Add benchmarks results into README.md

README.md

@@ -256,7 +256,7 @@ The textual representation of the IR after system independent optimizations:
 ```
 The visualized graph:
 
-![IR example](examples/mandelbrot.svg)
+![IR example](https://dstogov.github.io/mandelbrot.svg)
 
 The graph was generated by the commands:
 
@@ -342,6 +342,48 @@ clang -O -Xclang -disable-llvm-passes -c -emit-llvm -o tmp.bc ./dynasm/minilua.c
 opt tmp.bc --passes='function(mem2reg)' -S -o minilua.ll
 ```
 
+## Performace
+
+The fllowing table shows the benchmarks execution time in comparison to the same benchmarks compiled by ``gcc -O2`` (the more the better). The C benchmark were compiled by CLAG into LLVM code (without any LLVM optimizations, only SSA construction is necessary now) and then loaded, compiled and executed by IR framework.
+
+| Benhcmark       | Execution time (relative to GCC -02) | 
+| --------------- | ------------------------------------ |
+| array           |                                 0.86 |
+| binary-trees    |                                 0.97 |
+| funnkuch-reduce |                                 1.03 |
+| hash            |                                 1.13 |
+| hash2           |                                 0.88 |
+| heapsort        |                                 1.02 |
+| lists           |                                 0.95 |
+| matrix          |                                 1.10 |
+| method-call     |                                 0.92 |
+| mandelbrot      |                                 0.95 |
+| nbody           |                                 0.77 |
+| sieve           |                                 0.92 |
+| spectral-norm   |                                 0.94 |
+| strcat          |                                 0.96 |
+| oggenc          |                                 0.86 |
+| minilua         |                                 0.96 |
+| AVERAGE         |                                 0.95 |
+| GEOMEAN         |                                 0.95 |
+
+Most of the bechmarks a very simple (few screems of code), oggenc and minilua are real applications. As you can see, IR produces code that in average 5% slower than ``GCC -O2``, in worst case the code was ~25% slower, and on some benchmarks it even faster. The folloing chart shows the same data graphically.
+
+![IR benchmark chart](https://dstogov.github.io/ir_bench.svg)
+
+
+The following table shows the comilation time relative to ``gcc -O2`` (the more the better)
+
+| Benhcmark       | Compilation time (relative to GCC -02) | 
+| --------------- | -------------------------------------- |
+| oggenc          |                                  26.42 |
+| minilua         |                                  18.00 |
+| AVERAGE         |                                  22.21 |
+
+This comparison is not completely fair, because GCC compiles C source, but IR takes precompiled LLVM asm
+
+Anyway, **IR framework provides code that is in average 5% slower, but does this ~20 times faster**.
+
 ## PHP JIT based on IR
 
 A new experimental JIT for PHP based on this project is developed at [master](https://github.com/php/php-src/tree/master/ext/opcache/jit) php-src branch.