[PATCH] [LoopUnrollRuntime] Avoid high-cost trip count computation.

[Justin Bogner]

Sanjoy Das <sanjoy at playingwithpointers.com> writes:
> Hi atrick,
>
> Runtime unrolling of loops needs to emit an expression to compute the
> loop's runtime trip-count.  Avoid runtime unrolling if this computation
> will be expensive.

This looks pretty obvious to me (especially since you told me it fixes
the performance regression I mentioned from r233881).

>
> Depends on D8993.
>
> http://reviews.llvm.org/D8994
>
> Files:
>   lib/Transforms/Utils/LoopUnrollRuntime.cpp
>   test/Transforms/LoopUnroll/high-cost-trip-count-computation.ll
>   test/Transforms/LoopUnroll/runtime-loop.ll
>   test/Transforms/LoopUnroll/runtime-loop4.ll
>
> Index: lib/Transforms/Utils/LoopUnrollRuntime.cpp
> ===================================================================
> --- lib/Transforms/Utils/LoopUnrollRuntime.cpp
> +++ lib/Transforms/Utils/LoopUnrollRuntime.cpp
> @@ -312,6 +312,12 @@
>    if (isa<SCEVCouldNotCompute>(TripCountSC))
>      return false;
>  
> +  BasicBlock *Header = L->getHeader();
> +  const DataLayout &DL = Header->getModule()->getDataLayout();
> +  SCEVExpander Expander(*SE, DL, "loop-unroll");
> +  if (Expander.isHighCostExpansion(TripCountSC, L))
> +    return false;
> +
>    // We only handle cases when the unroll factor is a power of 2.
>    // Count is the loop unroll factor, the number of extra copies added + 1.
>    if (!isPowerOf2_32(Count))
> @@ -332,18 +338,15 @@
>    auto *DT = DTWP ? &DTWP->getDomTree() : nullptr;
>  
>    BasicBlock *PH = L->getLoopPreheader();
> -  BasicBlock *Header = L->getHeader();
>    BasicBlock *Latch = L->getLoopLatch();
>    // It helps to splits the original preheader twice, one for the end of the
>    // prolog code and one for a new loop preheader
>    BasicBlock *PEnd = SplitEdge(PH, Header, DT, LI);
>    BasicBlock *NewPH = SplitBlock(PEnd, PEnd->getTerminator(), DT, LI);
>    BranchInst *PreHeaderBR = cast<BranchInst>(PH->getTerminator());
> -  const DataLayout &DL = Header->getModule()->getDataLayout();
>  
>    // Compute the number of extra iterations required, which is:
>    //  extra iterations = run-time trip count % (loop unroll factor + 1)
> -  SCEVExpander Expander(*SE, DL, "loop-unroll");
>    Value *TripCount = Expander.expandCodeFor(TripCountSC, TripCountSC->getType(),
>                                              PreHeaderBR);
>    Value *BECount = Expander.expandCodeFor(BECountSC, BECountSC->getType(),
> Index: test/Transforms/LoopUnroll/high-cost-trip-count-computation.ll
> ===================================================================
> --- /dev/null
> +++ test/Transforms/LoopUnroll/high-cost-trip-count-computation.ll
> @@ -0,0 +1,27 @@
> +; RUN: opt -S -unroll-runtime -loop-unroll < %s | FileCheck %s
> +
> +target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
> +
> +;; Check that we don't emit expensive instructions to compute trip
> +;; counts when unrolling loops.
> +
> +define i32 @test(i64 %v12, i8* %array, i64* %loc) {
> +; CHECK-LABEL: @test(
> +; CHECK-NOT: udiv
> +entry:
> +  %step = load i64, i64* %loc, !range !0
> +  br label %loop
> +
> +loop:                                           ; preds = %entry, %loop
> +  %k.015 = phi i64 [ %v15, %loop ], [ %v12, %entry ]
> +  %v14 = getelementptr inbounds i8, i8* %array, i64 %k.015
> +  store i8 0, i8* %v14
> +  %v15 = add nuw nsw i64 %k.015, %step
> +  %v16 = icmp slt i64 %v15, 8193
> +  br i1 %v16, label %loop, label %loopexit
> +
> +loopexit:                             ; preds = %loop
> +  ret i32 0
> +}
> +
> +!0 = !{i64 1, i64 100}
> Index: test/Transforms/LoopUnroll/runtime-loop.ll
> ===================================================================
> --- test/Transforms/LoopUnroll/runtime-loop.ll
> +++ test/Transforms/LoopUnroll/runtime-loop.ll
> @@ -1,5 +1,7 @@
>  ; RUN: opt < %s -S -loop-unroll -unroll-runtime=true | FileCheck %s
>  
> +target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64-S128"
> +
>  ; Tests for unrolling loops with run-time trip counts
>  
>  ; CHECK: %xtraiter = and i32 %n
> Index: test/Transforms/LoopUnroll/runtime-loop4.ll
> ===================================================================
> --- test/Transforms/LoopUnroll/runtime-loop4.ll
> +++ test/Transforms/LoopUnroll/runtime-loop4.ll
> @@ -20,7 +20,8 @@
>    br label %loop2.header
>  
>  loop2.header:
> -  br label %loop2
> +  %e = icmp uge i32 %iter, 1
> +  br i1 %e, label %loop2, label %exit2
>  
>  loop2:
>    %iv2 = phi i32 [ 0, %loop2.header ], [ %inc2, %loop2 ]
>
> EMAIL PREFERENCES
>   http://reviews.llvm.org/settings/panel/emailpreferences/
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at cs.uiuc.edu
> http://lists.cs.uiuc.edu/mailman/listinfo/llvm-commits