[PATCH] Choose the best consecutive candidate for a store instruction in SLP vectorizer

Wei Mi wmi at google.com
Mon Jun 22 11:36:57 PDT 2015


Ping.

On Tue, Jun 16, 2015 at 10:14 AM, Wei Mi <wmi at google.com> wrote:
> I changed the patch to remove duplicate code.
>
> Then I run the testsuite with BENCHMARKING_ONLY defined and with turbo mode and address random turned off. It is very helpful and make the perf analysis a lot easier. Thanks!
> I didn't see much perf change with the patch.
>
> Thanks,
> Wei.
>
>
> REPOSITORY
>   rL LLVM
>
> http://reviews.llvm.org/D10445
>
> Files:
>   lib/Transforms/Vectorize/SLPVectorizer.cpp
>   test/Transforms/SLPVectorizer/X86/pr23510.ll
>
> Index: lib/Transforms/Vectorize/SLPVectorizer.cpp
> ===================================================================
> --- lib/Transforms/Vectorize/SLPVectorizer.cpp
> +++ lib/Transforms/Vectorize/SLPVectorizer.cpp
> @@ -3243,15 +3243,26 @@
>
>    // Do a quadratic search on all of the given stores and find
>    // all of the pairs of stores that follow each other.
> +  SmallVector<unsigned, 16> IndexQueue;
>    for (unsigned i = 0, e = Stores.size(); i < e; ++i) {
> -    for (unsigned j = 0; j < e; ++j) {
> -      if (i == j)
> -        continue;
> -      const DataLayout &DL = Stores[i]->getModule()->getDataLayout();
> -      if (R.isConsecutiveAccess(Stores[i], Stores[j], DL)) {
> -        Tails.insert(Stores[j]);
> +    const DataLayout &DL = Stores[i]->getModule()->getDataLayout();
> +    IndexQueue.clear();
> +    // If a store has multiple consectutive store candidates, search Stores
> +    // array according to the sequence: from i+1 to e, then from i-1 to 0.
> +    // This is because usually pairing with immediate succeeding or preceding
> +    // candidate create the best chance to find slp vectorization opportunity.
> +    unsigned j = 0;
> +    for (j = i + 1; j < e; ++j)
> +      IndexQueue.push_back(j);
> +    for (j = i; j > 0; --j)
> +      IndexQueue.push_back(j - 1);
> +
> +    for (auto &k : IndexQueue) {
> +      if (R.isConsecutiveAccess(Stores[i], Stores[k], DL)) {
> +        Tails.insert(Stores[k]);
>          Heads.insert(Stores[i]);
> -        ConsecutiveChain[Stores[i]] = Stores[j];
> +        ConsecutiveChain[Stores[i]] = Stores[k];
> +        break;
>        }
>      }
>    }
> Index: test/Transforms/SLPVectorizer/X86/pr23510.ll
> ===================================================================
> --- test/Transforms/SLPVectorizer/X86/pr23510.ll
> +++ test/Transforms/SLPVectorizer/X86/pr23510.ll
> @@ -0,0 +1,37 @@
> +; PR23510
> +; RUN: opt < %s -basicaa -slp-vectorizer -S | FileCheck %s
> +
> +target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
> +target triple = "x86_64-unknown-linux-gnu"
> +
> +; CHECK: lshr <2 x i64>
> +; CHECK: lshr <2 x i64>
> +
> + at total = global i64 0, align 8
> +
> +define void @_Z3fooPml(i64* nocapture %a, i64 %i) {
> +entry:
> +  %tmp = load i64, i64* %a, align 8
> +  %shr = lshr i64 %tmp, 4
> +  store i64 %shr, i64* %a, align 8
> +  %arrayidx1 = getelementptr inbounds i64, i64* %a, i64 1
> +  %tmp1 = load i64, i64* %arrayidx1, align 8
> +  %shr2 = lshr i64 %tmp1, 4
> +  store i64 %shr2, i64* %arrayidx1, align 8
> +  %arrayidx3 = getelementptr inbounds i64, i64* %a, i64 %i
> +  %tmp2 = load i64, i64* %arrayidx3, align 8
> +  %tmp3 = load i64, i64* @total, align 8
> +  %add = add i64 %tmp3, %tmp2
> +  store i64 %add, i64* @total, align 8
> +  %tmp4 = load i64, i64* %a, align 8
> +  %shr5 = lshr i64 %tmp4, 4
> +  store i64 %shr5, i64* %a, align 8
> +  %tmp5 = load i64, i64* %arrayidx1, align 8
> +  %shr7 = lshr i64 %tmp5, 4
> +  store i64 %shr7, i64* %arrayidx1, align 8
> +  %tmp6 = load i64, i64* %arrayidx3, align 8
> +  %tmp7 = load i64, i64* @total, align 8
> +  %add9 = add i64 %tmp7, %tmp6
> +  store i64 %add9, i64* @total, align 8
> +  ret void
> +}
>
> EMAIL PREFERENCES
>   http://reviews.llvm.org/settings/panel/emailpreferences/



More information about the llvm-commits mailing list