This pass is a draft prototype!
For the moment, it provides an invariance degrees for instructions, inner branches and loops (we call them chunks). It can also peel (disabled by default because it's not stable yet) the chunks regarding to the degrees computed previously.
You need LLVM with
clang, I use the release_40 branch.
Build:
$ cd lqicm_pass
$ edit build.sh → give the right LLVM-DIR path to cmake
$ ./build.sh
To use initialize…Pass(…) registration I added
initializeLegacyLQICMPassPass(PassRegistry*); in
include/llvm/InitializePasses.h it should exist a better way…
Run:
$ edit run.sh → give the right LLVM_BUILD path
$ ./run.sh something.(c|ll)
Example:
$ ./run.sh test/files/Fact.c
↑ This is a basic same factorial computed n times. This pass detects
the invariance of the entire loop and can peel it (see also Fact4.c
which has the same factorial but of invariance degree equal to 4,
means that the loop needs to be peeled four times before hoisting the
factorial loop).
The pass is enabled by default when the libLQICMPass.so is loaded to
opt or directly into clang (using -load option).
Registered as EP_ModuleOptimizerEarly it can provides statistics
(with -mllvm -stats flags) on quasi-invariants detected before loop
optimizations.
By adding the lqicm analysis pass before each iteration of licm we
can compare the number of invariants detected.
We decided to compare them with the -Oz optimization level to avoid
loop rotations (lqicm can't analyze this form of loop for the
moment). Here we compile vim with the flags -fno-vectorize -fno-slp-vectorize -Oz:
First, it's important to know that, at -Oz, licm is performed 3
times and hoists ~1500 instructions less than at -O3.
| 7m45,496s | user time |
|---|---|
| 3808 | number of loops at EP_ModuleOptimizerEarly |
| 13178 | total number of times lqicm LoopPass was launched |
| 8062 | number of all Instructions hoisted by licm |
| 160 | number of instructions sunk by licm |
| 6525 | total number of times lqicm has analyzed the loop (over 13178) |
| 7632 | number of all quasi-invariants detected |
| 808 | number of quasi-invariants Chunk detected |
| 6824 | number of quasi-invariants Instructions |
| 43 | number of chunks with invariance degree >= 2 |
Note that there are some restrictions on the form of the loop: some of the loops with several exit blocks, latches, rotated or without header are not analyzed. In the previous statistics it's ~50% of the time, this is why results are biased.
Over only half of the loop analyzes licm performs, we find 6824
quasi-invariants instructions (~85% of the total invariants currently
hoisted by licm). We can suppose that over all loop analyzes we
should find more instructions than licm. We are currently working
for the flexibility of our analysis.