README.txt
1 //===-- README.txt - Notes for WebAssembly code gen -----------------------===// 2 3 This WebAssembly backend is presently in a very early stage of development. 4 The code should build and not break anything else, but don't expect a lot more 5 at this point. 6 7 For more information on WebAssembly itself, see the design documents: 8 * https://github.com/WebAssembly/design/blob/master/README.md 9 10 The following documents contain some information on the planned semantics and 11 binary encoding of WebAssembly itself: 12 * https://github.com/WebAssembly/design/blob/master/AstSemantics.md 13 * https://github.com/WebAssembly/design/blob/master/BinaryEncoding.md 14 15 The backend is built, tested and archived on the following waterfall: 16 https://build.chromium.org/p/client.wasm.llvm/console 17 18 The backend's bringup is done using the GCC torture test suite first since it 19 doesn't require C library support. Current known failures are in 20 known_gcc_test_failures.txt, all other tests should pass. The waterfall will 21 turn red if not. Once most of these pass, further testing will use LLVM's own 22 test suite. The tests can be run locally using: 23 github.com/WebAssembly/experimental/blob/master/buildbot/torture_test.py 24 25 Interesting work that remains to be done: 26 * Write a pass to restructurize irreducible control flow. This needs to be done 27 before register allocation to be efficient, because it may duplicate basic 28 blocks and WebAssembly performs register allocation at a whole-function 29 level. Note that LLVM's GPU code has such a pass, but it linearizes control 30 flow (e.g. both sides of branches execute and are masked) which is undesirable 31 for WebAssembly. 32 33 //===---------------------------------------------------------------------===// 34 35 set_local instructions have a return value. We should (a) model this, 36 and (b) write optimizations which take advantage of it. Keep in mind that 37 many set_local instructions are implicit! 38 39 //===---------------------------------------------------------------------===// 40 41 Br, br_if, and tableswitch instructions can support having a value on the 42 expression stack across the jump (sometimes). We should (a) model this, and 43 (b) extend the stackifier to utilize it. 44 45 //===---------------------------------------------------------------------===// 46 47 The min/max operators aren't exactly a<b?a:b because of NaN and negative zero 48 behavior. The ARM target has the same kind of min/max instructions and has 49 implemented optimizations for them; we should do similar optimizations for 50 WebAssembly. 51 52 //===---------------------------------------------------------------------===// 53 54 AArch64 runs SeparateConstOffsetFromGEPPass, followed by EarlyCSE and LICM. 55 Would these be useful to run for WebAssembly too? Also, it has an option to 56 run SimplifyCFG after running the AtomicExpand pass. Would this be useful for 57 us too? 58 59 //===---------------------------------------------------------------------===// 60 61 When is it profitable to set isAsCheapAsAMove on instructions in WebAssembly? 62 63 //===---------------------------------------------------------------------===// 64 65 Register stackification uses the EXPR_STACK physical register to impose 66 ordering dependencies on instructions with stack operands. This is pessimistic; 67 we should consider alternate ways to model stack dependencies. 68 69 //===---------------------------------------------------------------------===// 70 71 Lots of things could be done in WebAssemblyTargetTransformInfo.cpp. Similarly, 72 there are numerous optimization-related hooks that can be overridden in 73 WebAssemblyTargetLowering. 74 75 //===---------------------------------------------------------------------===// 76 77 Instead of the OptimizeReturned pass, which should consider preserving the 78 "returned" attribute through to MachineInstrs and extending the StoreResults 79 pass to do this optimization on calls too. That would also let the 80 WebAssemblyPeephole pass clean up dead defs for such calls, as it does for 81 stores. 82 83 //===---------------------------------------------------------------------===// 84 85 Memset/memcpy/memmove should be marked with the "returned" attribute somehow, 86 even when they are translated through intrinsics. 87 88 //===---------------------------------------------------------------------===// 89
