[llvm][CodeGen] Machine Function Splitter

[snehasish]

Summary

We introduce a codegen optimization pass which splits functions into hot and cold
parts. This pass leverages the basic block sections feature recently
introduced in LLVM from the Propeller project. The pass targets
functions with profile coverage, identifies cold blocks and moves them
to a separate section. The linker groups all cold blocks across
functions together, decreasing fragmentation and improving icache and
itlb utilization.

We evaluated the Machine Function Splitter pass on clang bootstrap and SPECInt 2017.
For clang bootstrap we observe a mean 2.33% runtime improvement with a
~32% reduction in itlb and stlb misses. Additionally, l1 icache misses
reduced by 9.5% while l2 instruction misses reduced by 20%.
For SPECInt we report the change in IntRate the C/C++
benchmarks. All benchmarks apart from mcf and x264 improve, on average
by 0.6% with the max for deepsjeng at 1.6%.

Benchmark               % Change (IntRate)
500.perlbench_r          0.78
502.gcc_r                0.82
505.mcf_r               -0.30
520.omnetpp_r            0.18
523.xalancbmk_r          0.37
525.x264_r              -0.46
531.deepsjeng_r          1.61
541.leela_r              0.83
557.xz_r                 0.15

[tmsriram]

llvm/lib/CodeGen/BBSectionsPrepare.cpp

72 ↗

(On Diff #283424)

Should we rename the .h and .cpp so that they have the same prefix? Maybe BasicBlockSections.h and BasicBlockSections.cpp?

llvm/lib/CodeGen/TargetPassConfig.cpp

218

Why not call this split-machine-functions too for consistency?

llvm/test/CodeGen/X86/machine-function-splitter.ll

9

Also, check if the block is moved to the cold region has the expected call instruction?

[efriedma]

We probably need to discuss how to make basic-block-section stuff work on non-X86 targets at some point, but I guess we don't have to do it in this patch if it's off by default.

[xbolva00]

I am wondering what is is your opinion on machine unroller/reroller? Aggressive loop unrolling may destroy code cache too.

[davidxl]

Regarding reroller -- compiler with PGO will adjust the agressiveness of the unroller based on instruction workset size estimation. Doing this in later pass or in Propeller can help catch cases that are mis-handled.

[davidxl]

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

90

Add an internal option here (the coldness threshold) for experimental purpose.

I also suggest add an option to specify programSummary based coldness threshold such as 99.99 percentile coldness. The default cutoff is 99.9999% defined in ProfileSummaryInfo.cpp: ProfileSummaryCutoffCold

[hiraditya]

Please share performance numbers for publicly available workload(s).

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

78

Do we need to renumber?

[tmsriram]

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

78

Renumbering makes the sorting easy. The sorting will preserve the basic block order for the blocks that are not split.

[snehasish]

Updated diff based on review comments.

• Added two mllvm options to control cold count and threshold based split.
• Added tests for the new options.
• Updated test to check for the absence of unexpected blocks.
• Renamed BBSectionsPrepare pass and rebased this diff on the change.

[snehasish]

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

78

We need to ensure that the order is preserved so that we don't perturb the decisions made by prior passes such as MachineBlockPlacement. Renumbering simplifies the code that needs to be shared with the BasicBlockSections pass.

llvm/lib/CodeGen/TargetPassConfig.cpp

218

We can't register two options with the same string, i.e. "split-machine-functions".

[snehasish]

Simplify the cold count check.

[hiraditya]

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

78

I see we are renumbering both in MachineFunctionSplitter and BasicBlockSections

[hiraditya]

We evaluated the Machine Function Splitter pass on clang bootstrap and SPECInt 2017.

Could you share the details of the machine as well? The improvements are well within noise.

For clang bootstrap we observe a mean 2.33% runtime improvement with a
~32% reduction in itlb and stlb misses

While itlb reduction looks quite impressive, it doesn't seem to translate quite well to the runtime improvement. Did we see consistent >2% improvement with multiple runs? Please share the numbers.

[tmsriram]

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

78

The "bbsections-prepare" pass and the machine function splitter pass are intentionally made mutually exclusive. If bbsections is explicitly requested, machine function splitter does not apply. Please see the change in TargetPassConfig.cpp

[snehasish]

Could you share the details of the machine as well?

Sure, these were measured on a Lenovo P920 workstation -- Intel Skylake based Xeon(R) Gold 6154 CPU.

The improvements are well within noise.

For SPEC, the reported intrate improvement numbers are an average across 5 iterations. Note that SPEC binaries are tiny in size may only improve code locality in some cases.

While itlb reduction looks quite impressive, it doesn't seem to translate quite well to the runtime improvement.

It stands to reason that removing the itlb bottleneck will expose the next one :) We could dig deeper by looking into how the top down profile changes with and without splitting.

Did we see consistent >2% improvement with multiple runs? Please share the numbers.

We see consistent 2%+ improvements over FDO optimized binaries. The numbers reported are averaged across 10 runs, here is the data for one such experiment where 500 invocations of clang were executed and the overall end to end user time was measured. For completeness, I have included the data for a hot-cold-split optimized binary as well. Note this particular experiment does not use ThinLTO for any of the builds since I had some trouble running the hot-cold-split pass with ThinLTO enabled.

|----------------------------------|----------------------|----------------|-----------|
|                                  | User time in seconds ($ time run-commands.sh)     |
|----------------------------------|----------------------|----------------|-----------|
| Run #                            | FDO baseline         | Hot cold split | MFS       |
|                                1 |               484.65 |          479.2 |    466.93 |
|                                2 |                483.4 |         478.28 |    470.25 |
|                                3 |               485.57 |         479.15 |    470.36 |
|                                4 |               480.37 |         480.34 |    469.85 |
|                                5 |               482.97 |         478.18 |    471.93 |
|                                6 |               484.06 |         479.74 |    473.27 |
|                                7 |               482.67 |         477.42 |    472.56 |
|                                8 |               483.53 |         476.99 |    474.58 |
|                                9 |               486.43 |         480.76 |    473.92 |
|                               10 |               489.94 |         480.11 |    471.42 |
|----------------------------------|----------------------|----------------|-----------|
| 2 Tail Paired T-Test vs Baseline |                      |      0.0000636 | 0.0000006 |
|----------------------------------|----------------------|----------------|-----------|
| Average                          |              484.359 |        479.017 |   471.507 |
|----------------------------------|----------------------|----------------|-----------|
| % Change                         |                      |           1.10 |      2.65 |
|----------------------------------|----------------------|----------------|-----------|

Here is the data for TLB and icache. Each event was collected independently along with instructions to ensure no multiplexing. The variance reported by perf was less than 1% for each event (often less than 0.5%).

|-----------|--------------------------------------------|--------------------------------------------------------|
|           | $ perf stat -r 3 -e frontend_retired.${EVENT}:u,instructions:u -- run-commands.sh                   |
|-----------|--------------------------------------------|--------------------------------------------------------|
|           | Machine Function Splitter                  | FDO Baseline                                           |
|-----------|--------------------------------------------|--------------------------------------------------------|
| EVENT     | Misses        | Instructions      | MPKI   | Misses         | Instructions      | MPKI   | % Change |
| itlb_miss | 1,411,325,040 | 1,618,495,692,919 | 0.8720 |  2,066,003,373 | 1,618,097,715,534 | 1.2768 |    31.70 |
| stlb_miss |   131,949,440 | 1,618,466,757,079 | 0.0815 |    195,471,938 | 1,618,061,281,016 | 0.1208 |    32.51 |
| l1i_miss  | 9,678,255,804 | 1,618,479,987,914 | 5.9798 | 10,698,143,090 | 1,618,081,273,918 | 6.6116 |     9.56 |
| l2_miss   |   434,287,963 | 1,618,443,723,597 | 0.2683 |    542,869,835 | 1,618,081,904,973 | 0.3355 |    20.02 |
|-----------|--------------------------------------------|--------------------------------------------------------|

[snehasish]

Update PSI metadata to fix assert failure.

[hiraditya]

Thanks for adding the results, could you share the script to measure bootstrap numbers?

In HCS the ability to keep cold functions in a separate section was added in: D85331 (cc: @rjf ), can we try with -mllvm -enable-cold-section to compare with MachineFuncionSplitter.

[snehasish]

In D85368#2205653, @hiraditya wrote:

Thanks for adding the results, could you share the script to measure bootstrap numbers?

I've uploaded a Makefile here which will allow you to run the bootstrap benchmarks. Applying this patch on a local llvm repo and pointing the Makefile at it should be sufficient to get you going.

In HCS the ability to keep cold functions in a separate section was added in: D85331 (cc: @rjf ), can we try with -mllvm -enable-cold-section to compare with MachineFuncionSplitter.

We already incorporate this in our evaluation since we link using lld along with the flag -z,keep-text-section-prefix. Since the extracted functions are marked cold, they are assigned a .text.unlikely prefix. Passing -z,keep-text-section-prefix to lld ensures that these functions are placed in the appropriate output section achieving the same goal of improving locality for hot code. The impact of this can be seen in the binary characteristics we shared in the original RFC which showed a 41% and 47% decrease in size of .text and .text.hot respectively for the hot cold split pass.

We applied patch D85331 and find similar results. Comparing the sections of the binary (hot cold split vs PGO baseline), we find a new __llvm_cold section along with similar fractions of code extracted from .text and .text.hot --

   FILE SIZE        VM SIZE    
--------------  -------------- 
 [NEW] +7.31Mi  [NEW] +7.31Mi    __llvm_cold
  +64% +3.21Mi   +64% +3.21Mi    .eh_frame
  +26% +2.77Mi  [ = ]       0    .strtab
  +27%  +711Ki  [ = ]       0    .symtab
  +31%  +236Ki   +31%  +236Ki    .eh_frame_hdr
 +3.1%     +12  [ = ]       0    .shstrtab
  +29%      +9   +29%      +9    [LOAD #3 [RX]]
 +0.8%      +6  +0.8%      +6    [LOAD #2 [R]]
 -7.1%      -1  [ = ]       0    [Unmapped]
 -0.0%    -246  -0.0%    -246    .dynstr
 -0.1% -4.05Ki  -0.1% -4.05Ki    .rodata
 -1.4% -5.73Ki  -1.4% -5.73Ki    .text.startup
 -1.6%  -536Ki  -1.6%  -536Ki    .text.unlikely
-46.7% -2.51Mi -46.7% -2.51Mi    .text.hot
-43.8% -2.89Mi -43.8% -2.89Mi    .text
  +10% +8.28Mi  +7.2% +4.82Mi    TOTAL

Running the benchmarks with the patch enabled and comparing against the MachineFunctionSplitter. We see similar performance numbers --

|----------|----------|----------------|---------------------------|
| Run #    | Baseline | Hot Cold Split | Machine Function Splitter |
|----------|----------|----------------|---------------------------|
|        1 |   501.25 |         490.84 |                    489.48 |
|        2 |   504.22 |         491.66 |                    493.42 |
|        3 |   500.04 |          492.7 |                    489.18 |
|        4 |    499.4 |         493.31 |                    489.47 |
|        5 |   495.62 |          496.1 |                    488.79 |
|        6 |   500.62 |         495.61 |                    488.41 |
|        7 |   501.81 |         494.45 |                    487.67 |
|        8 |   496.96 |         495.91 |                    490.91 |
|        9 |   500.22 |         497.17 |                       488 |
|       10 |   499.66 |         493.81 |                    489.65 |
|----------|----------|----------------|---------------------------|
| Average  |   499.98 |        494.156 |                   489.498 |
| % Change |          |           1.16 |                      2.10 |
|----------|----------|----------------|---------------------------|

[snehasish]

Add an option to exclude specific sections.

• Add option -mfs-excluded-sections to allow users to specify section names to exclude.
• Add a test for the option.

[hiraditya]

Overall approach looks good to me even when we don't see good SPEC-17 numbers as the optimization is intended to reduce page faults. The improvements would be more pronounced in large applications. I'll review the code in more detail in the next few days. Thanks for working on this.

[hiraditya]

please run clang-format on the patch.

llvm/lib/CodeGen/BasicBlockSections.cpp

232

nit: tab?

277

Is this comment necessary?

[hiraditya]

Can we add more test cases to include eh_pad, invokeinst,

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

113

std::find?

[tmsriram]

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

111

I think excluding sections needs a bit more thought and we should do this as a separate patch if it is useful but I think a linker solution would be more favorable.

1. From what I understand, when a user specifies section names using the section keyword, then the expectation is that all functions marked with that section name will be grouped together. With function splitting, since you attach the ".cold" suffix to such sections that are split, there is no guarantee that the linker will place them together as these are not prefixed as ".text".

2. To overcome the above problem, the option to exclude such sections from being split is not ideal either as it moves the burden to the user to get this right with appropriate options.

3. I think the temporary fix is to not split sections which are not prefixed as ".text". You can add a "FIXME:" comment here to describe why you are doing this. Moving forward, we can look at a linker solution where '.' is treated as a valid section name separator and sections with identical prefix before the "." are always grouped together even if they are not named ".text".

4. I think we can move this handling as an enhancement in another patch.

[tmsriram]

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

111

Correction: I meant ".unlikely" and not ".cold".

[efriedma]

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

111

I think I'd rather make splitting for functions with an explicit section attribute opt-in, rather than opt-out. The user might have a strong need to emit a function in a particular section (for example, if the name is mentioned in a linker script). If someone is messing with section attributes in the first place, I'd like to be conservative by default.

llvm/lib/CodeGen/TargetPassConfig.cpp

218

Why do we need two options to control the same thing?

[snehasish]

Address review comments.

• Remove excluded section attribute option.
• Don't split functions which have a section attribute.
• Add a test for ehpads.
• Remove redundant comments and clang-format.

[snehasish]

Thanks for the reviews!
@hiraditya I've added a test to ensure ehpads are not split out. Let me know if there are additional cases you want to cover.

llvm/lib/CodeGen/BasicBlockSections.cpp

232

This seems to be the output from clang-format. The diff does not have a tab but it looks like phabricator is showing it as one?

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

111

@tmsriram I think you meant to exclude such functions from being split. I agree, taking into consideration @efriedma's comment, I've removed the option and made this a conservative check on the section attribute. Any function with the section attribute set is not split.

113

Code referencing this comment was removed.

llvm/lib/CodeGen/TargetPassConfig.cpp

218

In this patch we added two options

1. An option in llvm/lib/CodeGen/CommandFlags.cpp "split-machine-functions" so that llc can be used to invoke it in the tests.
2. We added a temporary option in llvm/lib/CodeGen/TargetPassConfig.cpp so that it can be invoked when running with clang or lld (for LTO).

AFAICT we cant use (2) for tests and having (1) makes it easy to compile things without an intermediate llc step. We plan on removing (2) in a future patch which will add appropriate options to clang (-fsplit-machine-functions) and lld (--lto-split-machine-functions).

[snehasish]

llvm/lib/CodeGen/TargetPassConfig.cpp

218

Correction: We can use (2) for tests by passing "-enable-split-machine-functions" to llc however since we plan to introduce clang and lld flags in the near future it seems cleaner to leave the llc flag in place and just remove (2) when that happens rather than reintroduce it. WDYT?

[efriedma]

llvm/lib/CodeGen/TargetPassConfig.cpp

218

clang doesn't call RegisterCodeGenFlags? That seems like something we should consider changing.

[snehasish]

llvm/lib/CodeGen/TargetPassConfig.cpp

218

The clang driver does not register the codegen flags, the only clang tool which does is clang-fuzzer. A small patch like the one below would do the trick for basic functionality. More plumbing might be needed to print the appropriate flags from the driver. I think this is probably worth more discussion and beyond the scope of this patch.

diff --git a/clang/tools/driver/cc1as_main.cpp b/clang/tools/driver/cc1as_main.cpp
index 87047be3c2b..0b9b5673d3e 100644
--- a/clang/tools/driver/cc1as_main.cpp
+++ b/clang/tools/driver/cc1as_main.cpp
@@ -21,6 +21,7 @@
 #include "llvm/ADT/STLExtras.h"
 #include "llvm/ADT/StringSwitch.h"
 #include "llvm/ADT/Triple.h"
+#include "llvm/CodeGen/CommandFlags.h"
 #include "llvm/IR/DataLayout.h"
 #include "llvm/MC/MCAsmBackend.h"
 #include "llvm/MC/MCAsmInfo.h"
@@ -61,6 +62,8 @@ using namespace clang::driver::options;
 using namespace llvm;
 using namespace llvm::opt;
 
+static codegen::RegisterCodeGenFlags CGF;
+
 namespace {
 
 /// Helper class for representing a single invocation of the assembler.

[tmsriram]

LGTM, thanks for doing this!

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

110

Maybe a comment here that this is useful while sorting the blocks?

118

Do we need a fixme comment here to say that we could split out landing pads if the exception patch for bb sections lands?

[snehasish]

Add a comment to explain renumbering, FIXME for ehpads.

• Added a comment to explain why renumbering blocks is necessary.
• Added a FIXME and pointer to exceptions splitting patch.

[snehasish]

Thanks for the comments.
@efriedma @hiraditya - Let me know if you have any further comments, I will wait till EOD Thursday 08/26. If not I'll take that as go ahead to commit this change.

[hiraditya]

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

124

nit: redundant braces,

llvm/lib/CodeGen/TargetPassConfig.cpp

216

Remove this FIXME. most passes continue to have cl::opt anyways.

[hiraditya]

nit: If FIXME's are mostly future works, then please replace them with TODOs.

[snehasish]

Address reviewer comments.

• Remove redundant braces.
• Remove FIXME for cl::opt.
• s/FIXME/TODO for future work.

[snehasish]

Thanks for the comments all! The builds look green and I'm going to go ahead and push this.

llvm/include/llvm/CodeGen/BasicBlockSectionUtils.h

+//===- BasicBlockSectionUtils.h - Utilities for basic block sections     --===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_CODEGEN_BASICBLOCKSECTIONUTILS_H
+#define LLVM_CODEGEN_BASICBLOCKSECTIONUTILS_H
+
+#include "llvm/ADT/STLExtras.h"
+
+namespace llvm {
+
+class MachineFunction;
+class MachineBasicBlock;
+
+using MachineBasicBlockComparator =
+    function_ref<bool(const MachineBasicBlock &, const MachineBasicBlock &)>;
+
+void sortBasicBlocksAndUpdateBranches(MachineFunction &MF,
+                                      MachineBasicBlockComparator MBBCmp);
+
+} // end namespace llvm
+
+#endif // LLVM_CODEGEN_BASICBLOCKSECTIONUTILS_H

llvm/include/llvm/CodeGen/CommandFlags.h

 llvm::DebuggerKind getDebuggerTuningOpt();
 
 bool getEnableStackSizeSection();
 
 bool getEnableAddrsig();
 
 bool getEmitCallSiteInfo();
 
+bool getEnableMachineFunctionSplitter();
+
 bool getEnableDebugEntryValues();
 
 bool getForceDwarfFrameSection();
 
 bool getXRayOmitFunctionIndex();
 
 /// Create this object with static storage to register codegen-related command
 /// line options.

llvm/include/llvm/CodeGen/MachineFunction.h

public:
   StringRef getName() const;
 
   /// getFunctionNumber - Return a unique ID for the current function.
   unsigned getFunctionNumber() const { return FunctionNumber; }
 
   /// Returns true if this function has basic block sections enabled.
   bool hasBBSections() const {
     return (BBSectionsType == BasicBlockSection::All ||
-            BBSectionsType == BasicBlockSection::List);
+            BBSectionsType == BasicBlockSection::List ||
+            BBSectionsType == BasicBlockSection::Preset);
   }
 
   /// Returns true if basic block labels are to be generated for this function.
   bool hasBBLabels() const {
     return BBSectionsType == BasicBlockSection::Labels;
   }
 
   void setBBSectionsType(BasicBlockSection V) { BBSectionsType = V; }

llvm/include/llvm/CodeGen/Passes.h

namespace llvm {
   FunctionPass *createUnreachableBlockEliminationPass();
 
   /// createBasicBlockSections Pass - This pass assigns sections to machine
   /// basic blocks and is enabled with -fbasic-block-sections. Buf is a memory
   /// buffer that contains the list of functions and basic block ids to
   /// selectively enable basic block sections.
   MachineFunctionPass *createBasicBlockSectionsPass(const MemoryBuffer *Buf);
 
+  /// createMachineFunctionSplitterPass - This pass splits machine functions
+  /// using profile information.
+  MachineFunctionPass *createMachineFunctionSplitterPass();
+
   /// MachineFunctionPrinter pass - This pass prints out the machine function to
   /// the given stream as a debugging tool.
   MachineFunctionPass *
   createMachineFunctionPrinterPass(raw_ostream &OS,
                                    const std::string &Banner ="");
 
   /// MIRPrinting pass - this pass prints out the LLVM IR into the given stream
   /// using the MIR serialization format.

llvm/include/llvm/InitializePasses.h

 void initializeMachineBlockPlacementStatsPass(PassRegistry&);
 void initializeMachineBranchProbabilityInfoPass(PassRegistry&);
 void initializeMachineCSEPass(PassRegistry&);
 void initializeMachineCombinerPass(PassRegistry&);
 void initializeMachineCopyPropagationPass(PassRegistry&);
 void initializeMachineDominanceFrontierPass(PassRegistry&);
 void initializeMachineDominatorTreePass(PassRegistry&);
 void initializeMachineFunctionPrinterPassPass(PassRegistry&);
+void initializeMachineFunctionSplitterPass(PassRegistry &);
 void initializeMachineLICMPass(PassRegistry&);
 void initializeMachineLoopInfoPass(PassRegistry&);
 void initializeMachineModuleInfoWrapperPassPass(PassRegistry &);
 void initializeMachineOptimizationRemarkEmitterPassPass(PassRegistry&);
 void initializeMachineOutlinerPass(PassRegistry&);
 void initializeMachinePipelinerPass(PassRegistry&);
 void initializeMachinePostDominatorTreePass(PassRegistry&);
 void initializeMachineRegionInfoPassPass(PassRegistry&);

llvm/include/llvm/Target/TargetOptions.h

enum class BasicBlockSection {
     All,    // Use Basic Block Sections for all basic blocks.  A section
             // for every basic block can significantly bloat object file sizes.
     List,   // Get list of functions & BBs from a file. Selectively enables
             // basic block sections for a subset of basic blocks which can be
             // used to control object size bloats from creating sections.
     Labels, // Do not use Basic Block Sections but label basic blocks.  This
             // is useful when associating profile counts from virtual addresses
             // to basic blocks.
+    Preset, // Similar to list but the blocks are identified by passes which
+            // seek to use Basic Block Sections, e.g. MachineFunctionSplitter.
+            // This option cannot be set via the command line.
     None    // Do not use Basic Block Sections.
   };
 
   enum class EABI {
     Unknown,
     Default, // Default means not specified
     EABI4,   // Target-specific (either 4, 5 or gnu depending on triple).
     EABI5,
TargetOptions()
           GuaranteedTailCallOpt(false), StackSymbolOrdering(true),
           EnableFastISel(false), EnableGlobalISel(false), UseInitArray(false),
           DisableIntegratedAS(false), RelaxELFRelocations(false),
           FunctionSections(false), DataSections(false),
           UniqueSectionNames(true), UniqueBasicBlockSectionNames(false),
           TrapUnreachable(false), NoTrapAfterNoreturn(false), TLSSize(0),
           EmulatedTLS(false), ExplicitEmulatedTLS(false), EnableIPRA(false),
           EmitStackSizeSection(false), EnableMachineOutliner(false),
-          SupportsDefaultOutlining(false), EmitAddrsig(false),
-          EmitCallSiteInfo(false), SupportsDebugEntryValues(false),
-          EnableDebugEntryValues(false), ForceDwarfFrameSection(false),
-          XRayOmitFunctionIndex(false),
+          EnableMachineFunctionSplitter(false), SupportsDefaultOutlining(false),
+          EmitAddrsig(false), EmitCallSiteInfo(false),
+          SupportsDebugEntryValues(false), EnableDebugEntryValues(false),
+          ForceDwarfFrameSection(false), XRayOmitFunctionIndex(false),
           FPDenormalMode(DenormalMode::IEEE, DenormalMode::IEEE) {}
 
     /// DisableFramePointerElim - This returns true if frame pointer elimination
     /// optimization should be disabled for the given machine function.
     bool DisableFramePointerElim(const MachineFunction &MF) const;
 
     /// UnsafeFPMath - This flag is enabled when the
     /// -enable-unsafe-fp-math flag is specified on the command line.  When
public:
     unsigned EnableIPRA : 1;
 
     /// Emit section containing metadata on function stack sizes.
     unsigned EmitStackSizeSection : 1;
 
     /// Enables the MachineOutliner pass.
     unsigned EnableMachineOutliner : 1;
 
+    /// Enables the MachineFunctionSplitter pass.
+    unsigned EnableMachineFunctionSplitter : 1;
+
     /// Set if the target supports default outlining behaviour.
     unsigned SupportsDefaultOutlining : 1;
 
     /// Emit address-significance table.
     unsigned EmitAddrsig : 1;
 
     /// Emit basic blocks into separate sections.
     BasicBlockSection BBSections = BasicBlockSection::None;

llvm/lib/CodeGen/BasicBlockSections.cpp

 //
 //===----------------------------------------------------------------------===//
 
 #include "llvm/ADT/Optional.h"
 #include "llvm/ADT/SmallSet.h"
 #include "llvm/ADT/SmallVector.h"
 #include "llvm/ADT/StringMap.h"
 #include "llvm/ADT/StringRef.h"
+#include "llvm/CodeGen/BasicBlockSectionUtils.h"
 #include "llvm/CodeGen/MachineFunction.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
 #include "llvm/CodeGen/MachineModuleInfo.h"
 #include "llvm/CodeGen/Passes.h"
 #include "llvm/CodeGen/TargetInstrInfo.h"
 #include "llvm/InitializePasses.h"
 #include "llvm/Support/Error.h"
 #include "llvm/Support/LineIterator.h"

 // All explicitly specified clusters of basic blocks will be ordered
 // accordingly. All non-specified BBs go into a separate "Cold" section.
 // Additionally, if exception handling landing pads end up in more than one
 // clusters, they are moved into a single "Exception" section. Eventually,
 // clusters are ordered in increasing order of their IDs, with the "Exception"
 // and "Cold" succeeding all other clusters.
 // FuncBBClusterInfo represent the cluster information for basic blocks. If this
 // is empty, it means unique sections for all basic blocks in the function.
-static bool assignSectionsAndSortBasicBlocks(
-    MachineFunction &MF,
+static void
+assignSections(MachineFunction &MF,
                const std::vector<Optional<BBClusterInfo>> &FuncBBClusterInfo) {
   assert(MF.hasBBSections() && "BB Sections is not set for function.");
   // This variable stores the section ID of the cluster containing eh_pads (if
   // all eh_pads are one cluster). If more than one cluster contain eh_pads, we
   // set it equal to ExceptionSectionID.
   Optional<MBBSectionID> EHPadsSectionID;
 
   for (auto &MBB : MF) {
     // With the 'all' option, every basic block is placed in a unique section.
assignSections(MachineFunction &MF,
   }
 
   // If EHPads are in more than one section, this places all of them in the
   // special exception section.
   if (EHPadsSectionID == MBBSectionID::ExceptionSectionID)
     for (auto &MBB : MF)
       if (MBB.isEHPad())
         MBB.setSectionID(EHPadsSectionID.getValue());
+}
 
+// This function is exposed externally by BasicBlockSectionUtils.h
+void llvm::sortBasicBlocksAndUpdateBranches(
+    MachineFunction &MF, MachineBasicBlockComparator MBBCmp) {
   SmallVector<MachineBasicBlock *, 4> PreLayoutFallThroughs(
       MF.getNumBlockIDs());
   for (auto &MBB : MF)
     PreLayoutFallThroughs[MBB.getNumber()] = MBB.getFallThrough();
 
+  MF.sort(MBBCmp);
+
+  // Set IsBeginSection and IsEndSection according to the assigned section IDs.
+  MF.assignBeginEndSections();
+
+  // After reordering basic blocks, we must update basic block branches to
+  // insert explicit fallthrough branches when required and optimize branches
+  // when possible.
+  updateBranches(MF, PreLayoutFallThroughs);
+}
+
+bool BasicBlockSections::runOnMachineFunction(MachineFunction &MF) {
+  auto BBSectionsType = MF.getTarget().getBBSectionsType();
+  assert(BBSectionsType != BasicBlockSection::None &&
+         "BB Sections not enabled!");
+  // Renumber blocks before sorting them for basic block sections.  This is
+  // useful during sorting, basic blocks in the same section will retain the
+  // default order.  This renumbering should also be done for basic block
+  // labels to match the profiles with the correct blocks.
+  MF.RenumberBlocks();
+
+  if (BBSectionsType == BasicBlockSection::Labels) {
+    MF.setBBSectionsType(BBSectionsType);
+    MF.createBBLabels();
+    return true;
+  }
+
+  std::vector<Optional<BBClusterInfo>> FuncBBClusterInfo;
+  if (BBSectionsType == BasicBlockSection::List &&
+      !getBBClusterInfoForFunction(MF, FuncAliasMap, ProgramBBClusterInfo,
+                                   FuncBBClusterInfo))
+    return true;
+  MF.setBBSectionsType(BBSectionsType);
+  MF.createBBLabels();
+  assignSections(MF, FuncBBClusterInfo);
+
   // We make sure that the cluster including the entry basic block precedes all
   // other clusters.
   auto EntryBBSectionID = MF.front().getSectionID();
 
   // Helper function for ordering BB sections as follows:
   //   * Entry section (section including the entry block).
   //   * Regular sections (in increasing order of their Number).
   //     ...
   //   * Exception section
   //   * Cold section
   auto MBBSectionOrder = [EntryBBSectionID](const MBBSectionID &LHS,
                                             const MBBSectionID &RHS) {
     // We make sure that the section containing the entry block precedes all the
     // other sections.
     if (LHS == EntryBBSectionID || RHS == EntryBBSectionID)
       return LHS == EntryBBSectionID;
     return LHS.Type == RHS.Type ? LHS.Number < RHS.Number : LHS.Type < RHS.Type;
   };
 
   // We sort all basic blocks to make sure the basic blocks of every cluster are
   // contiguous and ordered accordingly. Furthermore, clusters are ordered in
   // increasing order of their section IDs, with the exception and the
   // cold section placed at the end of the function.
-  MF.sort([&](MachineBasicBlock &X, MachineBasicBlock &Y) {
+  auto Comparator = [&](const MachineBasicBlock &X,
+                        const MachineBasicBlock &Y) {
     auto XSectionID = X.getSectionID();
     auto YSectionID = Y.getSectionID();
     if (XSectionID != YSectionID)
       return MBBSectionOrder(XSectionID, YSectionID);
     // If the two basic block are in the same section, the order is decided by
     // their position within the section.
     if (XSectionID.Type == MBBSectionID::SectionType::Default)
       return FuncBBClusterInfo[X.getNumber()]->PositionInCluster <
              FuncBBClusterInfo[Y.getNumber()]->PositionInCluster;
     return X.getNumber() < Y.getNumber();
-  });
-
-  // Set IsBeginSection and IsEndSection according to the assigned section IDs.
-  MF.assignBeginEndSections();
-
-  // After reordering basic blocks, we must update basic block branches to
-  // insert explicit fallthrough branches when required and optimize branches
-  // when possible.
-  updateBranches(MF, PreLayoutFallThroughs);
-
-  return true;
-}
-
-bool BasicBlockSections::runOnMachineFunction(MachineFunction &MF) {
-  auto BBSectionsType = MF.getTarget().getBBSectionsType();
-  assert(BBSectionsType != BasicBlockSection::None &&
-         "BB Sections not enabled!");
-  // Renumber blocks before sorting them for basic block sections.  This is
-  // useful during sorting, basic blocks in the same section will retain the
-  // default order.  This renumbering should also be done for basic block
-  // labels to match the profiles with the correct blocks.
-  MF.RenumberBlocks();
-
-  if (BBSectionsType == BasicBlockSection::Labels) {
-    MF.setBBSectionsType(BBSectionsType);
-    MF.createBBLabels();
-    return true;
-  }
+  };
 
-  std::vector<Optional<BBClusterInfo>> FuncBBClusterInfo;
-  if (BBSectionsType == BasicBlockSection::List &&
-      !getBBClusterInfoForFunction(MF, FuncAliasMap, ProgramBBClusterInfo,
-                                   FuncBBClusterInfo))
-    return true;
-  MF.setBBSectionsType(BBSectionsType);
-  MF.createBBLabels();
-  assignSectionsAndSortBasicBlocks(MF, FuncBBClusterInfo);
+  sortBasicBlocksAndUpdateBranches(MF, Comparator);
   return true;
 }
 
 // Basic Block Sections can be enabled for a subset of machine basic blocks.
 // This is done by passing a file containing names of functions for which basic
 // block sections are desired.  Additionally, machine basic block ids of the
 // functions can also be specified for a finer granularity. Moreover, a cluster
 // of basic blocks could be assigned to the same section.

llvm/lib/CodeGen/CMakeLists.txt

add_llvm_component_library(LLVMCodeGen
   MachineCSE.cpp
   MachineDebugify.cpp
   MachineDominanceFrontier.cpp
   MachineDominators.cpp
   MachineFrameInfo.cpp
   MachineFunction.cpp
   MachineFunctionPass.cpp
   MachineFunctionPrinterPass.cpp
+  MachineFunctionSplitter.cpp
   MachineInstrBundle.cpp
   MachineInstr.cpp
   MachineLICM.cpp
   MachineLoopInfo.cpp
   MachineLoopUtils.cpp
   MachineModuleInfo.cpp
   MachineModuleInfoImpls.cpp
   MachineOperand.cpp

llvm/lib/CodeGen/CommandFlags.cpp

 CGOPT(bool, EmulatedTLS)
 CGOPT(bool, UniqueSectionNames)
 CGOPT(bool, UniqueBasicBlockSectionNames)
 CGOPT(EABI, EABIVersion)
 CGOPT(DebuggerKind, DebuggerTuningOpt)
 CGOPT(bool, EnableStackSizeSection)
 CGOPT(bool, EnableAddrsig)
 CGOPT(bool, EmitCallSiteInfo)
+CGOPT(bool, EnableMachineFunctionSplitter)
 CGOPT(bool, EnableDebugEntryValues)
 CGOPT(bool, ForceDwarfFrameSection)
 CGOPT(bool, XRayOmitFunctionIndex)
 
 codegen::RegisterCodeGenFlags::RegisterCodeGenFlags() {
 #define CGBINDOPT(NAME)                                                        \
   do {                                                                         \
     NAME##View = std::addressof(NAME);                                         \
#define CGBINDOPT(NAME)                                                        \
   CGBINDOPT(EmitCallSiteInfo);
 
   static cl::opt<bool> EnableDebugEntryValues(
       "debug-entry-values",
       cl::desc("Enable debug info for the debug entry values."),
       cl::init(false));
   CGBINDOPT(EnableDebugEntryValues);
 
+  static cl::opt<bool> EnableMachineFunctionSplitter(
+      "split-machine-functions",
+      cl::desc("Split out cold basic blocks from machine functions based on "
+               "profile information"),
+      cl::init(false));
+  CGBINDOPT(EnableMachineFunctionSplitter);
+
   static cl::opt<bool> ForceDwarfFrameSection(
       "force-dwarf-frame-section",
       cl::desc("Always emit a debug frame section."), cl::init(false));
   CGBINDOPT(ForceDwarfFrameSection);
 
   static cl::opt<bool> XRayOmitFunctionIndex(
       "no-xray-index", cl::desc("Don't emit xray_fn_idx section"),
       cl::init(false));
TargetOptions codegen::InitTargetOptionsFromCodeGenFlags() {
   Options.BBSections = getBBSectionsMode(Options);
   Options.UniqueSectionNames = getUniqueSectionNames();
   Options.UniqueBasicBlockSectionNames = getUniqueBasicBlockSectionNames();
   Options.TLSSize = getTLSSize();
   Options.EmulatedTLS = getEmulatedTLS();
   Options.ExplicitEmulatedTLS = EmulatedTLSView->getNumOccurrences() > 0;
   Options.ExceptionModel = getExceptionModel();
   Options.EmitStackSizeSection = getEnableStackSizeSection();
+  Options.EnableMachineFunctionSplitter = getEnableMachineFunctionSplitter();
   Options.EmitAddrsig = getEnableAddrsig();
   Options.EmitCallSiteInfo = getEmitCallSiteInfo();
   Options.EnableDebugEntryValues = getEnableDebugEntryValues();
   Options.ForceDwarfFrameSection = getForceDwarfFrameSection();
   Options.XRayOmitFunctionIndex = getXRayOmitFunctionIndex();
 
   Options.MCOptions = mc::InitMCTargetOptionsFromFlags();
 

llvm/lib/CodeGen/MachineFunctionSplitter.cpp

+//===-- MachineFunctionSplitter.cpp - Split machine functions //-----------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// \file
+// Uses profile information to split out cold blocks.
+//
+// This pass splits out cold machine basic blocks from the parent function. This
+// implementation leverages the basic block section framework. Blocks marked
+// cold by this pass are grouped together in a separate section prefixed with
+// ".text.unlikely.*". The linker can then group these together as a cold
+// section. The split part of the function is a contiguous region identified by
+// the symbol "foo.cold". Grouping all cold blocks across functions together
+// decreases fragmentation and improves icache and itlb utilization. Note that
+// the overall changes to the binary size are negligible; only a small number of
+// additional jump instructions may be introduced.
+//
+// For the original, RFC of this pass please see
+// https://groups.google.com/d/msg/llvm-dev/RUegaMg-iqc/wFAVxa6fCgAJ 
+//===----------------------------------------------------------------------===//
+
+#include "llvm/ADT/Statistic.h"
+#include "llvm/Analysis/ProfileSummaryInfo.h"
+#include "llvm/CodeGen/BasicBlockSectionUtils.h"
+#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineBlockFrequencyInfo.h"
+#include "llvm/CodeGen/MachineFunction.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineModuleInfo.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/IR/Function.h"
+#include "llvm/IR/Module.h"
+#include "llvm/InitializePasses.h"
+#include "llvm/Support/CommandLine.h"
+
+using namespace llvm;
+
+static cl::opt<unsigned>
+    PercentileCutoff("mfs-psi-cutoff",
+                     cl::desc("Percentile profile summary cutoff used to "
+                              "determine cold blocks. Unused if set to zero."),
+                     cl::init(0), cl::Hidden);
+
+static cl::opt<unsigned> ColdCountThreshold(
+    "mfs-count-threshold",
+    cl::desc(
+        "Minimum number of times a block must be executed to be retained."),
+    cl::init(1), cl::Hidden);
+
+static cl::list<std::string> ExcludedSections(
+    "mfs-excluded-sections",
+    cl::desc(
+        "Comma separated list of section names to exclude from splitting."),
+    cl::Hidden);
+
+namespace {
+
+class MachineFunctionSplitter : public MachineFunctionPass {
+public:
+  static char ID;
+  MachineFunctionSplitter() : MachineFunctionPass(ID) {
+    initializeMachineFunctionSplitterPass(*PassRegistry::getPassRegistry());
+  }
+
+  StringRef getPassName() const override {
+    return "Machine Function Splitter Transformation";
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override;
+
+  bool runOnMachineFunction(MachineFunction &F) override;
+};
+} // end anonymous namespace
+
+static bool isColdBlock(MachineBasicBlock &MBB,
+                        const MachineBlockFrequencyInfo *MBFI,
+                        ProfileSummaryInfo *PSI) {
+  Optional<uint64_t> Count = MBFI->getBlockProfileCount(&MBB);
+  if (!Count.hasValue()) {
+    return true;
+  }
+
+  if (PercentileCutoff > 0) {
+    return PSI->isColdCountNthPercentile(PercentileCutoff, *Count);
+  }
+  return (*Count < ColdCountThreshold);
+}
+
+bool MachineFunctionSplitter::runOnMachineFunction(MachineFunction &MF) {
+  // FIXME: We only target functions with profile data. Static information may
+  // also be considered but we don't see performance improvements yet.
+  if (!MF.getFunction().hasProfileData()) {
+    return false;
+  }
+
+  // We don't want to proceed further for cold functions
+  // or functions of unknown hotness. Lukewarm functions have no prefix.
+  Optional<StringRef> SectionPrefix = MF.getFunction().getSectionPrefix();
+  if (SectionPrefix.hasValue() &&
+      (SectionPrefix.getValue().equals(".unlikely") ||
+       SectionPrefix.getValue().equals(".unknown"))) {
+    return false;
+  }
+
+  // Allow certain sections specified by name to be excluded from splitting.
+  // Commonly this will be empty and will only show up for section names
+  // specified by __attribute__((section(name)).
+  StringRef SectionName = MF.getFunction().getSection();
+  for (const auto &ES : ExcludedSections) {
+    if (SectionName.equals(ES)) {
+      return false;
+    }
+  }
+
+  MF.RenumberBlocks();
+  MF.setBBSectionsType(BasicBlockSection::Preset);
+  auto *MBFI = &getAnalysis<MachineBlockFrequencyInfo>();
+  auto *PSI = &getAnalysis<ProfileSummaryInfoWrapperPass>().getPSI();
+
+  for (auto &MBB : MF) {
+    // We retain the entry block and conservatively keep all landing pad blocks
+    // as part of the original function.
+    if ((MBB.pred_empty() || MBB.isEHPad()))
+      continue;
+    if (isColdBlock(MBB, MBFI, PSI)) {
+      MBB.setSectionID(MBBSectionID::ColdSectionID);
+    }
+  }
+
+  auto Comparator = [](const MachineBasicBlock &X, const MachineBasicBlock &Y) {
+    return X.getSectionID().Type < Y.getSectionID().Type;
+  };
+  llvm::sortBasicBlocksAndUpdateBranches(MF, Comparator);
+
+  return true;
+}
+
+void MachineFunctionSplitter::getAnalysisUsage(AnalysisUsage &AU) const {
+  AU.addRequired<MachineModuleInfoWrapperPass>();
+  AU.addRequired<MachineBlockFrequencyInfo>();
+  AU.addRequired<ProfileSummaryInfoWrapperPass>();
+}
+
+char MachineFunctionSplitter::ID = 0;
+INITIALIZE_PASS(MachineFunctionSplitter, "machine-function-splitter",
+                "Split machine functions using profile information", false,
+                false)
+
+MachineFunctionPass *llvm::createMachineFunctionSplitterPass() {
+  return new MachineFunctionSplitter();
+}

llvm/lib/CodeGen/TargetPassConfig.cpp

StopAfterOpt(StringRef(StopAfterOptName),
                  cl::desc("Stop compilation after a specific pass"),
                  cl::value_desc("pass-name"), cl::init(""), cl::Hidden);
 
 static cl::opt<std::string>
     StopBeforeOpt(StringRef(StopBeforeOptName),
                   cl::desc("Stop compilation before a specific pass"),
                   cl::value_desc("pass-name"), cl::init(""), cl::Hidden);
 
+/// Enable the machine function splitter pass.
+/// FIXME: Remove this once clang option for this feature has been added.
+static cl::opt<bool> EnableMachineFunctionSplitter(
+    "enable-split-machine-functions", cl::Hidden,
+    cl::desc("Split out cold blocks from machine functions based on profile "
+             "information."));
+
 /// Allow standard passes to be disabled by command line options. This supports
 /// simple binary flags that either suppress the pass or do nothing.
 /// i.e. -disable-mypass=false has no effect.
 /// These should be converted to boolOrDefault in order to use applyOverride.
 static IdentifyingPassPtr applyDisable(IdentifyingPassPtr PassID,
                                        bool Override) {
   if (Override)
     return IdentifyingPassPtr();
if (TM->Options.EnableMachineOutliner && getOptLevel() != CodeGenOpt::None &&
       EnableMachineOutliner != NeverOutline) {
     bool RunOnAllFunctions = (EnableMachineOutliner == AlwaysOutline);
     bool AddOutliner = RunOnAllFunctions ||
                        TM->Options.SupportsDefaultOutlining;
     if (AddOutliner)
       addPass(createMachineOutlinerPass(RunOnAllFunctions));
   }
 
-  if (TM->getBBSectionsType() != llvm::BasicBlockSection::None)
+  // Machine function splitter uses the basic block sections feature. Both
+  // cannot be enabled at the same time.
+  if (TM->Options.EnableMachineFunctionSplitter ||
+      EnableMachineFunctionSplitter) {
+    addPass(createMachineFunctionSplitterPass());
+  } else if (TM->getBBSectionsType() != llvm::BasicBlockSection::None) {
     addPass(llvm::createBasicBlockSectionsPass(TM->getBBSectionsFuncListBuf()));
+  }
 
   // Add passes that directly emit MI after all other MI passes.
   addPreEmitPass2();
 
   AddingMachinePasses = false;
 }
 
 /// Add passes that optimize machine instructions in SSA form.

llvm/test/CodeGen/X86/machine-function-splitter.ll

+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions | FileCheck %s -check-prefix=MFS-DEFAULTS
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions -mfs-count-threshold=2000 | FileCheck %s --dump-input=always -check-prefix=MFS-OPTS1
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions -mfs-psi-cutoff=950000 | FileCheck %s -check-prefix=MFS-OPTS2
+; RUN: llc < %s -mtriple=x86_64-unknown-linux-gnu -split-machine-functions -mfs-excluded-sections="nosplit" | FileCheck %s -check-prefix=MFS-OPTS3
+
+define void @foo1(i1 zeroext %0) nounwind !prof !14 !section_prefix !15 {
+;; Check that cold block is moved to .text.unlikely.
+; MFS-DEFAULTS-LABEL: foo1
+; MFS-DEFAULTS:       .section        .text.unlikely.foo1
+; MFS-DEFAULTS-NEXT:  foo1.cold:
+; MFS-DEFAULTS-NOT:   callq   bar
+; MFS-DEFAULTS-NEXT:  callq   baz
+  br i1 %0, label %2, label %4, !prof !17
+
+2:                                                ; preds = %1
+  %3 = call i32 @bar()
+  br label %6
+
+4:                                                ; preds = %1
+  %5 = call i32 @baz()
+  br label %6
+
+6:                                                ; preds = %4, %2
+  %7 = tail call i32 @qux()
+  ret void
+}
+
+define void @foo2(i1 zeroext %0) nounwind !prof !23 !section_prefix !16 {
+;; Check that function marked unlikely is not split.
+; MFS-DEFAULTS-LABEL: foo2
+; MFS-DEFAULTS-NOT:   foo2.cold:
+  br i1 %0, label %2, label %4, !prof !17
+
+2:                                                ; preds = %1
+  %3 = call i32 @bar()
+  br label %6
+
+4:                                                ; preds = %1
+  %5 = call i32 @baz()
+  br label %6
+
+6:                                                ; preds = %4, %2
+  %7 = tail call i32 @qux()
+  ret void
+}
+
+define void @foo3(i1 zeroext %0) nounwind !section_prefix !15 {
+;; Check that function without profile data is not split.
+; MFS-DEFAULTS-LABEL: foo3
+; MFS-DEFAULTS-NOT:   foo3.cold:
+  br i1 %0, label %2, label %4
+
+2:                                                ; preds = %1
+  %3 = call i32 @bar()
+  br label %6
+
+4:                                                ; preds = %1
+  %5 = call i32 @baz()
+  br label %6
+
+6:                                                ; preds = %4, %2
+  %7 = tail call i32 @qux()
+  ret void
+}
+
+define void @foo4(i1 zeroext %0, i1 zeroext %1) nounwind !prof !20 {
+;; Check that count threshold works.
+; MFS-OPTS1-LABEL: foo4
+; MFS-OPTS1:       .section        .text.unlikely.foo4
+; MFS-OPTS1-NEXT:  foo4.cold:
+; MFS-OPTS1-NOT:   callq    bar
+; MFS-OPTS1-NOT:   callq    baz
+; MFS-OPTS1-NEXT:  callq    bam
+  br i1 %0, label %3, label %7, !prof !18
+
+3:
+  %4 = call i32 @bar()
+  br label %7
+
+5:
+  %6 = call i32 @baz()
+  br label %7
+
+7:
+  br i1 %1, label %8, label %10, !prof !19
+
+8:
+  %9 = call i32 @bam()
+  br label %12
+
+10:
+  %11 = call i32 @baz()
+  br label %12
+
+12:
+  %13 = tail call i32 @qux()
+  ret void
+}
+
+define void @foo5(i1 zeroext %0, i1 zeroext %1) nounwind !prof !20 {
+;; Check that profile summary info cutoff works.
+; MFS-OPTS2-LABEL: foo5
+; MFS-OPTS2:       .section        .text.unlikely.foo5
+; MFS-OPTS2-NEXT:       foo5.cold:
+; MFS-OPTS2-NOT:   callq    bar
+; MFS-OPTS2-NOT:   callq    baz
+; MFS-OPTS2-NEXT:  callq    bam
+  br i1 %0, label %3, label %7, !prof !21
+
+3:
+  %4 = call i32 @bar()
+  br label %7
+
+5:
+  %6 = call i32 @baz()
+  br label %7
+
+7:
+  br i1 %1, label %8, label %10, !prof !22
+
+8:
+  %9 = call i32 @bam()
+  br label %12
+
+10:
+  %11 = call i32 @baz()
+  br label %12
+
+12:
+  %13 = call i32 @qux()
+  ret void
+}
+
+define void @foo6(i1 zeroext %0) nounwind section "nosplit" !prof !14 {
+;; Check that function in excluded section is not split.
+; MFS-OPTS3-LABEL: foo6
+; MFS-OPTS3-NOT:   foo6.cold:
+  br i1 %0, label %2, label %4, !prof !17
+
+2:                                                ; preds = %1
+  %3 = call i32 @bar()
+  br label %6
+
+4:                                                ; preds = %1
+  %5 = call i32 @baz()
+  br label %6
+
+6:                                                ; preds = %4, %2
+  %7 = tail call i32 @qux()
+  ret void
+}
+
+declare i32 @bar()
+declare i32 @baz()
+declare i32 @bam()
+declare i32 @qux()
+
+!llvm.module.flags = !{!0}
+!0 = !{i32 1, !"ProfileSummary", !1}
+!1 = !{!2, !3, !4, !5, !6, !7, !8, !9}
+!2 = !{!"ProfileFormat", !"InstrProf"}
+!3 = !{!"TotalCount", i64 10000}
+!4 = !{!"MaxCount", i64 10}
+!5 = !{!"MaxInternalCount", i64 1}
+!6 = !{!"MaxFunctionCount", i64 1000}
+!7 = !{!"NumCounts", i64 3}
+!8 = !{!"NumFunctions", i64 5}
+!9 = !{!"DetailedSummary", !10}
+!10 = !{!11, !12, !13}
+!11 = !{i32 10000, i64 100, i32 1}
+!12 = !{i32 999900, i64 100, i32 1}
+!13 = !{i32 999999, i64 1, i32 2}
+!14 = !{!"function_entry_count", i64 7000}
+!15 = !{!"function_section_prefix", !".hot"}
+!16 = !{!"function_section_prefix", !".unlikely"}
+!17 = !{!"branch_weights", i32 7000, i32 0}
+!18 = !{!"branch_weights", i32 3000, i32 4000}
+!19 = !{!"branch_weights", i32 1000, i32 6000}
+!20 = !{!"function_entry_count", i64 10000}
+!21 = !{!"branch_weights", i32 6000, i32 4000}
+!22 = !{!"branch_weights", i32 80, i32 9920}
+!23 = !{!"function_entry_count", i64 7}