OSDN Git Service

[MachineCombiner] Update instruction depths incrementally for large BBs.
authorFlorian Hahn <florian.hahn@arm.com>
Thu, 7 Sep 2017 12:49:39 +0000 (12:49 +0000)
committerFlorian Hahn <florian.hahn@arm.com>
Thu, 7 Sep 2017 12:49:39 +0000 (12:49 +0000)
commit651af0243784137d375a40c97325ffb383845493
treece4c0d2234ca78aabcf5c72743389f6734f7a7f0
parent4970c50abeb4a18ffcc65088870f739e2216856d
[MachineCombiner] Update instruction depths incrementally for large BBs.

Summary:
For large basic blocks with lots of combinable instructions, the
MachineTraceMetrics computations in MachineCombiner can dominate the compile
time, as computing the trace information is quadratic in the number of
instructions in a BB and it's relevant successors/predecessors.

In most cases, knowing the instruction depth should be enough to make
combination decisions. As we already iterate over all instructions in a basic
block, the instruction depth can be computed incrementally. This reduces the
cost of machine-combine drastically in cases where lots of instructions
are combined. The major drawback is that AFAIK, computing the critical path
length cannot be done incrementally. Therefore we only compute
instruction depths incrementally, for basic blocks with more
instructions than inc_threshold. The -machine-combiner-inc-threshold
option can be used to set the threshold and allows for easier
experimenting and checking if using incremental updates for all basic
blocks has any impact on the performance.

Reviewers: sanjoy, Gerolf, MatzeB, efriedma, fhahn

Reviewed By: fhahn

Subscribers: kiranchandramohan, javed.absar, efriedma, llvm-commits

Differential Revision: https://reviews.llvm.org/D36619

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@312719 91177308-0d34-0410-b5e6-96231b3b80d8
include/llvm/CodeGen/MachineTraceMetrics.h
lib/CodeGen/MachineCombiner.cpp
lib/CodeGen/MachineTraceMetrics.cpp
test/CodeGen/AArch64/machine-combiner.ll
test/CodeGen/X86/machine-combiner.ll
test/CodeGen/X86/mul-constant-result.ll