1 files changed, 70 insertions, 0 deletions
diff --git a/TODO.md b/TODO.md
new file mode 100644
index 00000000..02850276
--- /dev/null
+++ b/TODO.md
@@ -0,0 +1,70 @@
+
+## Roadmap 2.61
+
+Makefile:
+ - -march=native -Ofast -flto=full
+
+afl-fuzz:
+ - sync_fuzzers(): only masters sync from all, slaves only sync from master
+   (@andrea: be careful, often people run all slaves)
+ - ascii_only mode
+
+gcc_plugin:
+ - laf-intel
+ - better instrumentation
+
+qemu_mode:
+ - update to 4.x (probably this will be skipped :( )
+ - instrim for QEMU mode via static analysis (with r2pipe? or angr?)
+   Idea: The static analyzer outputs a map in which each edge that must be
+   skipped is marked with 1. QEMU loads it at startup in the parent process.
+ - rename qemu specific envs to AFL_QEMU (AFL_ENTRYPOINT, AFL_CODE_START/END, AFL_COMPCOV_LEVEL?)
+ - add AFL_QEMU_EXITPOINT (maybe multiple?)
+ - add/implement AFL_QEMU_INST_LIBLIST and AFL_QEMU_NOINST_PROGRAM
+
+custom_mutators:
+ - rip what Superion is doing into custom mutators for js, php, etc.
+ - uniform python and custom mutators API
+
+
+
+## The far away future:
+
+Problem: Average targets (tiff, jpeg, unrar) go through 1500 edges.
+         At afl's default map that means ~16 collisions and ~3 wrappings.
+
+ - Solution #1: increase map size.
+    every +1 decreases fuzzing speed by ~10% and halfs the collisions
+    birthday paradox predicts collisions at this # of edges:
+    
+    | mapsize | collisions |
+    | :-----: | :--------: |
+    | 2^16    | 302        |
+	  | 2^17    | 427        |
+	  | 2^18    | 603        |
+	  | 2^19    | 853        |
+	  | 2^20    | 1207       |
+	  | 2^21    | 1706       |
+	  | 2^22    | 2412       |
+	  | 2^23    | 3411       |
+	  | 2^24    | 4823       |
+
+    Increasing the map is an easy solution but also not a good one.
+
+ - Solution #2: use dynamic map size and collision free basic block IDs
+    This only works in llvm_mode and llvm >= 9 though
+    A potential good future solution. Heiko/hexcoder follows this up
+
+ - Solution #3: write instruction pointers to a big shared map
+    512kb/1MB shared map and the instrumented code writes the instruction
+    pointer into the map. Map must be big enough but could be command line
+    controlled.
+    
+    Good: complete coverage information, nothing is lost. choice of analysis
+          impacts speed, but this can be decided by user options
+    
+    Neutral: a little bit slower but no loss of coverage
+    
+    Bad: completely changes how afl uses the map and the scheduling.
+    Overall another very good solution, Marc Heuse/vanHauser follows this up
+