mirror of
https://github.com/c64scene-ar/llvm-6502.git
synced 2025-02-08 06:32:24 +00:00
Rewrite the shared implementation of BlockFrequencyInfo and MachineBlockFrequencyInfo entirely. The old implementation had a fundamental flaw: precision losses from nested loops (or very wide branches) compounded past loop exits (and convergence points). The @nested_loops testcase at the end of test/Analysis/BlockFrequencyAnalysis/basic.ll is motivating. This function has three nested loops, with branch weights in the loop headers of 1:4000 (exit:continue). The old analysis gives non-sensical results: Printing analysis 'Block Frequency Analysis' for function 'nested_loops': ---- Block Freqs ---- entry = 1.0 for.cond1.preheader = 1.00103 for.cond4.preheader = 5.5222 for.body6 = 18095.19995 for.inc8 = 4.52264 for.inc11 = 0.00109 for.end13 = 0.0 The new analysis gives correct results: Printing analysis 'Block Frequency Analysis' for function 'nested_loops': block-frequency-info: nested_loops - entry: float = 1.0, int = 8 - for.cond1.preheader: float = 4001.0, int = 32007 - for.cond4.preheader: float = 16008001.0, int = 128064007 - for.body6: float = 64048012001.0, int = 512384096007 - for.inc8: float = 16008001.0, int = 128064007 - for.inc11: float = 4001.0, int = 32007 - for.end13: float = 1.0, int = 8 Most importantly, the frequency leaving each loop matches the frequency entering it. The new algorithm leverages BlockMass and PositiveFloat to maintain precision, separates "probability mass distribution" from "loop scaling", and uses dithering to eliminate probability mass loss. I have unit tests for these types out of tree, but it was decided in the review to make the classes private to BlockFrequencyInfoImpl, and try to shrink them (or remove them entirely) in follow-up commits. The new algorithm should generally have a complexity advantage over the old. The previous algorithm was quadratic in the worst case. The new algorithm is still worst-case quadratic in the presence of irreducible control flow, but it's linear without it. The key difference between the old algorithm and the new is that control flow within a loop is evaluated separately from control flow outside, limiting propagation of precision problems and allowing loop scale to be calculated independently of mass distribution. Loops are visited bottom-up, their loop scales are calculated, and they are replaced by pseudo-nodes. Mass is then distributed through the function, which is now a DAG. Finally, loops are revisited top-down to multiply through the loop scales and the masses distributed to pseudo nodes. There are some remaining flaws. - Irreducible control flow isn't modelled correctly. LoopInfo and MachineLoopInfo ignore irreducible edges, so this algorithm will fail to scale accordingly. There's a note in the class documentation about how to get closer. See also the comments in test/Analysis/BlockFrequencyInfo/irreducible.ll. - Loop scale is limited to 4096 per loop (2^12) to avoid exhausting the 64-bit integer precision used downstream. - The "bias" calculation proposed on llvmdev is *not* incorporated here. This will be added in a follow-up commit, once comments from this review have been handled. git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@206548 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 r4, [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.