[HotColdSplit] Introduce a cost model to control splitting behavior

The main goal of the model is to avoid *increasing* function size, as
that would eradicate any memory locality benefits from splitting. This
happens when:

  - There are too many inputs or outputs to the cold region. Argument
    materialization and reloads of outputs have a cost.

  - The cold region has too many distinct exit blocks, causing a large
    switch to be formed in the caller.

  - The code size cost of the split code is less than the cost of a
    set-up call.

A secondary goal is to prevent excessive overall binary size growth.

With the cost model in place, I experimented to find a splitting
threshold that works well in practice. To make warm & cold code easily
separable for analysis purposes, I moved split functions to a "cold"
section. I experimented with thresholds between [0, 4] and set the
default to the threshold which minimized geomean __text size.

Experiment data from building LNT+externals for X86 (N = 639 programs,
all sizes in bytes):

| Configuration | __text geom size | __cold geom size | TEXT geom size |
| **-Os**       | 1736.3           | 0, n=0           | 10961.6        |
| -Os, thresh=0 | 1740.53          | 124.482, n=134   | 11014          |
| -Os, thresh=1 | 1734.79          | 57.8781, n=90    | 10978.6        |
| -Os, thresh=2 | ** 1733.85 **    | 65.6604, n=61    | 10977.6        |
| -Os, thresh=3 | 1733.85          | 65.3071, n=61    | 10977.6        |
| -Os, thresh=4 | 1735.08          | 67.5156, n=54    | 10965.7        |
| **-Oz**       | 1554.4           | 0, n=0           | 10153          |
| -Oz, thresh=2 | ** 1552.2 **     | 65.633, n=61     | 10176          |
| **-O3**       | 2563.37          | 0, n=0           | 13105.4        |
| -O3, thresh=2 | ** 2559.49 **    | 71.1072, n=61    | 13162.4        |

Picking thresh=2 reduces the geomean __text section size by 0.14% at
-Os, -Oz, and -O3 and causes ~0.2% growth in the TEXT segment. Note that
TEXT size is page-aligned, whereas section sizes are byte-aligned.

Experiment data from building LNT+externals for ARM64 (N = 558 programs,
all sizes in bytes):

| Configuration | __text geom size | __cold geom size | TEXT geom size |
| **-Os**       | 1763.96          | 0, n=0           | 42934.9        |
| -Os, thresh=2 | ** 1760.9 **     | 76.6755, n=61    | 42934.9        |

Picking thresh=2 reduces the geomean __text section size by 0.17% at
-Os and causes no growth in the TEXT segment.

Measurements were done with D57082 (r352080) applied.

Differential Revision: https://reviews.llvm.org/D57125

llvm-svn: 352228

lib/Transforms/IPO/HotColdSplitting.cpp

@@ -80,9 +80,9 @@ static cl::opt<bool> EnableStaticAnalyis("hot-cold-static-analysis",
                               cl::init(true), cl::Hidden);
 
 static cl::opt<int>
-    SplittingThreshold("hotcoldsplit-threshold", cl::init(3), cl::Hidden,
+    SplittingThreshold("hotcoldsplit-threshold", cl::init(2), cl::Hidden,
-                       cl::desc("Code size threshold for splitting cold code "
+                       cl::desc("Base penalty for splitting cold code (as a "
-                                "(as a multiple of TCC_Basic)"));
+                                "multiple of TCC_Basic)"));
 
 namespace {
 
@@ -139,31 +139,6 @@ static bool mayExtractBlock(const BasicBlock &BB) {
          !isa<InvokeInst>(BB.getTerminator());
 }
 
-/// Check whether \p Region is profitable to outline.
-static bool isProfitableToOutline(const BlockSequence &Region,
-                                  TargetTransformInfo &TTI) {
-  // If the splitting threshold is set at or below zero, skip the usual
-  // profitability check.
-  if (SplittingThreshold <= 0)
-    return true;
-
-  if (Region.size() > 1)
-    return true;
-
-  int Cost = 0;
-  const BasicBlock &BB = *Region[0];
-  for (const Instruction &I : BB) {
-    if (isa<DbgInfoIntrinsic>(&I) || &I == BB.getTerminator())
-      continue;
-
-    Cost += TTI.getInstructionCost(&I, TargetTransformInfo::TCK_CodeSize);
-
-    if (Cost >= (SplittingThreshold * TargetTransformInfo::TCC_Basic))
-      return true;
-  }
-  return false;
-}
-
 /// Mark \p F cold. Based on this assumption, also optimize it for minimum size.
 /// Return true if the function is changed.
 static bool markFunctionCold(Function &F) {
@@ -247,6 +222,82 @@ bool HotColdSplitting::shouldOutlineFrom(const Function &F) const {
   return true;
 }
 
+/// Get the benefit score of outlining \p Region.
+static int getOutliningBenefit(ArrayRef<BasicBlock *> Region,
+                               TargetTransformInfo &TTI) {
+  // Sum up the code size costs of non-terminator instructions. Tight coupling
+  // with \ref getOutliningPenalty is needed to model the costs of terminators.
+  int Benefit = 0;
+  for (BasicBlock *BB : Region)
+    for (Instruction &I : BB->instructionsWithoutDebug())
+      if (&I != BB->getTerminator())
+        Benefit +=
+            TTI.getInstructionCost(&I, TargetTransformInfo::TCK_CodeSize);
+
+  return Benefit;
+}
+
+/// Get the penalty score for outlining \p Region.
+static int getOutliningPenalty(ArrayRef<BasicBlock *> Region,
+                               unsigned NumInputs, unsigned NumOutputs) {
+  int Penalty = SplittingThreshold;
+  LLVM_DEBUG(dbgs() << "Applying penalty for splitting: " << Penalty << "\n");
+
+  // If the splitting threshold is set at or below zero, skip the usual
+  // profitability check.
+  if (SplittingThreshold <= 0)
+    return Penalty;
+
+  // The typical code size cost for materializing an argument for the outlined
+  // call.
+  LLVM_DEBUG(dbgs() << "Applying penalty for: " << NumInputs << " inputs\n");
+  const int CostForArgMaterialization = TargetTransformInfo::TCC_Basic;
+  Penalty += CostForArgMaterialization * NumInputs;
+
+  // The typical code size cost for an output alloca, its associated store, and
+  // its associated reload.
+  LLVM_DEBUG(dbgs() << "Applying penalty for: " << NumOutputs << " outputs\n");
+  const int CostForRegionOutput = 3 * TargetTransformInfo::TCC_Basic;
+  Penalty += CostForRegionOutput * NumOutputs;
+
+  // Find the number of distinct exit blocks for the region. Use a conservative
+  // check to determine whether control returns from the region.
+  bool NoBlocksReturn = true;
+  SmallPtrSet<BasicBlock *, 2> SuccsOutsideRegion;
+  for (BasicBlock *BB : Region) {
+    // If a block has no successors, only assume it does not return if it's
+    // unreachable.
+    if (succ_empty(BB)) {
+      NoBlocksReturn &= isa<UnreachableInst>(BB->getTerminator());
+      continue;
+    }
+
+    for (BasicBlock *SuccBB : successors(BB)) {
+      if (find(Region, SuccBB) == Region.end()) {
+        NoBlocksReturn = false;
+        SuccsOutsideRegion.insert(SuccBB);
+      }
+    }
+  }
+
+  // Apply a `noreturn` bonus.
+  if (NoBlocksReturn) {
+    LLVM_DEBUG(dbgs() << "Applying bonus for: " << Region.size()
+                      << " non-returning terminators\n");
+    Penalty -= Region.size();
+  }
+
+  // Apply a penalty for having more than one successor outside of the region.
+  // This penalty accounts for the switch needed in the caller.
+  if (!SuccsOutsideRegion.empty()) {
+    LLVM_DEBUG(dbgs() << "Applying penalty for: " << SuccsOutsideRegion.size()
+                      << " non-region successors\n");
+    Penalty += (SuccsOutsideRegion.size() - 1) * TargetTransformInfo::TCC_Basic;
+  }
+
+  return Penalty;
+}
+
 Function *HotColdSplitting::extractColdRegion(const BlockSequence &Region,
                                               DominatorTree &DT,
                                               BlockFrequencyInfo *BFI,
@@ -261,6 +312,18 @@ Function *HotColdSplitting::extractColdRegion(const BlockSequence &Region,
                    /* AllowAlloca */ false,
                    /* Suffix */ "cold." + std::to_string(Count));
 
+  // Perform a simple cost/benefit analysis to decide whether or not to permit
+  // splitting.
+  SetVector<Value *> Inputs, Outputs, Sinks;
+  CE.findInputsOutputs(Inputs, Outputs, Sinks);
+  int OutliningBenefit = getOutliningBenefit(Region, TTI);
+  int OutliningPenalty =
+      getOutliningPenalty(Region, Inputs.size(), Outputs.size());
+  LLVM_DEBUG(dbgs() << "Split profitability: benefit = " << OutliningBenefit
+                    << ", penalty = " << OutliningPenalty << "\n");
+  if (OutliningBenefit <= OutliningPenalty)
+    return nullptr;
+
   Function *OrigF = Region[0]->getParent();
   if (Function *OutF = CE.extractCodeRegion()) {
     User *U = *OutF->user_begin();
@@ -556,14 +619,6 @@ bool HotColdSplitting::outlineColdRegions(Function &F, bool HasProfileSummary) {
     assert(!Region.empty() && "Empty outlining region in worklist");
     do {
       BlockSequence SubRegion = Region.takeSingleEntrySubRegion(*DT);
-      if (!isProfitableToOutline(SubRegion, TTI)) {
-        LLVM_DEBUG({
-          dbgs() << "Skipping outlining; not profitable to outline\n";
-          SubRegion[0]->dump();
-        });
-        continue;
-      }
-
       LLVM_DEBUG({
         dbgs() << "Hot/cold splitting attempting to outline these blocks:\n";
         for (BasicBlock *BB : SubRegion)

test/Transforms/HotColdSplit/X86/extraction-subregion-breaks-phis.ll

@@ -1,63 +0,0 @@
-; RUN: opt -S -hotcoldsplit -hotcoldsplit-threshold=1 < %s | FileCheck %s
-
-target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
-target triple = "x86_64-apple-macosx10.14.0"
-
-; CHECK-LABEL: define {{.*}}@foo(
-; CHECK: call {{.*}}@foo.cold.1(
-; CHECK: unreachable
-
-; CHECK-LABEL: define {{.*}}@foo.cold.1(
-; CHECK: switch i32 undef, label %sw.epilog.i
-define void @foo(i32 %QMM) {
-entry:
-  switch i32 %QMM, label %entry.if.end16_crit_edge [
-    i32 1, label %if.then
-  ]
-
-entry.if.end16_crit_edge:                         ; preds = %entry
-  br label %if.end16
-
-if.then:                                          ; preds = %entry
-  br i1 undef, label %cond.true.i.i, label %_ZN10StringView8popFrontEv.exit.i
-
-cond.true.i.i:                                    ; preds = %if.then
-  ret void
-
-_ZN10StringView8popFrontEv.exit.i:                ; preds = %if.then
-  switch i32 undef, label %sw.epilog.i [
-    i32 81, label %if.end16
-    i32 82, label %sw.bb4.i
-    i32 83, label %sw.bb8.i
-    i32 84, label %sw.bb12.i
-    i32 65, label %if.end16
-    i32 66, label %sw.bb20.i
-    i32 67, label %sw.bb24.i
-    i32 68, label %sw.bb28.i
-  ]
-
-sw.bb4.i:                                         ; preds = %_ZN10StringView8popFrontEv.exit.i
-  br label %if.end16
-
-sw.bb8.i:                                         ; preds = %_ZN10StringView8popFrontEv.exit.i
-  br label %if.end16
-
-sw.bb12.i:                                        ; preds = %_ZN10StringView8popFrontEv.exit.i
-  br label %if.end16
-
-sw.bb20.i:                                        ; preds = %_ZN10StringView8popFrontEv.exit.i
-  br label %if.end16
-
-sw.bb24.i:                                        ; preds = %_ZN10StringView8popFrontEv.exit.i
-  br label %if.end16
-
-sw.bb28.i:                                        ; preds = %_ZN10StringView8popFrontEv.exit.i
-  br label %if.end16
-
-sw.epilog.i:                                      ; preds = %_ZN10StringView8popFrontEv.exit.i
-  br label %if.end16
-
-if.end16:                                         ; preds = %sw.epilog.i, %sw.bb28.i, %sw.bb24.i, %sw.bb20.i, %sw.bb12.i, %sw.bb8.i, %sw.bb4.i, %_ZN10StringView8popFrontEv.exit.i, %_ZN10StringView8popFrontEv.exit.i, %entry.if.end16_crit_edge
-  %0 = phi i8 [ 0, %entry.if.end16_crit_edge ], [ 0, %_ZN10StringView8popFrontEv.exit.i ], [ 0, %_ZN10StringView8popFrontEv.exit.i ], [ 1, %sw.bb4.i ], [ 2, %sw.bb8.i ], [ 3, %sw.bb12.i ], [ 1, %sw.bb20.i ], [ 2, %sw.bb24.i ], [ 3, %sw.bb28.i ], [ 0, %sw.epilog.i ]
-  unreachable
-}

test/Transforms/HotColdSplit/X86/outline-expensive.ll

@@ -1,25 +0,0 @@
-; The magic number 6 comes from (1 * TCC_Expensive) + (1 * CostOfCallX86).
-; RUN: opt -hotcoldsplit -hotcoldsplit-threshold=6 -S < %s | FileCheck %s
-
-; Test that we outline even though there are only two cold instructions. TTI
-; should determine that they are expensive in terms of code size.
-
-target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
-target triple = "x86_64-apple-macosx10.14.0"
-
-; CHECK-LABEL: @fun
-; CHECK: call void @fun.cold.1
-define void @fun(i32 %x) {
-entry:
-  br i1 undef, label %if.then, label %if.else
-
-if.then:
-  ret void
-
-if.else:
-  %y = sdiv i32 %x, 111
-  call void @sink(i32 %y)
-  ret void
-}
-
-declare void @sink(i32 %x) cold

test/Transforms/HotColdSplit/addr-taken.ll

@@ -1,4 +1,4 @@
-; RUN: opt -hotcoldsplit -hotcoldsplit-threshold=0 -S < %s | FileCheck %s
+; RUN: opt -hotcoldsplit -hotcoldsplit-threshold=-1 -S < %s | FileCheck %s
 
 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-macosx10.14.0"

test/Transforms/HotColdSplit/apply-noreturn-bonus.ll

@@ -0,0 +1,26 @@
+; REQUIRES: asserts
+; RUN: opt -hotcoldsplit -debug-only=hotcoldsplit -S < %s -o /dev/null 2>&1 | FileCheck %s
+
+declare void @sink() cold
+
+define void @foo(i32 %arg) {
+entry:
+  br i1 undef, label %cold1, label %exit
+
+cold1:
+  ; CHECK: Applying bonus for: 4 non-returning terminators
+  call void @sink()
+  br i1 undef, label %cold2, label %cold3
+
+cold2:
+  br label %cold4
+
+cold3:
+  br label %cold4
+
+cold4:
+  unreachable
+
+exit:
+  ret void
+}

test/Transforms/HotColdSplit/apply-penalty-for-inputs.ll

@@ -0,0 +1,19 @@
+; REQUIRES: asserts
+; RUN: opt -hotcoldsplit -debug-only=hotcoldsplit -S < %s -o /dev/null 2>&1 | FileCheck %s
+
+declare void @sink(i32*, i32, i32) cold
+
+@g = global i32 0
+
+define void @foo(i32 %arg) {
+  %local = load i32, i32* @g
+  br i1 undef, label %cold, label %exit
+
+cold:
+  ; CHECK: Applying penalty for: 2 inputs
+  call void @sink(i32* @g, i32 %arg, i32 %local)
+  ret void
+
+exit:
+  ret void
+}

test/Transforms/HotColdSplit/apply-penalty-for-outputs.ll

@@ -0,0 +1,22 @@
+; REQUIRES: asserts
+; RUN: opt -hotcoldsplit -debug-only=hotcoldsplit -S < %s -o /dev/null 2>&1 | FileCheck %s
+
+declare void @sink() cold
+
+@g = global i32 0
+
+define i32 @foo(i32 %arg) {
+entry:
+  br i1 undef, label %cold, label %exit
+
+cold:
+  ; CHECK: Applying penalty for: 1 output
+  ; CHECK: Applying penalty for: 1 non-region successors
+  %local = load i32, i32* @g
+  call void @sink()
+  br label %exit
+
+exit:
+  %p = phi i32 [ %local, %cold ], [ 0, %entry ]
+  ret i32 %p
+}

test/Transforms/HotColdSplit/apply-successor-penalty.ll

@@ -0,0 +1,53 @@
+; REQUIRES: asserts
+; RUN: opt -hotcoldsplit -debug-only=hotcoldsplit -S < %s -o /dev/null 2>&1 | FileCheck %s
+
+declare void @sink() cold
+
+; CHECK-LABEL: Outlining in one_non_region_successor
+define void @one_non_region_successor(i32 %arg) {
+entry:
+  br i1 undef, label %cold1, label %exit
+
+cold1:
+  ; CHECK: Applying penalty for: 1 non-region successor
+  call void @sink()
+  br i1 undef, label %cold2, label %cold3
+
+cold2:
+  br i1 undef, label %cold4, label %exit
+
+cold3:
+  br i1 undef, label %cold4, label %exit
+
+cold4:
+  unreachable
+
+exit:
+  ret void
+}
+
+; CHECK-LABEL: Outlining in two_non_region_successor
+define void @two_non_region_successors(i32 %arg) {
+entry:
+  br i1 undef, label %cold1, label %exit1
+
+cold1:
+  ; CHECK: Applying penalty for: 2 non-region successors
+  call void @sink()
+  br i1 undef, label %cold2, label %cold3
+
+cold2:
+  br i1 undef, label %cold4, label %exit1
+
+cold3:
+  br i1 undef, label %cold4, label %exit2
+
+cold4:
+  unreachable
+
+exit1:
+  br label %exit2
+
+exit2:
+  ret void
+}

test/Transforms/HotColdSplit/outline-disjoint-diamonds.ll

@@ -1,4 +1,4 @@
-; RUN: opt -S -hotcoldsplit -hotcoldsplit-threshold=0 < %s 2>&1 | FileCheck %s
+; RUN: opt -S -hotcoldsplit -hotcoldsplit-threshold=-1 < %s 2>&1 | FileCheck %s
 
 ; CHECK-LABEL: define {{.*}}@fun
 ; CHECK: call {{.*}}@fun.cold.2(

test/Transforms/HotColdSplit/resume.ll

@@ -1,4 +1,4 @@
-; RUN: opt -hotcoldsplit -hotcoldsplit-threshold=0 -S < %s | FileCheck %s
+; RUN: opt -hotcoldsplit -hotcoldsplit-threshold=-1 -S < %s | FileCheck %s
 
 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
 target triple = "x86_64-apple-macosx10.14.0"

test/Transforms/HotColdSplit/split-cold-2.ll

@@ -1,5 +1,5 @@
-; RUN: opt -hotcoldsplit -hotcoldsplit-threshold=0 -pass-remarks=hotcoldsplit -S < %s 2>&1 | FileCheck %s
+; RUN: opt -hotcoldsplit -hotcoldsplit-threshold=-1 -pass-remarks=hotcoldsplit -S < %s 2>&1 | FileCheck %s
-; RUN: opt -hotcoldsplit-threshold=0 -passes=hotcoldsplit -pass-remarks=hotcoldsplit -S < %s 2>&1 | FileCheck %s
+; RUN: opt -passes=hotcoldsplit -hotcoldsplit-threshold=-1 -pass-remarks=hotcoldsplit -S < %s 2>&1 | FileCheck %s
 
 ; Make sure this compiles. This test used to fail with an invalid phi node: the
 ; two predecessors were outlined and the SSA representation was invalid.