[llvm] r303084 - CodeGen: BlockPlacement: Increase tail duplication size for O3.

Vitaly Buka via llvm-commits llvm-commits at lists.llvm.org
Mon May 15 19:20:22 PDT 2017


Thanks, I've updated the test in r303136

On Mon, May 15, 2017 at 7:14 PM Dmitry Vyukov <dvyukov at google.com> wrote:

> I've benchmarks this change on the tsan benchmark:
> projects/compiler-rt/lib/tsan/tests/rtl/TsanRtlTest
> --gtest_also_run_disabled_tests --gtest_filter=DISABLED_BENCH.Mop*
> (some numbers below)
>
> So let's just update check_analyze for the new reality.
>
>
> $ egrep "OK.*Mop1Read" /tmp/old
> [       OK ] DISABLED_BENCH.Mop1Read (5317 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5907 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5356 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5376 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5339 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5226 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5261 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5464 ms)
> $ egrep "OK.*Mop1Read" /tmp/new
> [       OK ] DISABLED_BENCH.Mop1Read (5324 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5238 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5344 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5327 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5373 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5208 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5332 ms)
> [       OK ] DISABLED_BENCH.Mop1Read (5351 ms)
>
>
> $ egrep "OK.*Mop8Write" /tmp/old
> [       OK ] DISABLED_BENCH.Mop8Write (1624 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1623 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1575 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1625 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1619 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1585 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1586 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1627 ms)
> $ egrep "OK.*Mop8Write" /tmp/new
> [       OK ] DISABLED_BENCH.Mop8Write (1760 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1634 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1630 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1632 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1601 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1599 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1589 ms)
> [       OK ] DISABLED_BENCH.Mop8Write (1625 ms)
>
>
>
> On Mon, May 15, 2017 at 5:10 PM, Vitaly Buka <vitalybuka at google.com>
> wrote:
> > +Dmitry Vyukov
> >
> > On Mon, May 15, 2017 at 5:09 PM Vitaly Buka <vitalybuka at google.com>
> wrote:
> >>
> >> This test is broken after the patch:
> >>
> http://lab.llvm.org:8011/builders/sanitizer-x86_64-linux-autoconf/builds/8452/steps/tsan%20analyze/logs/stdio
> >>
> >> On Mon, May 15, 2017 at 10:44 AM Kyle Butt via llvm-commits
> >> <llvm-commits at lists.llvm.org> wrote:
> >>>
> >>> Author: iteratee
> >>> Date: Mon May 15 12:30:47 2017
> >>> New Revision: 303084
> >>>
> >>> URL: http://llvm.org/viewvc/llvm-project?rev=303084&view=rev
> >>> Log:
> >>> CodeGen: BlockPlacement: Increase tail duplication size for O3.
> >>>
> >>> At O3 we are more willing to increase size if we believe it will
> improve
> >>> performance. The current threshold for tail-duplication of 2
> instructions
> >>> is
> >>> conservative, and can be relaxed at O3.
> >>>
> >>> Benchmark results:
> >>> llvm test-suite:
> >>> 6% improvement in aha, due to duplication of loop latch
> >>> 3% improvement in hexxagon
> >>>
> >>> 2% slowdown in lpbench. Seems related, but couldn't completely
> diagnose.
> >>>
> >>> Internal google benchmark:
> >>> Produces 4% improvement on internal google protocol buffer
> serialization
> >>> benchmarks.
> >>>
> >>> Differential-Revision: https://reviews.llvm.org/D32324
> >>>
> >>> Modified:
> >>>     llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp
> >>>     llvm/trunk/test/CodeGen/PowerPC/tail-dup-layout.ll
> >>>     llvm/trunk/test/CodeGen/X86/sse1.ll
> >>>
> >>> Modified: llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp
> >>> URL:
> >>>
> http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp?rev=303084&r1=303083&r2=303084&view=diff
> >>>
> >>>
> ==============================================================================
> >>> --- llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp (original)
> >>> +++ llvm/trunk/lib/CodeGen/MachineBlockPlacement.cpp Mon May 15
> 12:30:47
> >>> 2017
> >>> @@ -133,6 +133,14 @@ static cl::opt<unsigned> TailDupPlacemen
> >>>               "that won't conflict."), cl::init(2),
> >>>      cl::Hidden);
> >>>
> >>> +// Heuristic for aggressive tail duplication.
> >>> +static cl::opt<unsigned> TailDupPlacementAggressiveThreshold(
> >>> +    "tail-dup-placement-aggressive-threshold",
> >>> +    cl::desc("Instruction cutoff for aggressive tail duplication
> during
> >>> "
> >>> +             "layout. Used at -O3. Tail merging during layout is
> forced
> >>> to "
> >>> +             "have a threshold that won't conflict."), cl::init(3),
> >>> +    cl::Hidden);
> >>> +
> >>>  // Heuristic for tail duplication.
> >>>  static cl::opt<unsigned> TailDupPlacementPenalty(
> >>>      "tail-dup-placement-penalty",
> >>> @@ -2646,9 +2654,26 @@ bool MachineBlockPlacement::runOnMachine
> >>>    assert(BlockToChain.empty());
> >>>    assert(ComputedEdges.empty());
> >>>
> >>> +  unsigned TailDupSize = TailDupPlacementThreshold;
> >>> +  // If only the aggressive threshold is explicitly set, use it.
> >>> +  if (TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0 &&
> >>> +      TailDupPlacementThreshold.getNumOccurrences() == 0)
> >>> +    TailDupSize = TailDupPlacementAggressiveThreshold;
> >>> +
> >>> +  TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
> >>> +  // For agressive optimization, we can adjust some thresholds to be
> >>> less
> >>> +  // conservative.
> >>> +  if (PassConfig->getOptLevel() >= CodeGenOpt::Aggressive) {
> >>> +    // At O3 we should be more willing to copy blocks for tail
> >>> duplication. This
> >>> +    // increases size pressure, so we only do it at O3
> >>> +    // Do this unless only the regular threshold is explicitly set.
> >>> +    if (TailDupPlacementThreshold.getNumOccurrences() == 0 ||
> >>> +        TailDupPlacementAggressiveThreshold.getNumOccurrences() != 0)
> >>> +      TailDupSize = TailDupPlacementAggressiveThreshold;
> >>> +  }
> >>> +
> >>>    if (TailDupPlacement) {
> >>>      MPDT = &getAnalysis<MachinePostDominatorTree>();
> >>> -    unsigned TailDupSize = TailDupPlacementThreshold;
> >>>      if (MF.getFunction()->optForSize())
> >>>        TailDupSize = 1;
> >>>      TailDup.initMF(MF, MBPI, /* LayoutMode */ true, TailDupSize);
> >>> @@ -2658,7 +2683,6 @@ bool MachineBlockPlacement::runOnMachine
> >>>    buildCFGChains();
> >>>
> >>>    // Changing the layout can create new tail merging opportunities.
> >>> -  TargetPassConfig *PassConfig = &getAnalysis<TargetPassConfig>();
> >>>    // TailMerge can create jump into if branches that make CFG
> >>> irreducible for
> >>>    // HW that requires structured CFG.
> >>>    bool EnableTailMerge = !MF.getTarget().requiresStructuredCFG() &&
> >>> @@ -2666,7 +2690,7 @@ bool MachineBlockPlacement::runOnMachine
> >>>                           BranchFoldPlacement;
> >>>    // No tail merging opportunities if the block number is less than
> >>> four.
> >>>    if (MF.size() > 3 && EnableTailMerge) {
> >>> -    unsigned TailMergeSize = TailDupPlacementThreshold + 1;
> >>> +    unsigned TailMergeSize = TailDupSize + 1;
> >>>      BranchFolder BF(/*EnableTailMerge=*/true, /*CommonHoist=*/false,
> >>> *MBFI,
> >>>                      *MBPI, TailMergeSize);
> >>>
> >>>
> >>> Modified: llvm/trunk/test/CodeGen/PowerPC/tail-dup-layout.ll
> >>> URL:
> >>>
> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/PowerPC/tail-dup-layout.ll?rev=303084&r1=303083&r2=303084&view=diff
> >>>
> >>>
> ==============================================================================
> >>> --- llvm/trunk/test/CodeGen/PowerPC/tail-dup-layout.ll (original)
> >>> +++ llvm/trunk/test/CodeGen/PowerPC/tail-dup-layout.ll Mon May 15
> >>> 12:30:47 2017
> >>> @@ -1,4 +1,5 @@
> >>> -; RUN: llc -O2 < %s | FileCheck %s
> >>> +; RUN: llc -O2 -o - %s | FileCheck --check-prefix=CHECK
> >>> --check-prefix=CHECK-O2 %s
> >>> +; RUN: llc -O3 -o - %s | FileCheck --check-prefix=CHECK
> >>> --check-prefix=CHECK-O3 %s
> >>>  target datalayout = "e-m:e-i64:64-n32:64"
> >>>  target triple = "powerpc64le-grtev4-linux-gnu"
> >>>
> >>> @@ -99,11 +100,9 @@ exit:
> >>>  ; test1
> >>>  ; test2
> >>>  ; test3
> >>> -; test4
> >>>  ; optional1
> >>>  ; optional2
> >>>  ; optional3
> >>> -; optional4
> >>>  ; exit
> >>>  ; even for 50/50 branches.
> >>>  ; Tail duplication puts test n+1 at the end of optional n
> >>> @@ -157,6 +156,98 @@ test3:
> >>>    br i1 %tagbit3eq0, label %exit, label %optional3, !prof !1
> >>>  optional3:
> >>>    call void @c()
> >>> +  br label %exit
> >>> +exit:
> >>> +  ret void
> >>> +}
> >>> +
> >>> +; Intended layout:
> >>> +; The chain-of-triangles based duplicating produces the layout when 3
> >>> +; instructions are allowed for tail-duplication.
> >>> +; test1
> >>> +; test2
> >>> +; test3
> >>> +; optional1
> >>> +; optional2
> >>> +; optional3
> >>> +; exit
> >>> +;
> >>> +; Otherwise it produces the layout:
> >>> +; test1
> >>> +; optional1
> >>> +; test2
> >>> +; optional2
> >>> +; test3
> >>> +; optional3
> >>> +; exit
> >>> +
> >>> +;CHECK-LABEL: straight_test_3_instr_test:
> >>> +; test1 may have been merged with entry
> >>> +;CHECK: mr [[TAGREG:[0-9]+]], 3
> >>> +;CHECK: clrlwi {{[0-9]+}}, [[TAGREG]], 30
> >>> +;CHECK-NEXT: cmplwi {{[0-9]+}}, 2
> >>> +
> >>> +;CHECK-O3-NEXT: bne 0, .[[OPT1LABEL:[_0-9A-Za-z]+]]
> >>> +;CHECK-O3-NEXT: # %test2
> >>> +;CHECK-O3-NEXT: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 28, 29
> >>> +;CHECK-O3-NEXT: cmplwi {{[0-9]+}}, 8
> >>> +;CHECK-O3-NEXT: bne 0, .[[OPT2LABEL:[_0-9A-Za-z]+]]
> >>> +;CHECK-O3-NEXT: .[[TEST3LABEL:[_0-9A-Za-z]+]]: # %test3
> >>> +;CHECK-O3-NEXT: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 26, 27
> >>> +;CHECK-O3-NEXT: cmplwi {{[0-9]+}}, 32
> >>> +;CHECK-O3-NEXT: bne 0, .[[OPT3LABEL:[_0-9A-Za-z]+]]
> >>> +;CHECK-O3-NEXT: .[[EXITLABEL:[_0-9A-Za-z]+]]: # %exit
> >>> +;CHECK-O3: blr
> >>> +;CHECK-O3-NEXT: .[[OPT1LABEL]]:
> >>> +;CHECK-O3: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 28, 29
> >>> +;CHECK-O3-NEXT: cmplwi {{[0-9]+}}, 8
> >>> +;CHECK-O3-NEXT: beq 0, .[[TEST3LABEL]]
> >>> +;CHECK-O3-NEXT: .[[OPT2LABEL]]:
> >>> +;CHECK-O3: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 26, 27
> >>> +;CHECK-O3-NEXT: cmplwi {{[0-9]+}}, 32
> >>> +;CHECK-O3-NEXT: beq 0, .[[EXITLABEL]]
> >>> +;CHECK-O3-NEXT: .[[OPT3LABEL]]:
> >>> +;CHECK-O3: b .[[EXITLABEL]]
> >>> +
> >>> +;CHECK-O2-NEXT: beq 0, .[[TEST2LABEL:[_0-9A-Za-z]+]]
> >>> +;CHECK-O2-NEXT: # %optional1
> >>> +;CHECK-O2: .[[TEST2LABEL]]: # %test2
> >>> +;CHECK-O2-NEXT: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 28, 29
> >>> +;CHECK-O2-NEXT: cmplwi {{[0-9]+}}, 8
> >>> +;CHECK-O2-NEXT: beq 0, .[[TEST3LABEL:[_0-9A-Za-z]+]]
> >>> +;CHECK-O2-NEXT: # %optional2
> >>> +;CHECK-O2: .[[TEST3LABEL]]: # %test3
> >>> +;CHECK-O2-NEXT: rlwinm {{[0-9]+}}, [[TAGREG]], 0, 26, 27
> >>> +;CHECK-O2-NEXT: cmplwi {{[0-9]+}}, 32
> >>> +;CHECK-O2-NEXT: beq 0, .[[EXITLABEL:[_0-9A-Za-z]+]]
> >>> +;CHECK-O2-NEXT: # %optional3
> >>> +;CHECK-O2: .[[EXITLABEL:[_0-9A-Za-z]+]]: # %exit
> >>> +;CHECK-O2: blr
> >>> +
> >>> +
> >>> +define void @straight_test_3_instr_test(i32 %tag) {
> >>> +entry:
> >>> +  br label %test1
> >>> +test1:
> >>> +  %tagbit1 = and i32 %tag, 3
> >>> +  %tagbit1eq0 = icmp eq i32 %tagbit1, 2
> >>> +  br i1 %tagbit1eq0, label %test2, label %optional1, !prof !2
> >>> +optional1:
> >>> +  call void @a()
> >>> +  br label %test2
> >>> +test2:
> >>> +  %tagbit2 = and i32 %tag, 12
> >>> +  %tagbit2eq0 = icmp eq i32 %tagbit2, 8
> >>> +  br i1 %tagbit2eq0, label %test3, label %optional2, !prof !2
> >>> +optional2:
> >>> +  call void @b()
> >>> +  br label %test3
> >>> +test3:
> >>> +  %tagbit3 = and i32 %tag, 48
> >>> +  %tagbit3eq0 = icmp eq i32 %tagbit3, 32
> >>> +  br i1 %tagbit3eq0, label %exit, label %optional3, !prof !1
> >>> +optional3:
> >>> +  call void @c()
> >>>    br label %exit
> >>>  exit:
> >>>    ret void
> >>>
> >>> Modified: llvm/trunk/test/CodeGen/X86/sse1.ll
> >>> URL:
> >>>
> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/CodeGen/X86/sse1.ll?rev=303084&r1=303083&r2=303084&view=diff
> >>>
> >>>
> ==============================================================================
> >>> --- llvm/trunk/test/CodeGen/X86/sse1.ll (original)
> >>> +++ llvm/trunk/test/CodeGen/X86/sse1.ll Mon May 15 12:30:47 2017
> >>> @@ -66,7 +66,10 @@ define <4 x float> @vselect(<4 x float>*
> >>>  ; X32-NEXT:    jne .LBB1_8
> >>>  ; X32-NEXT:  .LBB1_7:
> >>>  ; X32-NEXT:    movss {{.*#+}} xmm3 = mem[0],zero,zero,zero
> >>> -; X32-NEXT:    jmp .LBB1_9
> >>> +; X32-NEXT:    unpcklps {{.*#+}} xmm2 =
> xmm2[0],xmm3[0],xmm2[1],xmm3[1]
> >>> +; X32-NEXT:    cmpl $0, {{[0-9]+}}(%esp)
> >>> +; X32-NEXT:    je .LBB1_10
> >>> +; X32-NEXT:    jmp .LBB1_11
> >>>  ; X32-NEXT:  .LBB1_1:
> >>>  ; X32-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
> >>>  ; X32-NEXT:    cmpl $0, {{[0-9]+}}(%esp)
> >>> @@ -77,11 +80,10 @@ define <4 x float> @vselect(<4 x float>*
> >>>  ; X32-NEXT:    je .LBB1_7
> >>>  ; X32-NEXT:  .LBB1_8: # %entry
> >>>  ; X32-NEXT:    xorps %xmm3, %xmm3
> >>> -; X32-NEXT:  .LBB1_9: # %entry
> >>>  ; X32-NEXT:    unpcklps {{.*#+}} xmm2 =
> xmm2[0],xmm3[0],xmm2[1],xmm3[1]
> >>>  ; X32-NEXT:    cmpl $0, {{[0-9]+}}(%esp)
> >>>  ; X32-NEXT:    jne .LBB1_11
> >>> -; X32-NEXT:  # BB#10:
> >>> +; X32-NEXT:  .LBB1_10:
> >>>  ; X32-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
> >>>  ; X32-NEXT:  .LBB1_11: # %entry
> >>>  ; X32-NEXT:    unpcklps {{.*#+}} xmm0 =
> xmm0[0],xmm1[0],xmm0[1],xmm1[1]
> >>> @@ -103,7 +105,10 @@ define <4 x float> @vselect(<4 x float>*
> >>>  ; X64-NEXT:    jne .LBB1_8
> >>>  ; X64-NEXT:  .LBB1_7:
> >>>  ; X64-NEXT:    movss {{.*#+}} xmm3 = mem[0],zero,zero,zero
> >>> -; X64-NEXT:    jmp .LBB1_9
> >>> +; X64-NEXT:    unpcklps {{.*#+}} xmm2 =
> xmm2[0],xmm3[0],xmm2[1],xmm3[1]
> >>> +; X64-NEXT:    testl %esi, %esi
> >>> +; X64-NEXT:    je .LBB1_10
> >>> +; X64-NEXT:    jmp .LBB1_11
> >>>  ; X64-NEXT:  .LBB1_1:
> >>>  ; X64-NEXT:    movss {{.*#+}} xmm1 = mem[0],zero,zero,zero
> >>>  ; X64-NEXT:    testl %edx, %edx
> >>> @@ -114,11 +119,10 @@ define <4 x float> @vselect(<4 x float>*
> >>>  ; X64-NEXT:    je .LBB1_7
> >>>  ; X64-NEXT:  .LBB1_8: # %entry
> >>>  ; X64-NEXT:    xorps %xmm3, %xmm3
> >>> -; X64-NEXT:  .LBB1_9: # %entry
> >>>  ; X64-NEXT:    unpcklps {{.*#+}} xmm2 =
> xmm2[0],xmm3[0],xmm2[1],xmm3[1]
> >>>  ; X64-NEXT:    testl %esi, %esi
> >>>  ; X64-NEXT:    jne .LBB1_11
> >>> -; X64-NEXT:  # BB#10:
> >>> +; X64-NEXT:  .LBB1_10:
> >>>  ; X64-NEXT:    movss {{.*#+}} xmm0 = mem[0],zero,zero,zero
> >>>  ; X64-NEXT:  .LBB1_11: # %entry
> >>>  ; X64-NEXT:    unpcklps {{.*#+}} xmm0 =
> xmm0[0],xmm1[0],xmm0[1],xmm1[1]
> >>>
> >>>
> >>> _______________________________________________
> >>> llvm-commits mailing list
> >>> llvm-commits at lists.llvm.org
> >>> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
>
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20170516/5ef3fb92/attachment.html>


More information about the llvm-commits mailing list