[llvm] r284311 - [GVN/PRE] Hoist global values outside of loops.
Philip Reames via llvm-commits
llvm-commits at lists.llvm.org
Thu Oct 20 18:48:21 PDT 2016
To be clear, I don't view this as needing an urgent revert. I'm happy
to discuss with this in tree, but only if the discussion actually
happens. If it doesn't, then the change should be reverted. Our
community standards are that code be reviewed and the author is expected
to be responsive to review comments both pre and post commit.
Philip
On 10/20/2016 06:35 PM, Davide Italiano wrote:
>
> I can revert this patch if you prefer.
>
>
> On Oct 20, 2016 6:24 PM, "Philip Reames" <listmail at philipreames.com
> <mailto:listmail at philipreames.com>> wrote:
>
> Was there a review thread for this? If so, I missed it, but would
> have made the following comments.
>
> Overall, this patch mixes functional and stylistic changes. The
> later should have been strictly separate and committed
> separately. Trying to review this code now is more difficult than
> it should be.
>
> I'm unclear that this patch is overall a good idea. I get that
> it's *correct*, but it's a break from our current philosophy
> around hoisting and sinking that needs discussed. If this
> discussion has happened, please provide a reference so I can catch
> up on the thread.
>
>
> On 10/15/2016 02:35 PM, Davide Italiano via llvm-commits wrote:
>
> Author: davide
> Date: Sat Oct 15 16:35:23 2016
> New Revision: 284311
>
> URL: http://llvm.org/viewvc/llvm-project?rev=284311&view=rev
> <http://llvm.org/viewvc/llvm-project?rev=284311&view=rev>
> Log:
> [GVN/PRE] Hoist global values outside of loops.
>
> In theory this could be generalized to move anything where
> we prove the operands are available, but that would require
> rewriting PRE. As NewGVN will hopefully come soon, and we're
> trying to rewrite PRE in terms of NewGVN+MemorySSA, it's probably
> not worth spending too much time on it. Fix provided by
> Daniel Berlin!
>
> Added:
> llvm/trunk/test/Transforms/GVN/PRE/hoist-loads.ll
> Modified:
> llvm/trunk/lib/Transforms/Scalar/GVN.cpp
> llvm/trunk/test/Transforms/GVN/PRE/pre-single-pred.ll
>
> Modified: llvm/trunk/lib/Transforms/Scalar/GVN.cpp
> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/GVN.cpp?rev=284311&r1=284310&r2=284311&view=diff
> <http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Transforms/Scalar/GVN.cpp?rev=284311&r1=284310&r2=284311&view=diff>
> ==============================================================================
> --- llvm/trunk/lib/Transforms/Scalar/GVN.cpp (original)
> +++ llvm/trunk/lib/Transforms/Scalar/GVN.cpp Sat Oct 15
> 16:35:23 2016
> @@ -122,69 +122,84 @@ template <> struct DenseMapInfo<GVN::Exp
> /// location of the instruction from which it was formed.
> struct llvm::gvn::AvailableValue {
> enum ValType {
> - SimpleVal, // A simple offsetted value that is accessed.
> - LoadVal, // A value produced by a load.
> - MemIntrin, // A memory intrinsic which is loaded from.
> - UndefVal // A UndefValue representing a value from dead
> block (which
> - // is not yet physically removed from the CFG).
> + SimpleVal, // A simple offsetted value that is accessed.
> + LoadVal, // A value produced by a load.
> + MemIntrinVal, // A memory intrinsic which is loaded from.
> + UndefVal, // A UndefValue representing a value from
> dead block (which
> + // is not yet physically removed from the CFG).
> + CreateLoadVal // A duplicate load can be created higher
> up in the CFG that
> + // will eliminate this one.
> };
> /// V - The value that is live out of the block.
> - PointerIntPair<Value *, 2, ValType> Val;
> + std::pair<Value *, ValType> Val;
>
> Why not use a PointerIntPair<Value *, 3>? I'm pretty sure our
> Value's are at least 8 byte aligned?
>
> Changing this, and the related code churn, should have been done
> separately.
>
> /// Offset - The byte offset in Val that is interesting
> for the load query.
> unsigned Offset;
> static AvailableValue get(Value *V, unsigned Offset = 0) {
> AvailableValue Res;
> - Res.Val.setPointer(V);
> - Res.Val.setInt(SimpleVal);
> + Res.Val.first = V;
> + Res.Val.second = SimpleVal;
> Res.Offset = Offset;
> return Res;
> }
> static AvailableValue getMI(MemIntrinsic *MI, unsigned
> Offset = 0) {
> AvailableValue Res;
> - Res.Val.setPointer(MI);
> - Res.Val.setInt(MemIntrin);
> + Res.Val.first = MI;
> + Res.Val.second = MemIntrinVal;
> Res.Offset = Offset;
> return Res;
> }
> + static AvailableValue getCreateLoad(LoadInst *LI) {
> + AvailableValue Res;
> + Res.Val.first = LI;
> + Res.Val.second = CreateLoadVal;
> + return Res;
> + }
> +
> static AvailableValue getLoad(LoadInst *LI, unsigned
> Offset = 0) {
> AvailableValue Res;
> - Res.Val.setPointer(LI);
> - Res.Val.setInt(LoadVal);
> + Res.Val.first = LI;
> + Res.Val.second = LoadVal;
> Res.Offset = Offset;
> return Res;
> }
> static AvailableValue getUndef() {
> AvailableValue Res;
> - Res.Val.setPointer(nullptr);
> - Res.Val.setInt(UndefVal);
> + Res.Val.first = nullptr;
> + Res.Val.second = UndefVal;
> Res.Offset = 0;
> return Res;
> }
> - bool isSimpleValue() const { return Val.getInt() ==
> SimpleVal; }
> - bool isCoercedLoadValue() const { return Val.getInt() ==
> LoadVal; }
> - bool isMemIntrinValue() const { return Val.getInt() ==
> MemIntrin; }
> - bool isUndefValue() const { return Val.getInt() == UndefVal; }
> + bool isSimpleValue() const { return Val.second == SimpleVal; }
> + bool isCoercedLoadValue() const { return Val.second ==
> LoadVal; }
> + bool isMemIntrinValue() const { return Val.second ==
> MemIntrinVal; }
> + bool isUndefValue() const { return Val.second == UndefVal; }
> + bool isCreateLoadValue() const { return Val.second ==
> CreateLoadVal; }
> +
> + LoadInst *getCreateLoadValue() const {
> + assert(isCreateLoadValue() && "Wrong accessor");
> + return cast<LoadInst>(Val.first);
> + }
> Value *getSimpleValue() const {
> assert(isSimpleValue() && "Wrong accessor");
> - return Val.getPointer();
> + return Val.first;
> }
> LoadInst *getCoercedLoadValue() const {
> assert(isCoercedLoadValue() && "Wrong accessor");
> - return cast<LoadInst>(Val.getPointer());
> + return cast<LoadInst>(Val.first);
> }
> MemIntrinsic *getMemIntrinValue() const {
> assert(isMemIntrinValue() && "Wrong accessor");
> - return cast<MemIntrinsic>(Val.getPointer());
> + return cast<MemIntrinsic>(Val.first);
> }
> /// Emit code at the specified insertion point to adjust
> the value defined
> @@ -1191,7 +1206,11 @@ Value *AvailableValue::MaterializeAdjust
> Value *Res;
> Type *LoadTy = LI->getType();
> const DataLayout &DL = LI->getModule()->getDataLayout();
> - if (isSimpleValue()) {
> + if (isCreateLoadValue()) {
> + Instruction *I = getCreateLoadValue()->clone();
> + I->insertBefore(InsertPt);
> + Res = I;
>
> This case is rare, please leave the simpleValue case first and put
> this at the bottom of the chain.
>
> + } else if (isSimpleValue()) {
> Res = getSimpleValue();
> if (Res->getType() != LoadTy) {
> Res = GetStoreValueForLoad(Res, Offset, LoadTy,
> InsertPt, DL);
> @@ -1379,7 +1398,7 @@ void GVN::AnalyzeLoadAvailability(LoadIn
> continue;
> }
> - if (!DepInfo.isDef() && !DepInfo.isClobber()) {
> + if (!DepInfo.isDef() && !DepInfo.isClobber() &&
> !DepInfo.isNonFuncLocal()) {
> UnavailableBlocks.push_back(DepBB);
> continue;
> }
> @@ -1390,12 +1409,25 @@ void GVN::AnalyzeLoadAvailability(LoadIn
> Value *Address = Deps[i].getAddress();
> AvailableValue AV;
> - if (AnalyzeLoadAvailability(LI, DepInfo, Address, AV)) {
> + // TODO: We can use anything where the operands are
> available, and we should
> + // learn to recreate operands in other blocks if they are
> available.
> + // Because we don't have the infrastructure in our PRE,
> we restrict this to
> + // global values, because we know the operands are always
> available.
> + if (DepInfo.isNonFuncLocal()) {
> + if (isSafeToSpeculativelyExecute(LI) &&
> + isa<GlobalValue>(LI->getPointerOperand())) {
>
> This is overly specific. This also works for function arguments,
> etc..
>
> + AV = AvailableValue::getCreateLoad(LI);
> + ValuesPerBlock.push_back(AvailableValueInBlock::get(
> + &LI->getParent()->getParent()->getEntryBlock(),
> std::move(AV)));
>
> The entry block may be a horrible place to put this load. How was
> this heuristic evaluated?
>
> + } else
> + UnavailableBlocks.push_back(DepBB);
> +
> + } else if (AnalyzeLoadAvailability(LI, DepInfo, Address,
> AV)) {
> // subtlety: because we know this was a non-local
> dependency, we know
> // it's safe to materialize anywhere between the
> instruction within
> // DepInfo and the end of it's block.
> - ValuesPerBlock.push_back(AvailableValueInBlock::get(DepBB,
> - std::move(AV)));
> + ValuesPerBlock.push_back(
> + AvailableValueInBlock::get(DepBB, std::move(AV)));
> } else {
> UnavailableBlocks.push_back(DepBB);
> }
>
> Added: llvm/trunk/test/Transforms/GVN/PRE/hoist-loads.ll
> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/GVN/PRE/hoist-loads.ll?rev=284311&view=auto
> <http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/GVN/PRE/hoist-loads.ll?rev=284311&view=auto>
> ==============================================================================
> --- llvm/trunk/test/Transforms/GVN/PRE/hoist-loads.ll (added)
> +++ llvm/trunk/test/Transforms/GVN/PRE/hoist-loads.ll Sat Oct
> 15 16:35:23 2016
> @@ -0,0 +1,53 @@
> +; RUN: opt -gvn %s -S -o - | FileCheck %s
> +
> +target datalayout = "e-m:e-i64:64-f80:128-n8:16:32:64-S128"
> +target triple = "x86_64-unknown-linux-gnu"
> +
> +; Check that the loads of @aaa, @bbb and @ccc are hoisted.
> +; CHECK-LABEL: define void @foo
> +; CHECK-NEXT: %2 = load i32, i32* @ccc, align 4
> +; CHECK-NEXT: %3 = load i32, i32* @bbb, align 4
> +; CHECK-NEXT: %4 = load i32, i32* @aaa, align 4
> +
> + at aaa = local_unnamed_addr global i32 10, align 4
> + at bbb = local_unnamed_addr global i32 20, align 4
> + at ccc = local_unnamed_addr global i32 30, align 4
> +
> +define void @foo(i32* nocapture readonly) local_unnamed_addr {
> + br label %2
> +
> + %.0 = phi i32* [ %0, %1 ], [ %3, %22 ]
> + %3 = getelementptr inbounds i32, i32* %.0, i64 1
> + %4 = load i32, i32* %.0, align 4
> + %5 = load i32, i32* @ccc, align 4
> + %6 = icmp sgt i32 %4, %5
> + br i1 %6, label %7, label %10
> +
> + %8 = load i32, i32* @bbb, align 4
> + %9 = add nsw i32 %8, %4
> + store i32 %9, i32* @bbb, align 4
> + br label %10
> +
> + %11 = load i32, i32* @bbb, align 4
> + %12 = icmp sgt i32 %4, %11
> + br i1 %12, label %13, label %16
> +
> + %14 = load i32, i32* @aaa, align 4
> + %15 = add nsw i32 %14, %4
> + store i32 %15, i32* @aaa, align 4
> + br label %16
> +
> + %17 = load i32, i32* @aaa, align 4
> + %18 = icmp sgt i32 %4, %17
> + br i1 %18, label %19, label %22
> +
> + %20 = load i32, i32* @ccc, align 4
> + %21 = add nsw i32 %20, %4
> + store i32 %21, i32* @ccc, align 4
> + br label %22
> +
> + %23 = icmp slt i32 %4, 0
> + br i1 %23, label %24, label %2
> +
> + ret void
> +}
>
> Please add several more test cases which exercise the trivial
> case, a case where we *can't* hoist, and a case where we can hoist
> because the load is explicitly marked invariant.
>
>
> Modified: llvm/trunk/test/Transforms/GVN/PRE/pre-single-pred.ll
> URL:
> http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/GVN/PRE/pre-single-pred.ll?rev=284311&r1=284310&r2=284311&view=diff
> <http://llvm.org/viewvc/llvm-project/llvm/trunk/test/Transforms/GVN/PRE/pre-single-pred.ll?rev=284311&r1=284310&r2=284311&view=diff>
> ==============================================================================
> --- llvm/trunk/test/Transforms/GVN/PRE/pre-single-pred.ll
> (original)
> +++ llvm/trunk/test/Transforms/GVN/PRE/pre-single-pred.ll Sat
> Oct 15 16:35:23 2016
> @@ -1,16 +1,9 @@
> ; RUN: opt < %s -gvn -enable-load-pre -S | FileCheck %s
> -; This testcase assumed we'll PRE the load into %for.cond,
> but we don't actually
> -; verify that doing so is safe. If there didn't _happen_ to
> be a load in
> -; %for.end, we would actually be lengthening the execution on
> some paths, and
> -; we were never actually checking that case. Now we actually
> do perform some
> -; conservative checking to make sure we don't make paths
> longer, but we don't
> -; currently get this case, which we got lucky on previously.
> -;
> -; Now that that faulty assumption is corrected, test that we
> DON'T incorrectly
> -; hoist the load. Doing the right thing for the wrong
> reasons is still a bug.
> @p = external global i32
> define i32 @f(i32 %n) nounwind {
> +; CHECK: entry:
> +; CHECK-NEXT: %0 = load i32, i32* @p
> entry:
> br label %for.cond
> @@ -22,8 +15,9 @@ for.cond: ; preds = %for.inc, %entry
> for.cond.for.end_crit_edge: ; preds = %for.cond
> br label %for.end
> +; The load of @p should be hoisted into the entry block.
> ; CHECK: for.body:
> -; CHECK-NEXT: %tmp3 = load i32, i32* @p
> +; CHECK-NEXT: %dec = add i32 %tmp3, -1
> for.body: ; preds = %for.cond
> %tmp3 = load i32, i32* @p ; <i32> [#uses=1]
> %dec = add i32 %tmp3, -1 ; <i32> [#uses=2]
>
>
> _______________________________________________
> llvm-commits mailing list
> llvm-commits at lists.llvm.org <mailto:llvm-commits at lists.llvm.org>
> http://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-commits
> <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/20161020/2906fbb6/attachment.html>
More information about the llvm-commits
mailing list