1 Date: Fri, 1 Jun 2001 17:08:44 -0500 (CDT) 2 From: Chris Lattner <sabre (a] nondot.org> 3 To: Vikram S. Adve <vadve (a] cs.uiuc.edu> 4 Subject: RE: Interesting: GCC passes 5 6 > That is very interesting. I agree that some of these could be done on LLVM 7 > at link-time, but it is the extra time required that concerns me. Link-time 8 > optimization is severely time-constrained. 9 10 If we were to reimplement any of these optimizations, I assume that we 11 could do them a translation unit at a time, just as GCC does now. This 12 would lead to a pipeline like this: 13 14 Static optimizations, xlation unit at a time: 15 .c --GCC--> .llvm --llvmopt--> .llvm 16 17 Link time optimizations: 18 .llvm --llvm-ld--> .llvm --llvm-link-opt--> .llvm 19 20 Of course, many optimizations could be shared between llvmopt and 21 llvm-link-opt, but the wouldn't need to be shared... Thus compile time 22 could be faster, because we are using a "smarter" IR (SSA based). 23 24 > BTW, about SGI, "borrowing" SSA-based optimizations from one compiler and 25 > putting it into another is not necessarily easier than re-doing it. 26 > Optimization code is usually heavily tied in to the specific IR they use. 27 28 Understood. The only reason that I brought this up is because SGI's IR is 29 more similar to LLVM than it is different in many respects (SSA based, 30 relatively low level, etc), and could be easily adapted. Also their 31 optimizations are written in C++ and are actually somewhat 32 structured... of course it would be no walk in the park, but it would be 33 much less time consuming to adapt, say, SSA-PRE than to rewrite it. 34 35 > But your larger point is valid that adding SSA based optimizations is 36 > feasible and should be fun. (Again, link time cost is the issue.) 37 38 Assuming linktime cost wasn't an issue, the question is: 39 Does using GCC's backend buy us anything? 40 41 > It also occurs to me that GCC is probably doing quite a bit of back-end 42 > optimization (step 16 in your list). Do you have a breakdown of that? 43 44 Not really. The irritating part of GCC is that it mixes it all up and 45 doesn't have a clean separation of concerns. A lot of the "back end 46 optimization" happens right along with other data optimizations (ie, CSE 47 of machine specific things). 48 49 As far as REAL back end optimizations go, it looks something like this: 50 51 1. Instruction combination: try to make CISCy instructions, if available 52 2. Register movement: try to get registers in the right places for the 53 architecture to avoid register to register moves. For example, try to get 54 the first argument of a function to naturally land in %o0 for sparc. 55 3. Instruction scheduling: 'nuff said :) 56 4. Register class preferencing: ?? 57 5. Local register allocation 58 6. global register allocation 59 7. Spilling 60 8. Local regalloc 61 9. Jump optimization 62 10. Delay slot scheduling 63 11. Branch shorting for CISC machines 64 12. Instruction selection & peephole optimization 65 13. Debug info output 66 67 But none of this would be usable for LLVM anyways, unless we were using 68 GCC as a static compiler. 69 70 -Chris 71 72
