patch) | |
tree | 11c2e70b1530c04bf701e7b508cda3545d30f582 /module/Kbuild.in | |
parent | bb736d98d133b4449a4e3bb97a914651677e6713 (diff) |
btree: Implement faster binary search algorithm
This implements a binary search algorithm for B-Trees that reduces
branching to the absolute minimum necessary for a binary search
algorithm. It also enables the compiler to inline the comparator to
ensure that the only slowdown when doing binary search is from waiting
for memory accesses. Additionally, it instructs the compiler to unroll
the loop, which gives an additional 40% improve with Clang and 8%
improvement with GCC.
Consumers must opt into using the faster algorithm. At present, only
B-Trees used inside kernel code have been modified to use the faster
algorithm.
Micro-benchmarks suggest that this can improve binary search performance
by up to 3.5 times when compiling with Clang 16 and up to 1.9 times when
compiling with GCC 12.2.
Reviewed-by: Alexander Motin <[email protected]>
Reviewed-by: Brian Behlendorf <[email protected]>
Signed-off-by: Richard Yao <[email protected]>
Closes #14866
Diffstat (limited to 'module/Kbuild.in')
-rw-r--r-- | module/Kbuild.in | 14 |
1 files changed, 14 insertions, 0 deletions
diff --git a/module/Kbuild.in b/module/Kbuild.in index 8d29f56c2..29a55c977 100644 --- a/module/Kbuild.in +++ b/module/Kbuild.in @@ -34,6 +34,20 @@ ifeq ($(CONFIG_KASAN),y) ZFS_MODULE_CFLAGS += -Wno-error=frame-larger-than= endif +# Generated binary search code is particularly bad with this optimization. +# Oddly, range_tree.c is not affected when unrolling is not done and dsl_scan.c +# is not affected when unrolling is done. +# Disable it until the following upstream issue is resolved: +# https://github.com/llvm/llvm-project/issues/62790 +ifeq ($(CONFIG_X86),y) +ifeq ($(CONFIG_CC_IS_CLANG),y) +CFLAGS_zfs/dsl_scan.o += -mllvm -x86-cmov-converter=false +CFLAGS_zfs/metaslab.o += -mllvm -x86-cmov-converter=false +CFLAGS_zfs/range_tree.o += -mllvm -x86-cmov-converter=false +CFLAGS_zfs/zap_micro.o += -mllvm -x86-cmov-converter=false +endif +endif + ifneq ($(KBUILD_EXTMOD),) @CONFIG_QAT_TRUE@ZFS_MODULE_CFLAGS += -I@QAT_SRC@/include @CONFIG_QAT_TRUE@KBUILD_EXTRA_SYMBOLS += @QAT_SYMBOLS@ |