[IPRA] Interprocedural Register Allocation - Analysis Passes

[vivekvpandya]

Summary

 This commit addes pass to change CodeGen order to be bottom up order on Call Graph. The patch for the same has been provided by dear Mehdi Amini.
 
	This commit added required analysis passes for IPRA.
 
 1) lib/CodeGen/RegUsageInfoCollector.cpp (MachineFunction Pass) to create RegMask based on actual register usage, scheduled at POST-RA
 2) lib/CodeGen/RegisterUsageInfo.cpp (Immutable Pass) to store RegMask details for functions which has been processed in bottom up order on Call Graph, Scheduled before RegAllocation.
   
 This commit also adds a method setRegMask() to MachineOperand class.
 This commit also adds command line option to enable IPRA -enable-ipra and debug type "ip-regalloc".
 Changes to adher to coding standards, correction for ownership of RegMask vector in related functions.

[mehdi_amini]

include/llvm/CodeGen/MachineOperand.h

537

Take a const ptr.

include/llvm/CodeGen/PhysicalRegisterUsageInfo.h

36 ↗

(On Diff #58887)

You need to do the same in RegUsageInfoPropagationPass

43 ↗

(On Diff #58887)

Take RegMasks by value here.

45 ↗

(On Diff #58887)

const std::vector<uint32_t> *getRegUsageInfo(StringRef MFName);

lib/CodeGen/PhysicalRegisterUsageInfo.cpp

32 ↗

(On Diff #58887)

RegMasks[MFName] = std::move(RegMask);

lib/CodeGen/RegUsageInfoCollector.cpp

100

PRUI->storeUpdateRegUsageInfo(MF.getName(), std::move(RegMask));

lib/CodeGen/TargetPassConfig.cpp

587

remove

603

Remove this, it is useless.

lib/Target/X86/X86RegUsageInfoPropagate.cpp

56 ↗

(On Diff #58887)

Take a const ptr

92 ↗

(On Diff #58887)

C++11 for-range:

for(auto &MBB : MF) {
  for(auto &MI : MBB) {

113 ↗

(On Diff #58887)

auto updateRegMask = [&](StringRef FuncName) {
  const auto *RegMask = PRUI->getRegUsageInfo(FuncName);
  if (RegMask) { // else skip optimization
    setRegMask(MI, &(*RegMask)[0]);
    changed = true;
  }
};
MachineOperand &Operand = MI.getOperand(0);
if (Operand.isGlobal()) {
  updateRegMask(Operand.getGlobal()->getName());
} else if (Operand.isSymbol()) {
  updateRegMask(Operand.getGlobal()->getName());
}

lib/Target/X86/X86TargetMachine.cpp

316 ↗

(On Diff #58887)

This is not the right place, it should be added right after createX86ISelDag

[vivekvpandya]

Changes to adher to coding standards, correction for ownership of RegMask vector in related functions.

[mehdi_amini]

include/llvm/CodeGen/PhysicalRegisterUsageInfo.h

28 ↗

(On Diff #58917)

This is not clang-formatted, can you please format it?

[vivekvpandya]

Files have been formatted with clang-format.

[mehdi_amini]

include/llvm/CodeGen/MachineOperand.h

147

It is not good practice to clang-format the whole file as part of your patch. You should only format the patch of the file you changed (git clang-format handles it automatically, some text editor as well, or you can give a range on the command line, or even copy paster on stdin the range you want to format).

[mehdi_amini]

lib/CodeGen/TargetPassConfig.cpp

502

Looks like we'll need a home for this pass, can't really leave it there. I'm not sure where to put it yet.

517

Are these dependency required?

[vivekvpandya]

Fix that makes RegUsageInfoCollector Calling Convention aware so that it does not mark CalleeSaved register as clobbered if it is not used.

[qcolombet]

Hi Vivek,

A couple of minor comments, the main logic looks reasonable.

Cheers,
-Quentin

include/llvm/CodeGen/MachineOperand.h

537

I would say in a comment to double check MachineOperand::CreateRegMask for the characteristic of RegMaskPtr.

The reason why I think it is important is because that method clearly states the expected life time of the pointer and if we get it wrong we may end with nasty bug. Therefore, we should do our best to have the API clearly documented.

include/llvm/CodeGen/Passes.h

361

Typo: machine

include/llvm/CodeGen/PhysicalRegisterUsageInfo.h

1 ↗

(On Diff #59113)

.cpp => .h

2 ↗

(On Diff #59113)

Wrapped line.

51 ↗

(On Diff #59113)

Period.

51 ↗

(On Diff #59113)

Please use doxygen style comment: ///

52 ↗

(On Diff #59113)

Typo: function

53 ↗

(On Diff #59113)

Period

include/llvm/InitializePasses.h

343

I don’t know why this is not the case for last one, but we usually sort the initializers by name.

lib/CodeGen/CMakeLists.txt

103

Looks like this patch could be split:

• One patch for RegisterUsageInfo.
• One patch for InfoCollector.

lib/CodeGen/PhysicalRegisterUsageInfo.cpp

27 ↗

(On Diff #59113)

ip-regalloc

45 ↗

(On Diff #59113)

Remove the else per the coding convention.

lib/CodeGen/RegUsageInfoCollector.cpp

17

Wrapping is strange here.
We split in the middle of a sentence without hitting the 80-col limit.

82

Call getPassName.

88

That is a bit strange to state the comment like that.
What about: Compute the size of the bit vector to represent all the registers. The bit vector is broken into 32-bit chunks, thus takes the ceil of the number of registers divided by 32 for the size.

95

&& instead of nested ifs.

96

Period.

98

Encapsulate that loop into a setRegister thing.

107

Capital letter at the beginning, period at the end of the sentence.

lib/Target/X86/X86RegUsageInfoPropagate.cpp

2 ↗

(On Diff #59113)

Why is that pass X86 specific?
It should be generic, right?

12 ↗

(On Diff #59113)

Use doxygen comments: ///

16 ↗

(On Diff #59113)

queries

[mehdi_amini]

Somehow I had unsubmitted comments on phab, some are still relevant.

lib/CodeGen/PhysicalRegisterUsageInfo.cpp

2 ↗

(On Diff #59113)

This should be on one line (and fits within 80 chars).

lib/CodeGen/RegUsageInfoCollector.cpp

3

same

38

Doesn't this need to be in a header and be called somewhere?
It's not clear to me how is this pass registered?

110

Why not initializing RegMask with 0xFFFFFFFF in resize() and having the loop setting the bit to zero instead?

116

comment

lib/CodeGen/TargetPassConfig.cpp

503

No capital letters here.

584

s/register/registers/

585

Comment could be simply: Collect register usage information and produce a register mask of clobbered registers, to be used to optimize call sites.

lib/Target/X86/X86.h

86 ↗

(On Diff #59113)

s/call site/call sites/

lib/Target/X86/X86RegUsageInfoPropagate.cpp

2 ↗

(On Diff #59113)

When we first looked at it, is seemed to me that we would have to switch over the various possible calls. But it was a mistake on my side, and now that it is implemented it does not seem to need any target specific bit indeed.

[MatzeB]

The general approach seems fine at a first glance. I haven't reviewed this in-depth for correctness yet though, as I first have a bunch of nit-picks and coding style issues:

include/llvm/CodeGen/MachineOperand.h

537

+1

include/llvm/CodeGen/PhysicalRegisterUsageInfo.h

1 ↗

(On Diff #59113)

Does all of this need to be exposed in a public header anyway? Wouldn't it be enough to have the initialization stuff in Passes.h and keep the actual pass definitions private to the .cpp file in this case?

(if we decide no, then I'd rename this to something shorter like RegisterUsageInfo...)

11–17 ↗

(On Diff #59113)

Should use doxygen style comment here (see coding convention). It may also be wise not to mention specific filenames here as those could be moved or renamed in the future and people generally do not catch comment like this when doing so.

54 ↗

(On Diff #59113)

@Mehdi: I've seen you arguing for a stringmap here before, but I would argue that a reference to a compiler internal object would be more stable here, avoid string comparisons and also work with unnamed functions (not sure if that actually allowed in llvm, but I could see this happening for JITs...).

As to what object to use: MachineFunction* would be nice but I am not sure that would well as long as the MachineFunction gets created in the MachineFunctionAnalysisPass. So maybe tie it to the Functions GlobalObject for now and change it to MachineFunction* later when we switch to the new pass manager and hopefully can deal with MachineFunctions as a first-class object rather than an analysis.

59 ↗

(On Diff #59113)

Check your editor settings so the newline at the end of file is added.

lib/CodeGen/PhysicalRegisterUsageInfo.cpp

1–2 ↗

(On Diff #59113)

strange linebreak. I think in .cpp we also don't need those magic markers for the emacs users (my understanding is that they are just used in the .h files so emacs knows they contain C++ code and not just C code).

11–17 ↗

(On Diff #59113)

Use doxygen comments.

lib/CodeGen/RegUsageInfoCollector.cpp

2–3

strange linebreak.

12–20

doxygen comments.

56

add "end of anonymous namespace" comment (see coding conventions).

61

It is clear that this is about a pass so no need for the " Pass" suffix in the explanation.

77

This is always false anyway, just use return false; instead of a variable.

88

We could make the code here and in MachineOperand more robust by having a "typedef uint32_t RegMaskType" and then using sizeof(RegMaskType) * CHAR_BIT instead of hardcoding 32...
Though as that also hits existing code in MachineOperand a separate patch would be warranted.

89

how about uint32_t RegMask[regMaskSize] instead of using a std::vector here so we get a stack allocation instead of an unnecessary heap allocation of the vector?

93

We tend to introduce a new variable (like PRegE = TRI->getNumRegs()) in loops like this and compare with it to avoid getNumRegs() getting called in every iteration of the loop (see coding conventions).

112

No space after the *

lib/CodeGen/TargetPassConfig.cpp

117

I feel like this comment is just stating the very obvious and can be left out.

119

I think we can should add cl::Hidden for now until this is proven and stable.

502

Looks like we'll need a home for this pass, can't really leave it there. I'm not sure where to put it yet.

Just put the pass into CallGraphSCCPass.h as it is not specific to codegen (just happens to be used
there)?

lib/Target/X86/X86RegUsageInfoPropagate.cpp

93 ↗

(On Diff #59113)

for(auto &MBB : MF) {
for(auto &MI : MBB) {

Please do not use auto in contexts where it is not obvious what type it represents. It is friendlier for the readers to use MachineBasicBlock &MBB and MachineInstr &MI here.

101 ↗

(On Diff #59113)

We start local variables with uppercase letters (coding convention).

103–104 ↗

(On Diff #59113)

same comment as above.

114 ↗

(On Diff #59113)

This does not seem complicated enough to me to warrant a lambda. What about:

const char *FuncName = nullptr;
if (Operand.isGlobal()) {
  Name = Operand.getGlobal()->getName();
} else if (Operand.isSymbol()) {
  Name = Operand.getGlobal()->getName();
}
if (FuncName != nullptr && RegMask) {
  setRegMask(MI, &(*RegMask)[0]);
  changed = true;
}

128–129 ↗

(On Diff #59113)

odd line break (clang-format may be confused because of the macro...)

131–132 ↗

(On Diff #59113)

`DEBUG(dbgs() << MI << '\n');

[MatzeB]

@Mehdi: explicit ping so you do not miss my comment about the register mask map modeling among all the nitpicks.

[mehdi_amini]

include/llvm/CodeGen/PhysicalRegisterUsageInfo.h

54 ↗

(On Diff #59113)

I pointed a StringMap in order to not have the backend depends on the IR, now if you and Quentin thinks it is fine to go through a GlobalVariable *, I don't have strong feeling about it.
(and yes, unnamed global function are allowed)

[mehdi_amini]

lib/Target/X86/X86RegUsageInfoPropagate.cpp

93 ↗

(On Diff #59113)

I feel that it is clearly obvious that MI is a MachineInstr when reading for(auto &MI : MBB) {, I'll be explicit if it was something else. But I won't pick this fight :)

114 ↗

(On Diff #59113)

I guess it is a matter of personal style, I'd try to avoid this construct when possible, but I don't really care here.

[llvm-commits]

I know some of the issues are more a matter of personal preference than there being one true way. I just felt I should point out how I feel about them before people start writing them down as the one true way in the coding conventions :) I will of course not block the patches if we end up with a different style here.

[hfinkel]

lib/CodeGen/PhysicalRegisterUsageInfo.cpp

11 ↗

(On Diff #59113)

No need to duplicate this comment in both the source file and the header.

lib/CodeGen/RegUsageInfoCollector.cpp

89

regMaskSize is not a constant, and we can't use VLAs. We could use SmallVector with a reasonable default, however.

lib/CodeGen/TargetPassConfig.cpp

583

Why here? It seems much too early. Backends can use the register scavenger to use otherwise-unused registers until the very end. I think this needs to be after the call below to:

addPreEmitPass();

if not at the very end.

lib/Target/X86/X86TargetMachine.cpp

283 ↗

(On Diff #59113)

Does this need to be here so that targets can opt-in? Could we reasonably schedule this early in TargetPassConfig::addMachinePasses?

[hfinkel]

include/llvm/CodeGen/PhysicalRegisterUsageInfo.h

10–11 ↗

(On Diff #59113)

How about:

// This pass is required to take advantage of the interprocedural-register-allocation infrastructure.

15 ↗

(On Diff #59113)

If you change from using a StringMap, update this comment.

54 ↗

(On Diff #59113)

I'd also prefer a GlobalVariable * here, as that currently establishes our canonical function identify. If we work to further separate the MI level from the IR level, I suspect we'd establish some other function object which would have pointer-based identify, and we can always then update this code to use that instead.

[mehdi_amini]

include/llvm/CodeGen/PhysicalRegisterUsageInfo.h

54 ↗

(On Diff #59113)

Since MatzeB commented about the allocation in std::vector in the MachineFunctionPass, it seems equally important to override doInitialization(Module &M) in order to reserve the size of the map with the number of functions defined in the module, and doFinalization to clear the map.
(clearing the map is actually a correctness issue)

lib/CodeGen/RegUsageInfoCollector.cpp

89

What is regMaskSize on ARM and X86?
Before moving on with stack allocation here, I think we have to consider that:

• First this is ran once per function, so having one malloc per function during codegen does not make it an expensive analysis.
• Second the vector will be moved in the immutable pass map. So having a SmallVector makes it less inefficient to move and store (we may not want to have a DenseMap anymore, and thus we'd have to make an extra malloc there!).

[vivekvpandya]

include/llvm/CodeGen/MachineOperand.h

147

I will consider it, do we required to revert those changes ? If yes then please suggest easy way, I am using git.

include/llvm/CodeGen/PhysicalRegisterUsageInfo.h

29 ↗

(On Diff #58934)

Actually I run git clang-format , but that was after committing my changes so it did not changed any thing. I have applied clang-format on each file.

lib/CodeGen/CMakeLists.txt

103

@qcolombet Could you please explain why is that required?

lib/CodeGen/PhysicalRegisterUsageInfo.cpp

1–2 ↗

(On Diff #59113)

oh yes this is already mentioned in coding standards but I really didn't check for it.
Sorry.

lib/CodeGen/RegUsageInfoCollector.cpp

38

Isn't this will be called by method generated due to macroINITIALIZE_PASS_BEGIN ?

lib/CodeGen/TargetPassConfig.cpp

502

Have thought of any thing for this?

517

I believe CallGraphWrapperPass is required, though I have not look at this code closely as it was given by you :D

517

No these dependencies are not required, I am yet to find why not technically but I just removed it and compiled llvm again and things are working.

lib/Target/X86/X86RegUsageInfoPropagate.cpp

101 ↗

(On Diff #59113)

I thought it is not for primitive type. I will change it.

[mehdi_amini]

include/llvm/CodeGen/PhysicalRegisterUsageInfo.h

30 ↗

(On Diff #59113)

git clang-format takes a commit range, so you can format patches that are already committed.
(i.e, git clang-format HEAD~ formats the last commit, you can just amend it afterward)

lib/CodeGen/RegUsageInfoCollector.cpp

38

If there is something in the macro expansion that makes you think it is called, please elaborate.

[vivekvpandya]

lib/CodeGen/RegUsageInfoCollector.cpp

38

No, I look carefully at that macro it provides definition for initialize.. method but yes some where we need to call that method. Also I haven't mentioned this pass as dependency for any other pass with INITIALIZE_PASS_DEPENDENCY other wise it can call that function. Now I am also having same question as you.

[vivekvpandya]

lib/Target/X86/X86TargetMachine.cpp

283 ↗

(On Diff #59113)

@hfinkel X86RegUsageInfoPropagate.cpp will be generic pass soon then I will do this change too.

[vivekvpandya]

Correction in typos , more adhering to coding standards.
PhysicalRegisterUsageInfo.cpp -> RegisterUsageInfo.cpp
DenseMap of GlobelVariable * to RegMask is used instead of StringMap.
DummyCGSCCPass moved to CallGraphSCCPass.h and .cpp

[mehdi_amini]

We're getting there, the patch is looking really good now! :)

Still a little bit of work ahead, see inline comments.

include/llvm/Analysis/CallGraphSCCPass.h

115–130

Describe the class with a doxygen comment.

include/llvm/CodeGen/MachineOperand.h

537

Do not repeat the name of the function in the doxygen comment (other places that do it are outdated).

include/llvm/CodeGen/Passes.h

366

s/used/preserved/

include/llvm/CodeGen/RegisterUsageInfo.h

14

You shouldn't introduce implementation details in high-level description.

lib/CodeGen/CMakeLists.txt

103–105

It is a good practice to decoupled software component. Having separate patches helps to make sure we indeed have correctly separated the components.
It also forces to make sure the component are individually testable, and make sure we actually test them.

So I agree with Quentin on the principle, and I think it is also a good exercise for you to split the patch an submit the analysis alone and tested. Have a look at "CostModelAnalysis::print()" in lib/Analysis/CostModel.cpp and see how it is tested in test/Analysis/CostModel/X86/cast.ll

lib/CodeGen/RegUsageInfoCollector.cpp

12

Wrapping

lib/CodeGen/TargetPassConfig.cpp

119

Why "compile time"? I'd just write "Enable inter-procedural register allocation"

516

You forgot to remove this hunk

[vivekvpandya]

lib/CodeGen/CMakeLists.txt

103–105

@mehdi_amini I understand what you explain above but here I think RegisterUsageInfo is not tastable alone because it just holds RegMasks, RegisterInfoCollector is a trigger to IP regalloc and both of them can be tested together. Also to test X86RegUsageInfoPropagate it requires both of the above mentioned passes. But we can separate patches for RegisterUsageInfo + InfoCollector and X86RegUsageInfoPropagate ( condition that first patch is required to test second one) .
Is there any better plan in your mind?

[mehdi_amini]

lib/CodeGen/CMakeLists.txt

103–105

Splitting in two is what I had in mind: the analysis part on one side, the transformation part on another.

[vivekvpandya]

lib/CodeGen/CMakeLists.txt

103–105

Just to make sure RegisterUsageInfo.cpp and RegUsageInfoCollector.cpp both are part of analysis so there is not need of separate patch for them but I will separate changes related to X86RegUsageInfoPropagate.cpp in other patch.

[vivekvpandya]

@mehdi_amini I have made all suggested changes , I am writing a simple test for RegUsageInfoCollector so that next review you can also look at that. Thanks for your patience.

[vivekvpandya]

patch is splited into analysis and optimization pass. Simple test cases has been added.

[mehdi_amini]

include/llvm/CodeGen/RegisterUsageInfo.h

50

add a method print() like other analysis, and call it here behind a flag.

55

Doxygen

58

Doxygen

test/CodeGen/Generic/reg-usage-info.ll

3

I don't like that we can't pass the test in release mode. I suggested that the dump occurs in doFinalization() for the analysis pass (i.e. PhysicalRegisterUsageInfo). You need to implement a proper cl::opt that controls a dump there, that could be used in release mode.

5

I'm worried that we don't provide any stability guarantee on the numbers printed here.
Having a nicer textual form would be better.
I just don't see how to do it other than keep a pointer to the TRI in the DenseMap in the immutable pass to be able to get the register name.

[vivekvpandya]

test/CodeGen/Generic/reg-usage-info.ll

3

Aren't we refering

RegUsageInfoCollector

as analysis pass,

RegisterUsageInfo

is there for keeping data around.

5

If RegUsageInfoCollector is considered as analysis pass then we can add cl::opt into that file and have analysis printed from that.

[mehdi_amini]

test/CodeGen/Generic/reg-usage-info.ll

3

Technically in LLVM the analysis is the pass that you require and query the result from (i.e. a pass you're using as getAnalysis()). But yeah terminology can be fuzzy.

[vivekvpandya]

test/CodeGen/Generic/reg-usage-info.ll

3

SO What is your suggestion? Which file should be changed?

[mehdi_amini]

test/CodeGen/Generic/reg-usage-info.ll

3

I think I already told you: the way analysis are tested at the IR level is using the -analyze flag in opt. Since we don't have -analyze in llc, the closest I can imagine is to implement the analysis (in the LLVM sense) with the "print(...)" method as other analyses, and call it in doFinalization() when a cl::opt flag is set.

[vivekvpandya]

test/CodeGen/Generic/reg-usage-info.ll

5

Will TRI pointer assigned a value by the MachineFunction pass ? Or Is there any way to get TargetMachine or TargetRegisterInfo from Module object? Also if PReg name is required to be printed then MCRegisterInfo will be required.

[vivekvpandya]

a new command line option dump-regusage has been added so that reg usage information can be printed in release build. The test case is also changed accordingly.

[mehdi_amini]

include/llvm/CodeGen/RegisterUsageInfo.h

51

Is the TRI dependent on the Target or the SubTarget?
If it is the latter, it can change on each function and thus needs to be in the map (at which point we better have a dedicated class for the map value)

56

Same question.

lib/CodeGen/RegisterUsageInfo.cpp

24

I'm not sure the name of the option is great, but I'll leave the naming to Quentin/Matthias.

49

Can you implement the print in a separate print method with the same signature as the other analyses for consistency?
And call it from here being the DumpRegUsage flag?

[mehdi_amini]

lib/CodeGen/RegUsageInfoCollector.cpp

128

Mdl is not usual, M alone is more common (or sometimes Mod)

[vivekvpandya]

lib/CodeGen/RegUsageInfoCollector.cpp

128

Ok I will take care for naming , but that would be not required if we are going with Function * , because MF.getFunction() would be enough.

[mehdi_amini]

lib/CodeGen/RegUsageInfoCollector.cpp

130

This is not the right API: getNamedGlobal is for global *variables*, not function.

lib/CodeGen/RegisterUsageInfo.cpp

56

Add an assertion that MFGlobalVar is not null (and the name isn't very explicit)

[vivekvpandya]

lib/CodeGen/RegisterUsageInfo.cpp

56

now I think such assertion wold be for Function *, right ?

[vivekvpandya]

lib/CodeGen/RegUsageInfoCollector.cpp

130

I tried out

PRUI->storeUpdateRegUsageInfo(Mdl->getGlobalVariable(MF.getFunction()->getName(),true), std::move(RegMask));

and

PRUI->storeUpdateRegUsageInfo(Mdl->getGlobalVariable(MF.getFunction()->getName()), std::move(RegMask));

but both returns nullptr for some functions.

[mehdi_amini]

lib/CodeGen/RegUsageInfoCollector.cpp

130

My previous comment called out the fact that you were using an API for global *variables* instead of function, and you replaced with a call to getGlobalVariable?

There is a getFunction() API on the module that returns only functions, or there is a getNamedValue() that will return any symbol.

However you have MF.getFunction(), so you should not call any of these.

[vivekvpandya]

RegMasks uses Function * as key, print method and TargetMachine * have added to RegisterUsageInfo. Command line option -dump-regusage has been renamed to -print-regusage, test case is also updated to reflect the changes.

[mehdi_amini]

lib/CodeGen/RegisterUsageInfo.cpp

66

pair is not coding-convention friendly.

Also you are iterating on a map that is keyed on pointer values, which does not provided any ordering guarantee. Even with names as keys, the map is unordered anyway.
You need to generate a vector and sort it first.

test/CodeGen/Generic/reg-usage-info.ll

40

This test is not OK, for the reasons I explained when I provided the other test (stability).

[vivekvpandya]

Test case changed to simple test case which is originally provided by Mehdi Amini, print() now sorts analysis details before printing in RegisterUsageInfo.cpp

[mehdi_amini]

lib/CodeGen/RegUsageInfoCollector.cpp

93

I think MatzeB mentioned that TRI is a subclass of MCRI, so I'm not sure why you're using MCRI at all while you have TRI.

124

Why aren't you using MachineOperand::clobbersPhysReg here?

test/CodeGen/Generic/reg-usage-info.ll

9

What will be printed for bar1 and bar2?
Add the CHECK lines here.

[vivekvpandya]

git clang-formated , test case changes, MCRI is not required , clobberedReg expression removed.

[mehdi_amini]

include/llvm/CodeGen/RegisterUsageInfo.h

59–60

Doxygen the two public API above.

lib/CodeGen/RegUsageInfoCollector.cpp

80

No braces

lib/CodeGen/RegisterUsageInfo.cpp

73

No braces.

lib/CodeGen/TargetPassConfig.cpp

505

No braces

633

No braces

[mehdi_amini]

test/CodeGen/Generic/reg-usage-info.ll

16

trailing whitespace here.

[qcolombet]

Hi Vivek,

I was thinking about debug ability and usability of the produced code.
Something to consider in a following patch would be to add in the assembly output comments at the beginning of the function and/or related call site of the function listing what are the register preserved by the function.

The rationale is that now if we play with assembly files produced by different optimization phase, we want a quick way to check if two functions are equivalent.
E.g.,
Let say we have two assembly files produced for the same application:

• bad.s which fails, produced with ipra
• good.s which works, produced without ipra

When bisecting which function causes the failure, it is important to know if it is fine to take a function foo from good.s and replace it into bad.s. This may not be possible if the preserve register are not compatible.

Cheers,
-Quentin

[vivekvpandya]

@qcolombet. It will be very help full, I will look into that. Do we need to keep it default or a separate switch to get those comments ?

[mehdi_amini]

LGTM, with one inline comment.

include/llvm/CodeGen/RegisterUsageInfo.h

60

Strip the "this method is provided to" everywhere.
And here add that it will return null if the function is not known.

[vivekvpandya]

inline-asm related test case added Thanks to Mehdi Amini, and previous test case modified so that it do not fail due to transformation pass for IPRA.

[mehdi_amini]

Do you have commit access?

[vivekvpandya]

No I do not have commit access.

include/llvm/Analysis/CallGraphSCCPass.h

 //
 //===----------------------------------------------------------------------===//
 
 #ifndef LLVM_ANALYSIS_CALLGRAPHSCCPASS_H
 #define LLVM_ANALYSIS_CALLGRAPHSCCPASS_H
 
 #include "llvm/Analysis/CallGraph.h"
 #include "llvm/Pass.h"
+#include "llvm/PassSupport.h"
 
 namespace llvm {
 
 class CallGraphNode;
 class CallGraph;
 class PMStack;
 class CallGraphSCC;
 
public:
 
   typedef std::vector<CallGraphNode *>::const_iterator iterator;
   iterator begin() const { return Nodes.begin(); }
   iterator end() const { return Nodes.end(); }
 
   const CallGraph &getCallGraph() { return CG; }
 };
 
+void initializeDummyCGSCCPassPass(PassRegistry &);
+
+/// This pass is required by interprocedural register allocation. It forces
+/// codegen to follow bottom up order on call graph. 
+class DummyCGSCCPass : public CallGraphSCCPass {
+public:
+  static char ID;
+  DummyCGSCCPass() : CallGraphSCCPass(ID){
+    PassRegistry &Registry = *PassRegistry::getPassRegistry();
+    initializeDummyCGSCCPassPass(Registry);
+  };
+  bool runOnSCC(CallGraphSCC &SCC) override { return false; }
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.setPreservesAll();
+  }
+};
+
 } // End llvm namespace
             
 #endif

include/llvm/CodeGen/MachineOperand.h

private:
 
   /// SmallContents - This really should be part of the Contents union, but
   /// lives out here so we can get a better packed struct.
   /// MO_Register: Register number.
   /// OffsetedInfo: Low bits of offset.
   union {
     unsigned RegNo;           // For MO_Register.
     unsigned OffsetLo;        // Matches Contents.OffsetedInfo.OffsetHi.
   } SmallContents;
 
   /// ParentMI - This is the instruction that this operand is embedded into.
   /// This is valid for all operand types, when the operand is in an instr.
   MachineInstr *ParentMI;
 
   /// Contents union - This contains the payload for the various operand types.
   union {
     MachineBasicBlock *MBB;  // For MO_MachineBasicBlock.
void setIndex(int Idx) {
     Contents.OffsetedInfo.Val.Index = Idx;
   }
 
   void setMBB(MachineBasicBlock *MBB) {
     assert(isMBB() && "Wrong MachineOperand accessor");
     Contents.MBB = MBB;
   }
 
+  /// Sets value of register mask operand referencing Mask.  The
+  /// operand does not take ownership of the memory referenced by Mask, it must
+  /// remain valid for the lifetime of the operand. See CreateRegMask().
+  /// Any physreg with a 0 bit in the mask is clobbered by the instruction.
+  void setRegMask(const uint32_t *RegMaskPtr) {
+    assert(isRegMask() && "Wrong MachineOperand mutator");
+    Contents.RegMask = RegMaskPtr;
+  }
+
   //===--------------------------------------------------------------------===//
   // Other methods.
   //===--------------------------------------------------------------------===//
 
   /// isIdenticalTo - Return true if this operand is identical to the specified
   /// operand. Note: This method ignores isKill and isDead properties.
   bool isIdenticalTo(const MachineOperand &Other) const;
 

include/llvm/CodeGen/Passes.h

Pass *createGlobalMergePass(const TargetMachine *TM, unsigned MaximalOffset,
                               bool OnlyOptimizeForSize = false,
                               bool MergeExternalByDefault = false);
 
   /// This pass splits the stack into a safe stack and an unsafe stack to
   /// protect against stack-based overflow vulnerabilities.
   FunctionPass *createSafeStackPass(const TargetMachine *TM = nullptr);
 
   /// This pass detects subregister lanes in a virtual register that are used
   /// independently of other lanes and splits them into separate virtual
   /// registers.
   extern char &RenameIndependentSubregsID;
+  
+  /// This pass is executed POST-RA to collect which physical registers are
+  /// preserved by given machine function.
+  FunctionPass *createRegUsageInfoCollector();
 } // End llvm namespace
 
 /// Target machine pass initializer for passes with dependencies. Use with
 /// INITIALIZE_TM_PASS_END.
 #define INITIALIZE_TM_PASS_BEGIN INITIALIZE_PASS_BEGIN
 
 /// Target machine pass initializer for passes with dependencies. Use with
 /// INITIALIZE_TM_PASS_BEGIN.

include/llvm/CodeGen/RegisterUsageInfo.h

+//==- RegisterUsageInfo.h - Register Usage Informartion Storage -*- C++ -*-===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+/// \file
+/// This pass is required to take advantage of the interprocedural register 
+/// allocation infrastructure.
+///
+/// This pass is simple immutable pass which keeps RegMasks (calculated based on 
+/// actual register allocation) for functions in a module and provides simple API 
+/// to query this information.
+///
+//===----------------------------------------------------------------------===//
+
+#ifndef LLVM_CODEGEN_PHYSICALREGISTERUSAGEINFO_H
+#define LLVM_CODEGEN_PHYSICALREGISTERUSAGEINFO_H
+
+#include "llvm/Pass.h"
+#include "llvm/IR/Module.h"
+#include "llvm/ADT/DenseMap.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/IR/GlobalVariable.h"
+#include "llvm/Support/CommandLine.h"
+#include "llvm/Support/raw_ostream.h"
+
+namespace llvm {
+
+class PhysicalRegisterUsageInfo : public ImmutablePass {
+  virtual void anchor();
+
+public:
+  static char ID;
+
+  PhysicalRegisterUsageInfo() : ImmutablePass(ID) {
+    PassRegistry &Registry = *PassRegistry::getPassRegistry();
+    initializePhysicalRegisterUsageInfoPass(Registry);
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override {
+    AU.setPreservesAll();
+  }
+
+  /// This method is provided to set TargetRegisterInfo * , which is used to dump
+  /// analysis when command line option -dump-regusage is used.
+  void setTargetRegisterInfo(const TargetRegisterInfo *TRI_) {
+    TRI = TRI_;
+  }
+
+  /// This method is provided to set MCRegisterInfo * , which is used to dump
+  /// analysis when command line option -dump-regusage is used.
+  void setMCRegisterInfo(const MCRegisterInfo *MCRI_) {
+    MCRI = MCRI_;
+  }
+
+  bool doInitialization(Module &M) override;
+
+  bool doFinalization(Module &M) override;
+
+  void storeUpdateRegUsageInfo(const GlobalVariable* MFGlobalVar,
+                               std::vector<uint32_t> RegMasks);
+
+  const std::vector<uint32_t> *getRegUsageInfo(const GlobalVariable* MFGlobalVar);
+
+private:
+  /// A Dense map from GlobalVariable * for Function to RegMask.
+  /// In RegMask 0 means register used (clobbered) by function.
+  /// and 1 means content of register will be preserved around function call.
+  DenseMap<const GlobalVariable *, std::vector<uint32_t>> RegMasks;
+  const MCRegisterInfo *MCRI;
+  const TargetRegisterInfo *TRI;
+};
+}
+
+#endif

include/llvm/InitializePasses.h

 void initializeObjCARCExpandPass(PassRegistry&);
 void initializeObjCARCContractPass(PassRegistry&);
 void initializeObjCARCOptPass(PassRegistry&);
 void initializePAEvalPass(PassRegistry &);
 void initializeOptimizePHIsPass(PassRegistry&);
 void initializePartiallyInlineLibCallsLegacyPassPass(PassRegistry &);
 void initializePEIPass(PassRegistry&);
 void initializePHIEliminationPass(PassRegistry&);
+void initializePhysicalRegisterUsageInfoPass(PassRegistry &);
 void initializePartialInlinerPass(PassRegistry&);
 void initializePeepholeOptimizerPass(PassRegistry&);
 void initializePostDomOnlyPrinterPass(PassRegistry&);
 void initializePostDomOnlyViewerPass(PassRegistry&);
 void initializePostDomPrinterPass(PassRegistry&);
 void initializePostDomViewerPass(PassRegistry&);
 void initializePostDominatorTreeWrapperPassPass(PassRegistry&);
 void initializePostOrderFunctionAttrsLegacyPassPass(PassRegistry&);

 void initializeSjLjEHPreparePass(PassRegistry&);
 void initializeDemandedBitsWrapperPassPass(PassRegistry&);
 void initializeFuncletLayoutPass(PassRegistry &);
 void initializeLoopLoadEliminationPass(PassRegistry&);
 void initializeFunctionImportPassPass(PassRegistry &);
 void initializeLoopVersioningPassPass(PassRegistry &);
 void initializeWholeProgramDevirtPass(PassRegistry &);
 void initializePatchableFunctionPass(PassRegistry &);
 }
 
 #endif

lib/Analysis/CallGraphSCCPass.cpp

 }
 
 bool CallGraphSCCPass::skipSCC(CallGraphSCC &SCC) const {
   return !SCC.getCallGraph().getModule()
               .getContext()
               .getOptBisect()
               .shouldRunPass(this, SCC);
 }
+
+char DummyCGSCCPass::ID = 0;
+INITIALIZE_PASS(DummyCGSCCPass, "DummyCGSCCPass", "DummyCGSCCPass", false,
+                   false)
+

lib/CodeGen/CMakeLists.txt

add_llvm_library(LLVMCodeGen
   RegAllocBasic.cpp
   RegAllocFast.cpp
   RegAllocGreedy.cpp
   RegAllocPBQP.cpp
   RegisterClassInfo.cpp
   RegisterCoalescer.cpp
   RegisterPressure.cpp
   RegisterScavenging.cpp
   RenameIndependentSubregs.cpp
+  RegisterUsageInfo.cpp
+  RegUsageInfoCollector.cpp
   SafeStack.cpp
   ScheduleDAG.cpp
   ScheduleDAGInstrs.cpp
   ScheduleDAGPrinter.cpp
   ScoreboardHazardRecognizer.cpp
   ShadowStackGCLowering.cpp
   ShrinkWrap.cpp
   SjLjEHPrepare.cpp

lib/CodeGen/RegUsageInfoCollector.cpp

+//===- RegUsageInfoCollector.cpp - Register Usage Informartion Collector --===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+///
+/// This pass is required to take advantage of the interprocedural register 
+/// allocation infrastructure.
+///
+/// This pass is simple MachineFunction pass which collects register usage
+/// details by iterating through each physical registers and checking
+/// MRI::isPhysRegUsed() then creates a RegMask based on this details.
+/// The pass then stores this RegMask in PhysicalRegisterUsageInfo.cpp
+///
+//===----------------------------------------------------------------------===//
+
+#include "llvm/CodeGen/MachineBasicBlock.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstr.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/CodeGen/RegisterUsageInfo.h"
+#include "llvm/IR/Module.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "ip-regalloc"
+
+namespace llvm {
+void initializeRegUsageInfoCollectorPass(PassRegistry &);
+}
+
+namespace {
+class RegUsageInfoCollector : public MachineFunctionPass {
+public:
+  RegUsageInfoCollector() : MachineFunctionPass(ID) {
+    PassRegistry &Registry = *PassRegistry::getPassRegistry();
+    initializeRegUsageInfoCollectorPass(Registry);
+  }
+
+  const char *getPassName() const override {
+    return "Register Usage Information Collector Pass";
+  }
+
+  void getAnalysisUsage(AnalysisUsage &AU) const override;
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+
+  static char ID;
+
+private:
+  void markRegClobbered(const TargetRegisterInfo *TRI, uint32_t *RegMask,
+                        unsigned PReg);
+};
+} // end of anonymous namespace
+
+char RegUsageInfoCollector::ID = 0;
+
+INITIALIZE_PASS_BEGIN(RegUsageInfoCollector, "RegUsageInfoCollector",
+                      "Register Usage Information Collector", false, false)
+INITIALIZE_PASS_DEPENDENCY(PhysicalRegisterUsageInfo)
+INITIALIZE_PASS_END(RegUsageInfoCollector, "RegUsageInfoCollector",
+                    "Register Usage Information Collector", false, false)
+
+FunctionPass *llvm::createRegUsageInfoCollector() {
+  return new RegUsageInfoCollector();
+}
+
+void RegUsageInfoCollector::markRegClobbered(const TargetRegisterInfo *TRI,
+                                             uint32_t *RegMask, unsigned PReg) {
+  // If PReg is clobbered then all of its alias are also clobbered.
+  for (MCRegAliasIterator AI(PReg, TRI, true); AI.isValid(); ++AI) {
+    RegMask[*AI / 32] &= ~(1u << *AI % 32);
+    DEBUG(dbgs() << *AI << " ");
+  }
+}
+
+void RegUsageInfoCollector::getAnalysisUsage(AnalysisUsage &AU) const {
+  AU.addRequired<PhysicalRegisterUsageInfo>();
+  AU.setPreservesAll();
+  MachineFunctionPass::getAnalysisUsage(AU);
+}
+
+bool RegUsageInfoCollector::runOnMachineFunction(MachineFunction &MF) {
+  MachineRegisterInfo *MRI = &MF.getRegInfo();
+  TargetRegisterInfo *TRI =
+      (TargetRegisterInfo *)MF.getSubtarget().getRegisterInfo();
+  const TargetMachine &TM = MF.getTarget();
+  const MCRegisterInfo *MCRI = TM.getMCRegisterInfo();
+
+  DEBUG(dbgs() << " -------------------- " << getPassName()
+               << " ----------------------- \n");
+  DEBUG(dbgs() << "Function Name : " << MF.getName() << "\n");
+
+  std::vector<uint32_t> RegMask;
+
+  // Compute the size of the bit vector to represent all the registers.
+  // The bit vector is broken into 32-bit chunks, thus takes the ceil of
+  // the number of registers divided by 32 for the size.
+  unsigned regMaskSize = (TRI->getNumRegs() + 31) / 32;
+  RegMask.resize(regMaskSize, 0xFFFFFFFF);
+
+  PhysicalRegisterUsageInfo *PRUI = &getAnalysis<PhysicalRegisterUsageInfo>();
+  PRUI->setTargetRegisterInfo(TRI);
+  PRUI->setMCRegisterInfo(MCRI);
+
+  DEBUG(dbgs() << "Clobbered Registers: ");
+  for (unsigned PReg = 1, PRegE = TRI->getNumRegs(); PReg < PRegE; ++PReg) {
+    if (!MRI->reg_nodbg_empty(PReg) && MRI->isPhysRegUsed(PReg))
+      markRegClobbered(TRI, &RegMask[0], PReg);
+  }
+
+  const uint32_t *CallPreservedMask =
+      TRI->getCallPreservedMask(MF, MF.getFunction()->getCallingConv());
+  // Set callee saved register as preserved.
+  for (unsigned index = 0; index < regMaskSize; index++) {
+    RegMask[index] = RegMask[index] | CallPreservedMask[index];
+  }
+
+  DEBUG(dbgs() << " \n-----------------------------------------------------------"
+                  "------------------ \n");
+
+  const Module *Mdl = MF.getFunction()->getParent();
+
+  PRUI->storeUpdateRegUsageInfo(Mdl->getNamedGlobal(MF.getName()),
+                                std::move(RegMask));
+
+  return false;
+}

lib/CodeGen/RegisterUsageInfo.cpp

+//===- RegisterUsageInfo.cpp - Register Usage Informartion Storage --------===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+///
+/// This pass is required to take advantage of the interprocedural register 
+/// allocation infrastructure.
+///
+//===----------------------------------------------------------------------===//
+
+#include "llvm/IR/Module.h"
+#include "llvm/CodeGen/RegisterUsageInfo.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "ip-regalloc"
+
+cl::opt<bool> DumpRegUsage("dump-regusage", cl::init(false), cl::Hidden,
+            cl::desc("Dump register usage details collected for analysis."));
+
+INITIALIZE_PASS(PhysicalRegisterUsageInfo, "reg-usage-info",
+                "Register Usage Informartion Stroage", false, true)
+
+char PhysicalRegisterUsageInfo::ID = 0;
+
+void PhysicalRegisterUsageInfo::anchor() {}
+
+bool PhysicalRegisterUsageInfo::doInitialization(Module &M) {
+  RegMasks.grow(M.size());
+  return false;
+}
+
+bool PhysicalRegisterUsageInfo::doFinalization(Module &M) {
+  if (DumpRegUsage) {
+    for (auto pair : RegMasks) {
+      errs() << "Clobbered Registers: ";
+      for (unsigned PReg = 1, PRegE = TRI->getNumRegs(); PReg < PRegE; ++PReg) {
+        if (!(pair.second[PReg / 32] & 1u << PReg % 32))
+          errs() << MCRI->getName(PReg) << " ";
+      }
+      errs() << "\n";
+    }
+  }
+
+  RegMasks.shrink_and_clear();
+  return false;
+}
+
+void PhysicalRegisterUsageInfo::storeUpdateRegUsageInfo(
+    const GlobalVariable* MFGlobalVar, std::vector<uint32_t> RegMask) {
+  RegMasks[MFGlobalVar] = std::move(RegMask);
+}
+
+const std::vector<uint32_t> *
+PhysicalRegisterUsageInfo::getRegUsageInfo(const GlobalVariable* MFGlobalVar) {
+  if (RegMasks.find(MFGlobalVar) != RegMasks.end())
+    return &(RegMasks.find(MFGlobalVar)->second);
+  return nullptr;
+}

lib/CodeGen/TargetPassConfig.cpp

 // This file defines interfaces to access the target independent code
 // generation passes provided by the LLVM backend.
 //
 //===---------------------------------------------------------------------===//
 
 #include "llvm/CodeGen/TargetPassConfig.h"
 
 #include "llvm/Analysis/BasicAliasAnalysis.h"
+#include "llvm/Analysis/CallGraphSCCPass.h"
 #include "llvm/Analysis/CFLAliasAnalysis.h"
 #include "llvm/Analysis/Passes.h"
 #include "llvm/Analysis/ScopedNoAliasAA.h"
 #include "llvm/Analysis/TypeBasedAliasAnalysis.h"
 #include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/RegisterUsageInfo.h"
 #include "llvm/CodeGen/RegAllocRegistry.h"
 #include "llvm/IR/IRPrintingPasses.h"
 #include "llvm/IR/LegacyPassManager.h"
 #include "llvm/IR/Verifier.h"
 #include "llvm/MC/MCAsmInfo.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"

 // Experimental option to run live interval analysis early.
 static cl::opt<bool> EarlyLiveIntervals("early-live-intervals", cl::Hidden,
     cl::desc("Run live interval analysis earlier in the pipeline"));
 
 static cl::opt<bool> UseCFLAA("use-cfl-aa-in-codegen",
   cl::init(false), cl::Hidden,
   cl::desc("Enable the new, experimental CFL alias analysis in CodeGen"));
 
+cl::opt<bool>
+    UseIPRA("enable-ipra", cl::init(false), cl::Hidden,
+            cl::desc("Enable interprocedural register allocation "
+                     "to reduce load/store at procedure calls."));
+
 /// 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();
void TargetPassConfig::addCodeGenPrepare() {
   addPass(createRewriteSymbolsPass());
 }
 
 /// Add common passes that perform LLVM IR to IR transforms in preparation for
 /// instruction selection.
 void TargetPassConfig::addISelPrepare() {
   addPreISel();
 
+  if (UseIPRA) {
+    // Force codegen to run according to the callgraph.
+    addPass(new DummyCGSCCPass);
+  }
+
   // Add both the safe stack and the stack protection passes: each of them will
   // only protect functions that have corresponding attributes.
   addPass(createSafeStackPass(TM));
   addPass(createStackProtectorPass(TM));
 
   if (PrintISelInput)
     addPass(createPrintFunctionPass(
         dbgs(), "\n\n*** Final LLVM Code input to ISel ***\n"));
 
   // All passes which modify the LLVM IR are now complete; run the verifier
   // to ensure that the IR is valid.
   if (!DisableVerify)
     addPass(createVerifierPass());
 }
 
 /// Add the complete set of target-independent postISel code generator passes.
 ///
 /// This can be read as the standard order of major LLVM CodeGen stages. Stages
 /// with nontrivial configuration or multiple passes are broken out below in
 /// add%Stage routines.
 ///
 /// Any TargetPassConfig::addXX routine may be overriden by the Target. The
 /// addPre/Post methods with empty header implementations allow injecting
 /// target-specific fixups just before or after major stages. Additionally,
 /// targets have the flexibility to change pass order within a stage by
 /// overriding default implementation of add%Stage routines below. Each
 /// technique has maintainability tradeoffs because alternate pass orders are
 /// not well supported. addPre/Post works better if the target pass is easily
 /// tied to a common pass. But if it has subtle dependencies on multiple passes,
void TargetPassConfig::addMachinePasses() {
   if (getOptimizeRegAlloc())
     addOptimizedRegAlloc(createRegAllocPass(true));
   else
     addFastRegAlloc(createRegAllocPass(false));
 
   // Run post-ra passes.
   addPostRegAlloc();
 
   // Insert prolog/epilog code.  Eliminate abstract frame index references...
   if (getOptLevel() != CodeGenOpt::None)
     addPass(&ShrinkWrapID);
 
   // Prolog/Epilog inserter needs a TargetMachine to instantiate. But only
   // do so if it hasn't been disabled, substituted, or overridden.
   if (!isPassSubstitutedOrOverridden(&PrologEpilogCodeInserterID))
       addPass(createPrologEpilogInserterPass(TM));
 
   /// Add passes that optimize machine instructions after register allocation.
   if (getOptLevel() != CodeGenOpt::None)
     addMachineLateOptimization();
 
   // Expand pseudo instructions before second scheduling pass.
   addPass(&ExpandPostRAPseudosID);
 
   // Run pre-sched2 passes.
   addPreSched2();
 
   if (EnableImplicitNullChecks)
     addPass(&ImplicitNullChecksID);
 
   // Second pass scheduler.
   // Let Target optionally insert this pass by itself at some other
   // point.
   if (getOptLevel() != CodeGenOpt::None &&
       !TM->targetSchedulesPostRAScheduling()) {
     if (MISchedPostRA)
       addPass(&PostMachineSchedulerID);
     else
       addPass(&PostRASchedulerID);
   }
 
   // GC
   if (addGCPasses()) {
     if (PrintGCInfo)
       addPass(createGCInfoPrinter(dbgs()), false, false);
   }
 
   // Basic block placement.
   if (getOptLevel() != CodeGenOpt::None)
     addBlockPlacement();
 
   addPreEmitPass();
   
+  if (UseIPRA) {
+    // Collect register usage information and produce a register mask of 
+    // clobbered registers, to be used to optimize call sites.
+    addPass(createRegUsageInfoCollector());
+  }
+  
   addPass(&FuncletLayoutID, false);
 
   addPass(&StackMapLivenessID, false);
   addPass(&LiveDebugValuesID, false);
 
   addPass(&PatchableFunctionID, false);
 
   AddingMachinePasses = false;
void TargetPassConfig::addMachineSSAOptimization() {
 
   // With optimization, dead code should already be eliminated. However
   // there is one known exception: lowered code for arguments that are only
   // used by tail calls, where the tail calls reuse the incoming stack
   // arguments directly (see t11 in test/CodeGen/X86/sibcall.ll).
   addPass(&DeadMachineInstructionElimID);
 
   // Allow targets to insert passes that improve instruction level parallelism,
   // like if-conversion. Such passes will typically need dominator trees and
   // loop info, just like LICM and CSE below.
   addILPOpts();
 
   addPass(&MachineLICMID, false);
   addPass(&MachineCSEID, false);
   addPass(&MachineSinkingID);
 
   addPass(&PeepholeOptimizerID);

test/CodeGen/Generic/reg-usage-info.ll

+; RUN: llc -enable-ipra -dump-regusage -o /dev/null 2>&1 < %s  | FileCheck %s
+; CHECK: Clobbered Registers: AH AL AX DI DIL EAX EDI EFLAGS ESP RAX RDI RSP SP SPL
+
+
+target triple = "x86_64-apple-macosx10.11.0"
+
+; Function Attrs: nounwind ssp uwtable
+define i32 @fib(i32 %n) #0 {
+entry:
+  %retval = alloca i32, align 4
+  %n.addr = alloca i32, align 4
+  store i32 %n, i32* %n.addr, align 4
+  %0 = load i32, i32* %n.addr, align 4
+  %cmp = icmp eq i32 %0, 1
+  br i1 %cmp, label %if.then, label %lor.lhs.false
+
+lor.lhs.false:                                    ; preds = %entry
+  %1 = load i32, i32* %n.addr, align 4
+  %cmp1 = icmp eq i32 %1, 2
+  br i1 %cmp1, label %if.then, label %if.end
+
+if.then:                                          ; preds = %lor.lhs.false, %entry
+  store i32 1, i32* %retval, align 4
+  br label %return
+
+if.end:                                           ; preds = %lor.lhs.false
+  %2 = load i32, i32* %n.addr, align 4
+  %sub = sub nsw i32 %2, 1
+  %call = call i32 @fib(i32 %sub)
+  %3 = load i32, i32* %n.addr, align 4
+  %sub2 = sub nsw i32 %3, 2
+  %call3 = call i32 @fib(i32 %sub2)
+  %add = add nsw i32 %call, %call3
+  store i32 %add, i32* %retval, align 4
+  br label %return
+
+return:                                           ; preds = %if.end, %if.then
+  %4 = load i32, i32* %retval, align 4
+  ret i32 %4
+}
+
+; Function Attrs: nounwind ssp uwtable
+define i32 @main() #0 {
+entry:
+  %retval = alloca i32, align 4
+  %n = alloca i32, align 4
+  store i32 0, i32* %retval, align 4
+  store i32 10, i32* %n, align 4
+  %0 = load i32, i32* %n, align 4
+  %call = call i32 @fib(i32 %0)
+  ret i32 %call
+}
+
+attributes #0 = { nounwind ssp uwtable "disable-tail-calls"="false" "less-precise-fpmad"="false" "no-frame-pointer-elim"="true" "no-frame-pointer-elim-non-leaf" "no-infs-fp-math"="false" "no-jump-tables"="false" "no-nans-fp-math"="false" "stack-protector-buffer-size"="8" "target-cpu"="core2" "target-features"="+cx16,+fxsr,+mmx,+sse,+sse2,+sse3,+ssse3,+x87" "unsafe-fp-math"="false" "use-soft-float"="false" }