mirror of
https://github.com/RPCSX/llvm.git
synced 2025-02-27 16:38:30 +00:00

undefined result. This adds new ISD nodes for the new semantics, selecting them when the LLVM intrinsic indicates that the undef behavior is desired. The new nodes expand trivially to the old nodes, so targets don't actually need to do anything to support these new nodes besides indicating that they should be expanded. I've done this for all the operand types that I could figure out for all the targets. Owners of various targets, please review and let me know if any of these are incorrect. Note that the expand behavior is *conservatively correct*, and exactly matches LLVM's current behavior with these operations. Ideally this patch will not change behavior in any way. For example the regtest suite finds the exact same instruction sequences coming out of the code generator. That's why there are no new tests here -- all of this is being exercised by the existing test suite. Thanks to Duncan Sands for reviewing the various bits of this patch and helping me get the wrinkles ironed out with expanding for each target. Also thanks to Chris for clarifying through all the discussions that this is indeed the approach he was looking for. That said, there are likely still rough spots. Further review much appreciated. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@146466 91177308-0d34-0410-b5e6-96231b3b80d8
//===---------------------------------------------------------------------===// Common register allocation / spilling problem: mul lr, r4, lr str lr, [sp, #+52] ldr lr, [r1, #+32] sxth r3, r3 ldr r4, [sp, #+52] mla r4, r3, lr, r4 can be: mul lr, r4, lr mov r4, lr str lr, [sp, #+52] ldr lr, [r1, #+32] sxth r3, r3 mla r4, r3, lr, r4 and then "merge" mul and mov: mul r4, r4, lr str lr, [sp, #+52] ldr lr, [r1, #+32] sxth r3, r3 mla r4, r3, lr, r4 It also increase the likelihood the store may become dead. //===---------------------------------------------------------------------===// bb27 ... ... %reg1037 = ADDri %reg1039, 1 %reg1038 = ADDrs %reg1032, %reg1039, %NOREG, 10 Successors according to CFG: 0x8b03bf0 (#5) bb76 (0x8b03bf0, LLVM BB @0x8b032d0, ID#5): Predecessors according to CFG: 0x8b0c5f0 (#3) 0x8b0a7c0 (#4) %reg1039 = PHI %reg1070, mbb<bb76.outer,0x8b0c5f0>, %reg1037, mbb<bb27,0x8b0a7c0> Note ADDri is not a two-address instruction. However, its result %reg1037 is an operand of the PHI node in bb76 and its operand %reg1039 is the result of the PHI node. We should treat it as a two-address code and make sure the ADDri is scheduled after any node that reads %reg1039. //===---------------------------------------------------------------------===// Use local info (i.e. register scavenger) to assign it a free register to allow reuse: ldr r3, [sp, #+4] add r3, r3, #3 ldr r2, [sp, #+8] add r2, r2, #2 ldr r1, [sp, #+4] <== add r1, r1, #1 ldr r0, [sp, #+4] add r0, r0, #2 //===---------------------------------------------------------------------===// LLVM aggressively lift CSE out of loop. Sometimes this can be negative side- effects: R1 = X + 4 R2 = X + 7 R3 = X + 15 loop: load [i + R1] ... load [i + R2] ... load [i + R3] Suppose there is high register pressure, R1, R2, R3, can be spilled. We need to implement proper re-materialization to handle this: R1 = X + 4 R2 = X + 7 R3 = X + 15 loop: R1 = X + 4 @ re-materialized load [i + R1] ... R2 = X + 7 @ re-materialized load [i + R2] ... R3 = X + 15 @ re-materialized load [i + R3] Furthermore, with re-association, we can enable sharing: R1 = X + 4 R2 = X + 7 R3 = X + 15 loop: T = i + X load [T + 4] ... load [T + 7] ... load [T + 15] //===---------------------------------------------------------------------===// It's not always a good idea to choose rematerialization over spilling. If all the load / store instructions would be folded then spilling is cheaper because it won't require new live intervals / registers. See 2003-05-31-LongShifts for an example. //===---------------------------------------------------------------------===// With a copying garbage collector, derived pointers must not be retained across collector safe points; the collector could move the objects and invalidate the derived pointer. This is bad enough in the first place, but safe points can crop up unpredictably. Consider: %array = load { i32, [0 x %obj] }** %array_addr %nth_el = getelementptr { i32, [0 x %obj] }* %array, i32 0, i32 %n %old = load %obj** %nth_el %z = div i64 %x, %y store %obj* %new, %obj** %nth_el If the i64 division is lowered to a libcall, then a safe point will (must) appear for the call site. If a collection occurs, %array and %nth_el no longer point into the correct object. The fix for this is to copy address calculations so that dependent pointers are never live across safe point boundaries. But the loads cannot be copied like this if there was an intervening store, so may be hard to get right. Only a concurrent mutator can trigger a collection at the libcall safe point. So single-threaded programs do not have this requirement, even with a copying collector. Still, LLVM optimizations would probably undo a front-end's careful work. //===---------------------------------------------------------------------===// The ocaml frametable structure supports liveness information. It would be good to support it. //===---------------------------------------------------------------------===// The FIXME in ComputeCommonTailLength in BranchFolding.cpp needs to be revisited. The check is there to work around a misuse of directives in inline assembly. //===---------------------------------------------------------------------===// It would be good to detect collector/target compatibility instead of silently doing the wrong thing. //===---------------------------------------------------------------------===// It would be really nice to be able to write patterns in .td files for copies, which would eliminate a bunch of explicit predicates on them (e.g. no side effects). Once this is in place, it would be even better to have tblgen synthesize the various copy insertion/inspection methods in TargetInstrInfo. //===---------------------------------------------------------------------===// Stack coloring improvements: 1. Do proper LiveStackAnalysis on all stack objects including those which are not spill slots. 2. Reorder objects to fill in gaps between objects. e.g. 4, 1, <gap>, 4, 1, 1, 1, <gap>, 4 => 4, 1, 1, 1, 1, 4, 4 //===---------------------------------------------------------------------===// The scheduler should be able to sort nearby instructions by their address. For example, in an expanded memset sequence it's not uncommon to see code like this: movl $0, 4(%rdi) movl $0, 8(%rdi) movl $0, 12(%rdi) movl $0, 0(%rdi) Each of the stores is independent, and the scheduler is currently making an arbitrary decision about the order. //===---------------------------------------------------------------------===// Another opportunitiy in this code is that the $0 could be moved to a register: movl $0, 4(%rdi) movl $0, 8(%rdi) movl $0, 12(%rdi) movl $0, 0(%rdi) This would save substantial code size, especially for longer sequences like this. It would be easy to have a rule telling isel to avoid matching MOV32mi if the immediate has more than some fixed number of uses. It's more involved to teach the register allocator how to do late folding to recover from excessive register pressure.