[MC] Pass MCSubtargetInfo to fixupNeedsRelaxation and applyFixup

[peter.smith]

Summary

On targets like Arm some relaxations may only be performed when certain architectural features are available. As functions can be compiled with differing levels of architectural support we must make a judgement on whether we can relax based on the MCSubtargetInfo for the function. This change passes through the MCSubtargetInfo for the function to fixupNeedsRelaxation so that the decision on whether to relax can be made per function. In this patch, only the ARM backend makes use of this information. We must also pass the MCSubtargetInfo to applyFixup because some fixups skip error checking on the assumption that relaxation has occurred, to prevent code-generation errors applyFixup must see the same MCSubtargetInfo as fixupNeedsRelaxation.

This addresses part of PR36542 where each function has the cpu and target features it was compiled with in its attributes, these features included support for Thumb2, whereas the default cpu passed to the lto plugin did not. As the decision on whether to relax was using the default cpu, a branch, used for a tail call by a Thumb2 function did not get relaxed to a wide branch.

I don't think that there is an alternative to making the relaxation decision based on per function information as if we use some kind of aggregate MCSubtargetInfo determined from a combination of all the Functions and the module we run into two self contradictory use-cases

• I want to do runtime selection of a specific function, we should use the least capable target feature set for a per module relaxation decision.
• I want to run on a specific target architecture, we should use the target features of that arch/cpu for a per module relaxation decision.

As we can't tell which of these use-cases I think we should make the decision per function.

[fhahn]

Thanks Peter! I think using the function attributes here makes a lot of sense.

Currently, ARMAsmBackend still has a STI member variable, which gets initialized when constructing the backend. At most places, the current STI passed in, but at some places (e.g. adjustFixupValue, line 592) still use the member variable. Are there use-cases where the initial STI contains the correct features, but the current STI does not? I can't think of any off the top of my head and to avoid confusion over which STI is used, it might make sense to get rid of the member variable?

[peter.smith]

Thanks for pointing that out. I should check each use of the member to see if the information can be used incorrectly. I think it would be good if we can remove the MCSubtargetInfo, I have my suspicions that it might be used from Assembly and for BuildAttributes generation (on Arm), again some more checking is needed.

[efriedma]

I think ARMAsmBackend::writeNopData and ARMAsmBackend::applyFixup also need to use a passed-in STI rather than one stored on the ARMAsmBackend.

[echristo]

In D44928#1052244, @efriedma wrote:

I think ARMAsmBackend::writeNopData and ARMAsmBackend::applyFixup also need to use a passed-in STI rather than one stored on the ARMAsmBackend.

Agreed. I'm reasonably certain the cached one should just be removed.

[peter.smith]

Thanks for the comments and apologies for the delayed response (Easter break). The patch here is somewhat fortunate in that the MCRelaxableFragment already has a reference to the MCSubtargetInfo. The other instances ARMAsmBackend::writeNopData and ARMAsmBackend::applyFixup may be called on other Fragments that don't have the MCSubtargetInfo member such as MCAlignFragment and MCDataFragment. I'll do some investigation to see what the best options are.

[peter.smith]

I've done a bit of thinking about how best to get the necessary information to ARMAsmBackend::writeNopData and ARMAsmBackend::applyFixup. I made an attempt at passing MCSubtargetInfo through the necessary layers so that STI could be removed. I think that this is impractical for writeNopData, although possible for applyFixup. What I propose to do:

• The writeNopData() currently checks to see if the Architectural NOP instruction is supported, based on the MCSubtargetInfo in ARMAsmBackend. I think that an EmitAssemblerFlag that is used in the same way as the flag to switch from Thumb to Arm can toggle the support for the Architectural NOP without having to do large amounts of source code changes, most of them outside the ARM backend.
• All the fixups that test the MCSubTargetInfo for target features are for instruction fixups, these are covered by the MCRelaxableFragment type that already has STI embedded in it. Rather than adding STI to the Data Fragments with relaxations I propose to pass a pointer to STI that will be null for Data Fragments.

[efriedma]

Rather than adding STI to the Data Fragments with relaxations I propose to pass a pointer to STI that will be null for Data Fragments.

This seems fine.

I think that an EmitAssemblerFlag that is used in the same way as the flag to switch from Thumb to Arm can toggle the support for the Architectural NOP without having to do large amounts of source code changes, most of them outside the ARM backend.

This is really hacky... the fragments in question should really have an STI instead. Plus there are actually separate features; whether NOP is supported in ARM mode (which is what hasNOP queries), and whether NOP is supported in Thumb mode (i.e. whether Thumb2 is supported). Plus other architectures have similar problems (e.g. x86 changes the generated NOPs based on the target). If you're not going to fix it properly, please just leave it as-is for now, with a FIXME.

[peter.smith]

I think that an EmitAssemblerFlag that is used in the same way as the flag to switch from Thumb to Arm can toggle the support for the Architectural NOP without having to do large amounts of source code changes, most of them outside the ARM backend.

This is really hacky... the fragments in question should really have an STI instead. Plus there are actually separate features; whether NOP is supported in ARM mode (which is what hasNOP queries), and whether NOP is supported in Thumb mode (i.e. whether Thumb2 is supported). Plus other architectures have similar problems (e.g. x86 changes the generated NOPs based on the target). If you're not going to fix it properly, please just leave it as-is for now, with a FIXME.

Ok; I'd like to fix the problem as it would allow us to get rid of the STI in the ARM/X86 AsmBackend class. I think it would be better off as a separate patch as I don't have a good answer for how to achieve this right now. The main difficulty is that Nops can be written as padding for some fragments such as MCAlignFragment created by EmitCodeAlignment(). There are places where EmitCodeAlignment() is called where some calculation is needed to work out what the STI should be, for example the Constant Island emission, or the STI is not easily available such as parseDirectiveAlign(). These calculations might also need other MCFragments to record the STI so we can walk backwards to find the last STI. I think that these are soluble problems but I'm still thinking that I'm missing something simpler. I'd be very grateful if you have any ideas or suggestions?

[efriedma]

For the nop thing, fundamentally, I don't think there's any simpler approach. We need to annotate each MCAlignFragment created by MCObjectStreamer::EmitCodeAlignment with the correct STI, so we can figure out the correct nops later. We currently throw away that information, and there isn't any way to recover it later.

Consider something like the following; there are two alignment fragments, which each need a different STI.

.arch armv6
foo:
mov r1, r0
nop
.p2align 4
bx lr
.arch armv7-a
bar:
mov r1, r0
nop
.p2align 4
bx lr

I just tried binutils as for ARM, and it has an amusing bug: the nops used for all alignment directives in a file is based on the last ".arch" directive in that file.

[peter.smith]

I've updated the diff, description and title to include passing through MCSubtargetInfo to applyFixup() as well. I think that it is worth combining these changes into one as it is possible to introduce a code generation fault with just fixupNeedsRelaxation() receiving the STI.

I found out that it is possible for MCDataFragments to contain instructions with fixups so I've moved up the MCSubtargetInfo STI member from MCRelaxedFragment. To make sure that an STI is present when hasInstructions() is true, I've modified setHasInstructions to take an STI. This flushed out a few streamers that were not setting setHasInstructions() for MCDataFragments that contained instructions with fixups.

I note that the per fragment use of STI in the ARMAsmBackend::applyFixup() seems at best an optimisation and at worst redundant. As I understand it, each use is just skipping the out of range error message check, which we would never see for those subtargets because the instruction would have previously been relaxed. I can remove the tests that check STI->getFeatureBits() and all the tests still pass, I could be missing something though. I've added an extra test case that will silently generate incorrect code (skips an out of range immediate check) if STI is passed to fixupNeedsRelaxation() but not applyFixup().

I've not covered writeNopData() in this patch as I think that it would be best covered by a separate patch or patches.

[efriedma]

include/llvm/MC/MCAsmBackend.h

83 ↗

(On Diff #141320)

This comment needs to be updated?

lib/MC/MCELFStreamer.cpp

90 ↗

(On Diff #141320)

Is it possible for these two subtargets to be different?

[peter.smith]

Thanks for the comments. I think your second one brings up a couple of wider problems that I'll need to address:

• getOrCreateDataFragment() needs to start a new fragment when the original contains instructions and the new one has a different Subtarget. Otherwise the later Subtarget will overwrite the first, I don't think that this makes any practical difference for the MCAsmBackend functions changed in this patch, however it will do for the WriteNops().
• For instruction bundling in cases where the previous fragment must be used then it must either have the same Subtarget, or we permit the later Subtarget to overwrite the first.

I'm going to be away from the office for the next 7 days for a couple of conferences (including Euro LLVM) so I'll get back to this then.

[peter.smith]

I've updated the diff to account for instruction bundling. When an instruction must be put in the same fragment for the purposes of bundling we now check that the MCSubtargetInfo for the fragment is consistent. I've added some tests to check that we detect an attempt to change subtarget within a bundle.

My understanding is that instruction bundling was implemented for Arm and X86 NaCl targets https://developer.chrome.com/native-client/reference/sandbox_internals/arm-32-bit-sandbox . Given that NaCl on Arm is only supported on Arm V7 ARM only I think it is very unlikely that anyone has attempted to do this in production. On X86 I think that switching between 32-bit/64-bit mode within a bundle is also highly unlikely.

[efriedma]

I'm not sure I completely follow the bundle-related changes; adding Petr Hosek as a reviewer.

Otherwise LGTM, but I'd feel more comfortable if this got a second review.

[peter.smith]

Thanks for the review. I'll wait a bit to see if there is a second opinion, I too would value more input on instruction bundling as I've not used it myself. I've completed a set of patches that will pass through the subtarget to writeNops() and remove STI from the X86 and ARM AsmBackends, these have some further reorganisation of the position of instruction bundling inside the MCFragment hierarchy. I'll add these as dependent reviews to this one and link them in a comment.

[peter.smith]

I've added reviews for the support of writeNops(). My apologies for the size of the changes, the first three make sure the STI is recorded in the fragment so that we can retrieve it when calling writeNops. The final one passes STI through to writeNops and removes STI from the X86 and ARM AsmBackends.
1 [MC] Move bundling and MCSubtargetInfo to MCEncodedFragment [NFC] D45959
2 [MC] Add MCSubtargetInfo to MCPaddingFragment [NFC] D45960
3 [MC] Add MCSubtargetInfo to MCAlignFragment [NFC] D45961
4 [MC] Use local MCSubtargetInfo in writeNops D45962

[peter.smith]

I've not had any more comments this week on the approach to instruction bundling. Would anyone have any objections to committing this next week as this will fix pr36542, and will at least not make the writeNops case any worse?

[efriedma]

@phosek Are you planning to look at this?

[peter.smith]

I've rebased the patch against current trunk. The littleEndian parameter was removed from applyFixup, NFC for the patch.

[echristo]

Taking a quick glance through this I wonder if adding the subtarget to the fragment is necessary since we've got it at encodeInstruction time? I haven't done a thorough look, but I'm concerned we're missing something here.

Thoughts? Also one similar sort of inline comment.

lib/Target/ARM/MCTargetDesc/ARMAsmBackend.h

22–25 ↗

(On Diff #147010)

I think we should just rely on a passed in STI wherever necessary and only store module/object level caches here. Also applies to isThumbMode below :)

Thoughts?

[peter.smith]

Thanks very much for the comments. Although the STI is available at the time the instruction is encoded; due to relaxation being a separate pass resolved after all instructions have been emitted we don't have the STI to pass to applyFixups() as all we have is a list of fragments. For example the MCRelaxableFragment has to store the STI as it might need to re-encode the instruction during relaxation. Thinking about some possible alternatives:

• Make another pass over the input during relaxation (and at other times we need the STI, such as emitting NOPS in alignment padding), so that we know what STI is in effect at the time. I think that this would be a much larger change, in effect making the assembler 2 pass.
• Maintain some kind of mapping table between fragment and STI. I think that this is logically equivalent to storing the STI in the fragment though.
• Store the STI per MCInst, this seems wasteful and I don't think it would work for writeNops() as these come from alignment and padding fragments.

In short I've not been able to come up with anything better. If I have missed something and there is a better alternative I'd be happy to try implementing it.

lib/Target/ARM/MCTargetDesc/ARMAsmBackend.h

22–25 ↗

(On Diff #147010)

I've got some follow up patches that do this for STI; the other use is in writeNops() . Unfortunately the patches needed to do this are substantially larger and more complicated than this one. Given that writeNops() matters less for correctness than this one I thought it better to do in separate patches. I've not yet looked at isThumbMode but my intuition is that it should also be possible to do as a follow up patch.

Repeating the comment from earlier in the history:

I've added reviews for the support of writeNops(). My apologies for the size of the changes, the first three make sure the STI is recorded in the fragment so that we can retrieve it when calling writeNops. The final one passes STI through to writeNops and removes STI from the X86 and ARM AsmBackends.
1 [MC] Move bundling and MCSubtargetInfo to MCEncodedFragment [NFC] D45959
2 [MC] Add MCSubtargetInfo to MCPaddingFragment [NFC] D45960
3 [MC] Add MCSubtargetInfo to MCAlignFragment [NFC] D45961
4 [MC] Use local MCSubtargetInfo in writeNops D45962

[peter.smith]

Ping. Is anyone willing to approve or has any thoughts on a better design?

If I can summarise so far:

• Each MachineFunction can have its own Subtarget.
• On targets such as Arm, and I think also Mips, it is possible to change the Subtarget on the fly in assembler with a .arch directive.
• On Arm, relaxation and fixup calculation decisions can be affected by the Subtarget. PR36542 shows a real example with LTO where a default CPU of ARM7TDMI (Architecture v4t) was passed to MC, yet the MachineFunctions were generated using a much more recent Armv7A, this caused a fixup error as the code-generator was expecting the 2-byte Thumb1 Branch to be relaxed to a Thumb2 4-byte branch by MC.
• A legitimate use case of .arch is to do runtime selection of architecture so it is not appropriate to just pick either the lowest or highest Subtarget that we find in the Module.
• At the time we do relaxation and fixups (after layout) we don't know what the current active Subtarget is so we must have a way of recovering the Subtarget for relaxations and fixups.
• This patch extends the example of MCRelaxableFragment and stores the Subtarget in the MCDataFragment.
• There is at most one Subtarget per MCDataFragment so we create a new MCDataFragment when the Subtarget changes. This causes some complications with the AlignedBundling and RelaxAll used by NativeClient as all the instructions in a Bundle must be within the same fragment. This patch gives an error message if the Subtarget is changed within a bundle (possible in assembly only). A possible more complex alternative is to accept multiple Subtargets per Fragment but given the restrictions on target of NativeClient (ARM v7A instructions only) and deprecation of NativeClient it is extremely unlikely that this case will need to be handled.
• On both x86 and Arm the writeNops() function is also affected but is not handled by this patch due to the size of the change. There are a series of follow up patches (see earlier comments) for that.

At this stage I can't think of a logical alternative to making a mapping between Fragment and Subtarget. At present the Subtarget is stored in the fragment itself, it could be possible to maintain an external mapping table but I'm not sure whether that would be an improvement as it would need to be threaded through to where it is needed.

[echristo]

Right now I don't have a better design without quite a bit more in depth look. That said, this gets us through a correctness issue and can be revisited in the future so let's go with it.

Thanks for the patience.

-eric

llvm/trunk/include/llvm/MC/MCAsmBackend.h

virtual bool shouldForceRelocation(const MCAssembler &Asm,
                                      const MCValue &Target) {
     return false;
   }
 
   /// Apply the \p Value for given \p Fixup into the provided data fragment, at
   /// the offset specified by the fixup and following the fixup kind as
   /// appropriate. Errors (such as an out of range fixup value) should be
   /// reported via \p Ctx.
+  /// The  \p STI is present only for fragments of type MCRelaxableFragment and
+  /// MCDataFragment with hasInstructions() == true.
   virtual void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                           const MCValue &Target, MutableArrayRef<char> Data,
-                          uint64_t Value, bool IsResolved) const = 0;
+                          uint64_t Value, bool IsResolved,
+                          const MCSubtargetInfo *STI) const = 0;
 
   /// Check whether the given target requires emitting differences of two
   /// symbols as a set of relocations.
   virtual bool requiresDiffExpressionRelocations() const { return false; }
 
   /// @}
 
   /// \name Target Relaxation Interfaces
   /// @{
 
   /// Check whether the given instruction may need relaxation.
   ///
   /// \param Inst - The instruction to test.
-  virtual bool mayNeedRelaxation(const MCInst &Inst) const = 0;
+  /// \param STI - The MCSubtargetInfo in effect when the instruction was
+  /// encoded.
+  virtual bool mayNeedRelaxation(const MCInst &Inst,
+                                 const MCSubtargetInfo &STI) const = 0;
 
   /// Target specific predicate for whether a given fixup requires the
   /// associated instruction to be relaxed.
   virtual bool fixupNeedsRelaxationAdvanced(const MCFixup &Fixup, bool Resolved,
                                             uint64_t Value,
                                             const MCRelaxableFragment *DF,
                                             const MCAsmLayout &Layout,
                                             const bool WasForced) const;

llvm/trunk/include/llvm/MC/MCFragment.h

 
 protected:
   MCEncodedFragmentWithFixups(MCFragment::FragmentType FType,
                               bool HasInstructions,
                               MCSection *Sec)
       : MCEncodedFragmentWithContents<ContentsSize>(FType, HasInstructions,
                                                     Sec) {}
 
+  /// STI - The MCSubtargetInfo in effect when the instruction was encoded.
+  /// must be non-null for instructions.
+  const MCSubtargetInfo *STI = nullptr;
+
 public:
+
+  /// Retrieve the MCSubTargetInfo in effect when the instruction was encoded.
+  /// Guaranteed to be non-null if hasInstructions() == true
+  const MCSubtargetInfo *getSubtargetInfo() const { return STI; }
+
   using const_fixup_iterator = SmallVectorImpl<MCFixup>::const_iterator;
   using fixup_iterator = SmallVectorImpl<MCFixup>::iterator;
 
   SmallVectorImpl<MCFixup> &getFixups() { return Fixups; }
   const SmallVectorImpl<MCFixup> &getFixups() const { return Fixups; }
 
   fixup_iterator fixup_begin() { return Fixups.begin(); }
   const_fixup_iterator fixup_begin() const { return Fixups.begin(); }

 
 /// Fragment for data and encoded instructions.
 ///
 class MCDataFragment : public MCEncodedFragmentWithFixups<32, 4> {
 public:
   MCDataFragment(MCSection *Sec = nullptr)
       : MCEncodedFragmentWithFixups<32, 4>(FT_Data, false, Sec) {}
 
-  void setHasInstructions(bool V) { HasInstructions = V; }
+  /// Record that the fragment contains instructions with the MCSubtargetInfo in
+  /// effect when the instruction was encoded.
+  void setHasInstructions(const MCSubtargetInfo &STI) {
+    HasInstructions = true;
+    this->STI = &STI;
+  }
 
   static bool classof(const MCFragment *F) {
     return F->getKind() == MCFragment::FT_Data;
   }
 };
 
 /// This is a compact (memory-size-wise) fragment for holding an encoded
 /// instruction (non-relaxable) that has no fixups registered. When applicable,

 /// A relaxable fragment holds on to its MCInst, since it may need to be
 /// relaxed during the assembler layout and relaxation stage.
 ///
 class MCRelaxableFragment : public MCEncodedFragmentWithFixups<8, 1> {
 
   /// Inst - The instruction this is a fragment for.
   MCInst Inst;
 
-  /// STI - The MCSubtargetInfo in effect when the instruction was encoded.
-  const MCSubtargetInfo &STI;
-
 public:
   MCRelaxableFragment(const MCInst &Inst, const MCSubtargetInfo &STI,
                       MCSection *Sec = nullptr)
       : MCEncodedFragmentWithFixups(FT_Relaxable, true, Sec),
-        Inst(Inst), STI(STI) {}
+        Inst(Inst) { this->STI = &STI; }
 
   const MCInst &getInst() const { return Inst; }
   void setInst(const MCInst &Value) { Inst = Value; }
 
-  const MCSubtargetInfo &getSubtargetInfo() { return STI; }
-
   static bool classof(const MCFragment *F) {
     return F->getKind() == MCFragment::FT_Relaxable;
   }
 };
 
 class MCAlignFragment : public MCFragment {
   /// Alignment - The alignment to ensure, in bytes.
   unsigned Alignment;

llvm/trunk/include/llvm/MC/MCObjectStreamer.h

void insert(MCFragment *F) {
     flushPendingLabels(F);
     MCSection *CurSection = getCurrentSectionOnly();
     CurSection->getFragmentList().insert(CurInsertionPoint, F);
     F->setParent(CurSection);
   }
 
   /// Get a data fragment to write into, creating a new one if the current
   /// fragment is not a data fragment.
-  MCDataFragment *getOrCreateDataFragment();
+  /// Optionally a \p STI can be passed in so that a new fragment is created
+  /// if the Subtarget differs from the current fragment.
+  MCDataFragment *getOrCreateDataFragment(const MCSubtargetInfo* STI = nullptr);
   MCPaddingFragment *getOrCreatePaddingFragment();
 
 protected:
   bool changeSectionImpl(MCSection *Section, const MCExpr *Subsection);
 
   /// If any labels have been emitted but not assigned fragments, ensure that
   /// they get assigned, either to F if possible or to a new data fragment.
   /// Optionally, it is also possible to provide an offset \p FOffset, which
public:
   void EmitCVFileChecksumOffsetDirective(unsigned FileNo) override;
   void EmitDTPRel32Value(const MCExpr *Value) override;
   void EmitDTPRel64Value(const MCExpr *Value) override;
   void EmitTPRel32Value(const MCExpr *Value) override;
   void EmitTPRel64Value(const MCExpr *Value) override;
   void EmitGPRel32Value(const MCExpr *Value) override;
   void EmitGPRel64Value(const MCExpr *Value) override;
   bool EmitRelocDirective(const MCExpr &Offset, StringRef Name,
-                          const MCExpr *Expr, SMLoc Loc) override;
+                          const MCExpr *Expr, SMLoc Loc,
+                          const MCSubtargetInfo &STI) override;
   using MCStreamer::emitFill;
   void emitFill(const MCExpr &NumBytes, uint64_t FillValue,
                 SMLoc Loc = SMLoc()) override;
   void emitFill(const MCExpr &NumValues, int64_t Size, int64_t Expr,
                 SMLoc Loc = SMLoc()) override;
   void EmitFileDirective(StringRef Filename) override;
 
   void FinishImpl() override;

llvm/trunk/include/llvm/MC/MCStreamer.h

public:
   MCSection *getAssociatedXDataSection(const MCSection *TextSec);
 
   virtual void EmitSyntaxDirective();
 
   /// Emit a .reloc directive.
   /// Returns true if the relocation could not be emitted because Name is not
   /// known.
   virtual bool EmitRelocDirective(const MCExpr &Offset, StringRef Name,
-                                  const MCExpr *Expr, SMLoc Loc) {
+                                  const MCExpr *Expr, SMLoc Loc,
+                                  const MCSubtargetInfo &STI) {
     return true;
   }
 
   /// Emit the given \p Instruction into the current section.
   /// PrintSchedInfo == true then schedul comment should be added to output
   virtual void EmitInstruction(const MCInst &Inst, const MCSubtargetInfo &STI,
                                bool PrintSchedInfo = false);
 

llvm/trunk/lib/MC/MCAsmStreamer.cpp

public:
   void EmitInstruction(const MCInst &Inst, const MCSubtargetInfo &STI,
                        bool PrintSchedInfo) override;
 
   void EmitBundleAlignMode(unsigned AlignPow2) override;
   void EmitBundleLock(bool AlignToEnd) override;
   void EmitBundleUnlock() override;
 
   bool EmitRelocDirective(const MCExpr &Offset, StringRef Name,
-                          const MCExpr *Expr, SMLoc Loc) override;
+                          const MCExpr *Expr, SMLoc Loc,
+                          const MCSubtargetInfo &STI) override;
 
   /// If this file is backed by an assembly streamer, this dumps the specified
   /// string in the output .s file. This capability is indicated by the
   /// hasRawTextSupport() predicate.
   void EmitRawTextImpl(StringRef String) override;
 
   void FinishImpl() override;
 };

 }
 
 void MCAsmStreamer::EmitBundleUnlock() {
   OS << "\t.bundle_unlock";
   EmitEOL();
 }
 
 bool MCAsmStreamer::EmitRelocDirective(const MCExpr &Offset, StringRef Name,
-                                       const MCExpr *Expr, SMLoc) {
+                                       const MCExpr *Expr, SMLoc,
+                                       const MCSubtargetInfo &STI) {
   OS << "\t.reloc ";
   Offset.print(OS, MAI);
   OS << ", " << Name;
   if (Expr) {
     OS << ", ";
     Expr->print(OS, MAI);
   }
   EmitEOL();

llvm/trunk/lib/MC/MCAssembler.cpp

for (MCFragment &Frag : Sec) {
       // being templated makes this tricky.
       if (isa<MCEncodedFragment>(&Frag) &&
           isa<MCCompactEncodedInstFragment>(&Frag))
         continue;
       if (!isa<MCEncodedFragment>(&Frag) && !isa<MCCVDefRangeFragment>(&Frag))
         continue;
       ArrayRef<MCFixup> Fixups;
       MutableArrayRef<char> Contents;
+      const MCSubtargetInfo *STI = nullptr;
       if (auto *FragWithFixups = dyn_cast<MCDataFragment>(&Frag)) {
         Fixups = FragWithFixups->getFixups();
         Contents = FragWithFixups->getContents();
+        STI = FragWithFixups->getSubtargetInfo();
+        assert(!FragWithFixups->hasInstructions() || STI != nullptr);
       } else if (auto *FragWithFixups = dyn_cast<MCRelaxableFragment>(&Frag)) {
         Fixups = FragWithFixups->getFixups();
         Contents = FragWithFixups->getContents();
+        STI = FragWithFixups->getSubtargetInfo();
+        assert(!FragWithFixups->hasInstructions() || STI != nullptr);
       } else if (auto *FragWithFixups = dyn_cast<MCCVDefRangeFragment>(&Frag)) {
         Fixups = FragWithFixups->getFixups();
         Contents = FragWithFixups->getContents();
       } else
         llvm_unreachable("Unknown fragment with fixups!");
       for (const MCFixup &Fixup : Fixups) {
         uint64_t FixedValue;
         bool IsResolved;
         MCValue Target;
         std::tie(Target, FixedValue, IsResolved) =
             handleFixup(Layout, Frag, Fixup);
         getBackend().applyFixup(*this, Fixup, Target, Contents, FixedValue,
-                                IsResolved);
+                                IsResolved, STI);
       }
     }
   }
 }
 
 void MCAssembler::Finish() {
   // Create the layout object.
   MCAsmLayout Layout(*this);

 }
 
 bool MCAssembler::fragmentNeedsRelaxation(const MCRelaxableFragment *F,
                                           const MCAsmLayout &Layout) const {
   assert(getBackendPtr() && "Expected assembler backend");
   // If this inst doesn't ever need relaxation, ignore it. This occurs when we
   // are intentionally pushing out inst fragments, or because we relaxed a
   // previous instruction to one that doesn't need relaxation.
-  if (!getBackend().mayNeedRelaxation(F->getInst()))
+  if (!getBackend().mayNeedRelaxation(F->getInst(), *F->getSubtargetInfo()))
     return false;
 
   for (const MCFixup &Fixup : F->getFixups())
     if (fixupNeedsRelaxation(Fixup, F, Layout))
       return true;
 
   return false;
 }
 
 bool MCAssembler::relaxInstruction(MCAsmLayout &Layout,
                                    MCRelaxableFragment &F) {
   assert(getEmitterPtr() &&
          "Expected CodeEmitter defined for relaxInstruction");
   if (!fragmentNeedsRelaxation(&F, Layout))
     return false;
 
   ++stats::RelaxedInstructions;
 
   // FIXME-PERF: We could immediately lower out instructions if we can tell
   // they are fully resolved, to avoid retesting on later passes.
 
   // Relax the fragment.
 
   MCInst Relaxed;
-  getBackend().relaxInstruction(F.getInst(), F.getSubtargetInfo(), Relaxed);
+  getBackend().relaxInstruction(F.getInst(), *F.getSubtargetInfo(), Relaxed);
 
   // Encode the new instruction.
   //
   // FIXME-PERF: If it matters, we could let the target do this. It can
   // probably do so more efficiently in many cases.
   SmallVector<MCFixup, 4> Fixups;
   SmallString<256> Code;
   raw_svector_ostream VecOS(Code);
-  getEmitter().encodeInstruction(Relaxed, VecOS, Fixups, F.getSubtargetInfo());
+  getEmitter().encodeInstruction(Relaxed, VecOS, Fixups, *F.getSubtargetInfo());
 
   // Update the fragment.
   F.setInst(Relaxed);
   F.getContents() = Code;
   F.getFixups() = Fixups;
 
   return true;
 }

llvm/trunk/lib/MC/MCELFStreamer.cpp

void MCELFStreamer::mergeFragment(MCDataFragment *DF,
 
   flushPendingLabels(DF, DF->getContents().size());
 
   for (unsigned i = 0, e = EF->getFixups().size(); i != e; ++i) {
     EF->getFixups()[i].setOffset(EF->getFixups()[i].getOffset() +
                                  DF->getContents().size());
     DF->getFixups().push_back(EF->getFixups()[i]);
   }
-  DF->setHasInstructions(true);
+  if (DF->getSubtargetInfo() == nullptr && EF->getSubtargetInfo())
+    DF->setHasInstructions(*EF->getSubtargetInfo());
   DF->getContents().append(EF->getContents().begin(), EF->getContents().end());
 }
 
 void MCELFStreamer::InitSections(bool NoExecStack) {
   MCContext &Ctx = getContext();
   SwitchSection(Ctx.getObjectFileInfo()->getTextSection());
   EmitCodeAlignment(4);
 
void MCELFStreamer::EmitInstToFragment(const MCInst &Inst,
                                        const MCSubtargetInfo &STI) {
   this->MCObjectStreamer::EmitInstToFragment(Inst, STI);
   MCRelaxableFragment &F = *cast<MCRelaxableFragment>(getCurrentFragment());
 
   for (unsigned i = 0, e = F.getFixups().size(); i != e; ++i)
     fixSymbolsInTLSFixups(F.getFixups()[i].getValue());
 }
 
+// A fragment can only have one Subtarget, and when bundling is enabled we
+// sometimes need to use the same fragment. We give an error if there
+// are conflicting Subtargets.
+static void CheckBundleSubtargets(const MCSubtargetInfo *OldSTI,
+                                  const MCSubtargetInfo *NewSTI) {
+  if (OldSTI && NewSTI && OldSTI != NewSTI)
+    report_fatal_error("A Bundle can only have one Subtarget.");
+}
+
 void MCELFStreamer::EmitInstToData(const MCInst &Inst,
                                    const MCSubtargetInfo &STI) {
   MCAssembler &Assembler = getAssembler();
   SmallVector<MCFixup, 4> Fixups;
   SmallString<256> Code;
   raw_svector_ostream VecOS(Code);
   Assembler.getEmitter().encodeInstruction(Inst, VecOS, Fixups, STI);
 
   for (unsigned i = 0, e = Fixups.size(); i != e; ++i)
     fixSymbolsInTLSFixups(Fixups[i].getValue());
 
   // There are several possibilities here:
   //
   // If bundling is disabled, append the encoded instruction to the current data
   // fragment (or create a new such fragment if the current fragment is not a
-  // data fragment).
+  // data fragment, or the Subtarget has changed).
   //
   // If bundling is enabled:
   // - If we're not in a bundle-locked group, emit the instruction into a
   //   fragment of its own. If there are no fixups registered for the
   //   instruction, emit a MCCompactEncodedInstFragment. Otherwise, emit a
   //   MCDataFragment.
   // - If we're in a bundle-locked group, append the instruction to the current
   //   data fragment because we want all the instructions in a group to get into
   //   the same fragment. Be careful not to do that for the first instruction in
   //   the group, though.
   MCDataFragment *DF;
 
   if (Assembler.isBundlingEnabled()) {
     MCSection &Sec = *getCurrentSectionOnly();
-    if (Assembler.getRelaxAll() && isBundleLocked())
+    if (Assembler.getRelaxAll() && isBundleLocked()) {
       // If the -mc-relax-all flag is used and we are bundle-locked, we re-use
       // the current bundle group.
       DF = BundleGroups.back();
+      CheckBundleSubtargets(DF->getSubtargetInfo(), &STI);
+    }
     else if (Assembler.getRelaxAll() && !isBundleLocked())
       // When not in a bundle-locked group and the -mc-relax-all flag is used,
       // we create a new temporary fragment which will be later merged into
       // the current fragment.
       DF = new MCDataFragment();
-    else if (isBundleLocked() && !Sec.isBundleGroupBeforeFirstInst())
+    else if (isBundleLocked() && !Sec.isBundleGroupBeforeFirstInst()) {
       // If we are bundle-locked, we re-use the current fragment.
       // The bundle-locking directive ensures this is a new data fragment.
       DF = cast<MCDataFragment>(getCurrentFragment());
+      CheckBundleSubtargets(DF->getSubtargetInfo(), &STI);
+    }
     else if (!isBundleLocked() && Fixups.size() == 0) {
       // Optimize memory usage by emitting the instruction to a
       // MCCompactEncodedInstFragment when not in a bundle-locked group and
       // there are no fixups registered.
       MCCompactEncodedInstFragment *CEIF = new MCCompactEncodedInstFragment();
       insert(CEIF);
       CEIF->getContents().append(Code.begin(), Code.end());
       return;
     } else {
       DF = new MCDataFragment();
       insert(DF);
     }
     if (Sec.getBundleLockState() == MCSection::BundleLockedAlignToEnd) {
       // If this fragment is for a group marked "align_to_end", set a flag
       // in the fragment. This can happen after the fragment has already been
       // created if there are nested bundle_align groups and an inner one
       // is the one marked align_to_end.
       DF->setAlignToBundleEnd(true);
     }
 
     // We're now emitting an instruction in a bundle group, so this flag has
     // to be turned off.
     Sec.setBundleGroupBeforeFirstInst(false);
   } else {
-    DF = getOrCreateDataFragment();
+    DF = getOrCreateDataFragment(&STI);
   }
 
   // Add the fixups and data.
   for (unsigned i = 0, e = Fixups.size(); i != e; ++i) {
     Fixups[i].setOffset(Fixups[i].getOffset() + DF->getContents().size());
     DF->getFixups().push_back(Fixups[i]);
   }
-  DF->setHasInstructions(true);
+  DF->setHasInstructions(STI);
   DF->getContents().append(Code.begin(), Code.end());
 
   if (Assembler.isBundlingEnabled() && Assembler.getRelaxAll()) {
     if (!isBundleLocked()) {
-      mergeFragment(getOrCreateDataFragment(), DF);
+      mergeFragment(getOrCreateDataFragment(&STI), DF);
       delete DF;
     }
   }
 }
 
 void MCELFStreamer::EmitBundleAlignMode(unsigned AlignPow2) {
   assert(AlignPow2 <= 30 && "Invalid bundle alignment");
   MCAssembler &Assembler = getAssembler();
if (getAssembler().getRelaxAll()) {
     assert(!BundleGroups.empty() && "There are no bundle groups");
     MCDataFragment *DF = BundleGroups.back();
 
     // FIXME: Use BundleGroups to track the lock state instead.
     Sec.setBundleLockState(MCSection::NotBundleLocked);
 
     // FIXME: Use more separate fragments for nested groups.
     if (!isBundleLocked()) {
-      mergeFragment(getOrCreateDataFragment(), DF);
+      mergeFragment(getOrCreateDataFragment(DF->getSubtargetInfo()), DF);
       BundleGroups.pop_back();
       delete DF;
     }
 
     if (Sec.getBundleLockState() != MCSection::BundleLockedAlignToEnd)
       getOrCreateDataFragment()->setAlignToBundleEnd(false);
   } else
     Sec.setBundleLockState(MCSection::NotBundleLocked);

llvm/trunk/lib/MC/MCMachOStreamer.cpp

void MCMachOStreamer::EmitInstToData(const MCInst &Inst,
   raw_svector_ostream VecOS(Code);
   getAssembler().getEmitter().encodeInstruction(Inst, VecOS, Fixups, STI);
 
   // Add the fixups and data.
   for (MCFixup &Fixup : Fixups) {
     Fixup.setOffset(Fixup.getOffset() + DF->getContents().size());
     DF->getFixups().push_back(Fixup);
   }
+  DF->setHasInstructions(STI);
   DF->getContents().append(Code.begin(), Code.end());
 }
 
 void MCMachOStreamer::FinishImpl() {
   EmitFrames(&getAssembler().getBackend());
 
   // We have to set the fragment atom associations so we can relax properly for
   // Mach-O.

llvm/trunk/lib/MC/MCObjectStreamer.cpp

MCFragment *MCObjectStreamer::getCurrentFragment() const {
   assert(getCurrentSectionOnly() && "No current section!");
 
   if (CurInsertionPoint != getCurrentSectionOnly()->getFragmentList().begin())
     return &*std::prev(CurInsertionPoint);
 
   return nullptr;
 }
 
-MCDataFragment *MCObjectStreamer::getOrCreateDataFragment() {
-  MCDataFragment *F = dyn_cast_or_null<MCDataFragment>(getCurrentFragment());
+static bool CanReuseDataFragment(const MCDataFragment &F,
+                                 const MCAssembler &Assembler,
+                                 const MCSubtargetInfo *STI) {
+  if (!F.hasInstructions())
+    return true;
   // When bundling is enabled, we don't want to add data to a fragment that
   // already has instructions (see MCELFStreamer::EmitInstToData for details)
-  if (!F || (Assembler->isBundlingEnabled() && !Assembler->getRelaxAll() &&
-             F->hasInstructions())) {
+  if (Assembler.isBundlingEnabled())
+    return Assembler.getRelaxAll();
+  // If the subtarget is changed mid fragment we start a new fragment to record
+  // the new STI.
+  return !STI || F.getSubtargetInfo() == STI;
+}
+
+MCDataFragment *
+MCObjectStreamer::getOrCreateDataFragment(const MCSubtargetInfo *STI) {
+  MCDataFragment *F = dyn_cast_or_null<MCDataFragment>(getCurrentFragment());
+  if (!F || !CanReuseDataFragment(*F, *Assembler, STI)) {
     F = new MCDataFragment();
     insert(F);
   }
   return F;
 }
 
 MCPaddingFragment *MCObjectStreamer::getOrCreatePaddingFragment() {
   MCPaddingFragment *F =
void MCObjectStreamer::EmitInstructionImpl(const MCInst &Inst,
 
   // Now that a machine instruction has been assembled into this section, make
   // a line entry for any .loc directive that has been seen.
   MCCVLineEntry::Make(this);
   MCDwarfLineEntry::Make(this, getCurrentSectionOnly());
 
   // If this instruction doesn't need relaxation, just emit it as data.
   MCAssembler &Assembler = getAssembler();
-  if (!Assembler.getBackend().mayNeedRelaxation(Inst)) {
+  if (!Assembler.getBackend().mayNeedRelaxation(Inst, STI)) {
     EmitInstToData(Inst, STI);
     return;
   }
 
   // Otherwise, relax and emit it as data if either:
   // - The RelaxAll flag was passed
   // - Bundling is enabled and this instruction is inside a bundle-locked
   //   group. We want to emit all such instructions into the same data
   //   fragment.
   if (Assembler.getRelaxAll() ||
       (Assembler.isBundlingEnabled() && Sec->isBundleLocked())) {
     MCInst Relaxed;
     getAssembler().getBackend().relaxInstruction(Inst, STI, Relaxed);
-    while (getAssembler().getBackend().mayNeedRelaxation(Relaxed))
+    while (getAssembler().getBackend().mayNeedRelaxation(Relaxed, STI))
       getAssembler().getBackend().relaxInstruction(Relaxed, STI, Relaxed);
     EmitInstToData(Relaxed, STI);
     return;
   }
 
   // Otherwise emit to a separate fragment.
   EmitInstToFragment(Inst, STI);
 }
void MCObjectStreamer::EmitGPRel64Value(const MCExpr *Value) {
   flushPendingLabels(DF, DF->getContents().size());
 
   DF->getFixups().push_back(
       MCFixup::create(DF->getContents().size(), Value, FK_GPRel_4));
   DF->getContents().resize(DF->getContents().size() + 8, 0);
 }
 
 bool MCObjectStreamer::EmitRelocDirective(const MCExpr &Offset, StringRef Name,
-                                          const MCExpr *Expr, SMLoc Loc) {
+                                          const MCExpr *Expr, SMLoc Loc,
+                                          const MCSubtargetInfo &STI) {
   int64_t OffsetValue;
   if (!Offset.evaluateAsAbsolute(OffsetValue))
     llvm_unreachable("Offset is not absolute");
 
   if (OffsetValue < 0)
     llvm_unreachable("Offset is negative");
 
-  MCDataFragment *DF = getOrCreateDataFragment();
+  MCDataFragment *DF = getOrCreateDataFragment(&STI);
   flushPendingLabels(DF, DF->getContents().size());
 
   Optional<MCFixupKind> MaybeKind = Assembler->getBackend().getFixupKind(Name);
   if (!MaybeKind.hasValue())
     return true;
 
   MCFixupKind Kind = *MaybeKind;
 

llvm/trunk/lib/MC/MCParser/AsmParser.cpp

if (Lexer.is(AsmToken::Comma)) {
     if (!Expr->evaluateAsRelocatable(Value, nullptr, nullptr))
       return Error(ExprLoc, "expression must be relocatable");
   }
 
   if (parseToken(AsmToken::EndOfStatement,
                  "unexpected token in .reloc directive"))
       return true;
 
-  if (getStreamer().EmitRelocDirective(*Offset, Name, Expr, DirectiveLoc))
+  const MCTargetAsmParser &MCT = getTargetParser();
+  const MCSubtargetInfo &STI = MCT.getSTI();
+  if (getStreamer().EmitRelocDirective(*Offset, Name, Expr, DirectiveLoc, STI))
     return Error(NameLoc, "unknown relocation name");
 
   return false;
 }
 
 /// parseDirectiveValue
 ///  ::= (.byte | .short | ... ) [ expression (, expression)* ]
 bool AsmParser::parseDirectiveValue(StringRef IDVal, unsigned Size) {

llvm/trunk/lib/MC/MCWasmStreamer.cpp

 void MCWasmStreamer::mergeFragment(MCDataFragment *DF, MCDataFragment *EF) {
   flushPendingLabels(DF, DF->getContents().size());
 
   for (unsigned i = 0, e = EF->getFixups().size(); i != e; ++i) {
     EF->getFixups()[i].setOffset(EF->getFixups()[i].getOffset() +
                                  DF->getContents().size());
     DF->getFixups().push_back(EF->getFixups()[i]);
   }
-  DF->setHasInstructions(true);
+  if (DF->getSubtargetInfo() == nullptr && EF->getSubtargetInfo())
+    DF->setHasInstructions(*EF->getSubtargetInfo());
   DF->getContents().append(EF->getContents().begin(), EF->getContents().end());
 }
 
 void MCWasmStreamer::EmitAssemblerFlag(MCAssemblerFlag Flag) {
   // Let the target do whatever target specific stuff it needs to do.
   getAssembler().getBackend().handleAssemblerFlag(Flag);
 
   // Do any generic stuff we need to do.
void MCWasmStreamer::EmitInstToData(const MCInst &Inst,
   // new such fragment if the current fragment is not a data fragment).
   MCDataFragment *DF = getOrCreateDataFragment();
 
   // Add the fixups and data.
   for (unsigned i = 0, e = Fixups.size(); i != e; ++i) {
     Fixups[i].setOffset(Fixups[i].getOffset() + DF->getContents().size());
     DF->getFixups().push_back(Fixups[i]);
   }
-  DF->setHasInstructions(true);
+  DF->setHasInstructions(STI);
   DF->getContents().append(Code.begin(), Code.end());
 }
 
 void MCWasmStreamer::FinishImpl() {
   EmitFrames(nullptr);
 
   // Set fragment atoms so we can map from code fragment to defining symbol
   addFragmentAtoms();

llvm/trunk/lib/MC/MCWinCOFFStreamer.cpp

void MCWinCOFFStreamer::EmitInstToData(const MCInst &Inst,
   raw_svector_ostream VecOS(Code);
   getAssembler().getEmitter().encodeInstruction(Inst, VecOS, Fixups, STI);
 
   // Add the fixups and data.
   for (unsigned i = 0, e = Fixups.size(); i != e; ++i) {
     Fixups[i].setOffset(Fixups[i].getOffset() + DF->getContents().size());
     DF->getFixups().push_back(Fixups[i]);
   }
-
+  DF->setHasInstructions(STI);
   DF->getContents().append(Code.begin(), Code.end());
 }
 
 void MCWinCOFFStreamer::InitSections(bool NoExecStack) {
   // FIXME: this is identical to the ELF one.
   // This emulates the same behavior of GNU as. This makes it easier
   // to compare the output as the major sections are in the same order.
   SwitchSection(getContext().getObjectFileInfo()->getTextSection());

llvm/trunk/lib/Target/AArch64/MCTargetDesc/AArch64AsmBackend.cpp

const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const override {
 
     assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
            "Invalid kind!");
     return Infos[Kind - FirstTargetFixupKind];
   }
 
   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                   const MCValue &Target, MutableArrayRef<char> Data,
-                  uint64_t Value, bool IsResolved) const override;
+                  uint64_t Value, bool IsResolved,
+                  const MCSubtargetInfo *STI) const override;
 
-  bool mayNeedRelaxation(const MCInst &Inst) const override;
+  bool mayNeedRelaxation(const MCInst &Inst,
+                         const MCSubtargetInfo &STI) const override;
   bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
                             const MCRelaxableFragment *DF,
                             const MCAsmLayout &Layout) const override;
   void relaxInstruction(const MCInst &Inst, const MCSubtargetInfo &STI,
                         MCInst &Res) const override;
   bool writeNopData(raw_ostream &OS, uint64_t Count) const override;
 
   void HandleAssemblerFlag(MCAssemblerFlag Flag) {}
case AArch64::fixup_aarch64_pcrel_call26:
     // Instructions are always little endian
     return 0;
   }
 }
 
 void AArch64AsmBackend::applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                                    const MCValue &Target,
                                    MutableArrayRef<char> Data, uint64_t Value,
-                                   bool IsResolved) const {
+                                   bool IsResolved,
+                                   const MCSubtargetInfo *STI) const {
   unsigned NumBytes = getFixupKindNumBytes(Fixup.getKind());
   if (!Value)
     return; // Doesn't change encoding.
   MCFixupKindInfo Info = getFixupKindInfo(Fixup.getKind());
   MCContext &Ctx = Asm.getContext();
   // Apply any target-specific value adjustments.
   Value = adjustFixupValue(Fixup, Value, Ctx, TheTriple, IsResolved);
 
if (FulleSizeInBytes == 0) {
     assert(NumBytes <= FulleSizeInBytes && "Invalid fixup size!");
     for (unsigned i = 0; i != NumBytes; ++i) {
       unsigned Idx = FulleSizeInBytes - 1 - i;
       Data[Offset + Idx] |= uint8_t((Value >> (i * 8)) & 0xff);
     }
   }
 }
 
-bool AArch64AsmBackend::mayNeedRelaxation(const MCInst &Inst) const {
+bool AArch64AsmBackend::mayNeedRelaxation(const MCInst &Inst,
+                                          const MCSubtargetInfo &STI) const {
   return false;
 }
 
 bool AArch64AsmBackend::fixupNeedsRelaxation(const MCFixup &Fixup,
                                              uint64_t Value,
                                              const MCRelaxableFragment *DF,
                                              const MCAsmLayout &Layout) const {
   // FIXME:  This isn't correct for AArch64. Just moving the "generic" logic

llvm/trunk/lib/Target/AMDGPU/MCTargetDesc/AMDGPUAsmBackend.cpp

 class AMDGPUAsmBackend : public MCAsmBackend {
 public:
   AMDGPUAsmBackend(const Target &T) : MCAsmBackend(support::little) {}
 
   unsigned getNumFixupKinds() const override { return AMDGPU::NumTargetFixupKinds; };
 
   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                   const MCValue &Target, MutableArrayRef<char> Data,
-                  uint64_t Value, bool IsResolved) const override;
+                  uint64_t Value, bool IsResolved,
+                  const MCSubtargetInfo *STI) const override;
   bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
                             const MCRelaxableFragment *DF,
                             const MCAsmLayout &Layout) const override {
     return false;
   }
   void relaxInstruction(const MCInst &Inst, const MCSubtargetInfo &STI,
                         MCInst &Res) const override {
     llvm_unreachable("Not implemented");
   }
-  bool mayNeedRelaxation(const MCInst &Inst) const override { return false; }
+  bool mayNeedRelaxation(const MCInst &Inst,
+                         const MCSubtargetInfo &STI) const override {
+    return false;
+  }
 
   unsigned getMinimumNopSize() const override;
   bool writeNopData(raw_ostream &OS, uint64_t Count) const override;
 
   const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const override;
 };
 
 } //End anonymous namespace
static uint64_t adjustFixupValue(const MCFixup &Fixup, uint64_t Value,
   default:
     llvm_unreachable("unhandled fixup kind");
   }
 }
 
 void AMDGPUAsmBackend::applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                                   const MCValue &Target,
                                   MutableArrayRef<char> Data, uint64_t Value,
-                                  bool IsResolved) const {
+                                  bool IsResolved,
+                                  const MCSubtargetInfo *STI) const {
   Value = adjustFixupValue(Fixup, Value, &Asm.getContext());
   if (!Value)
     return; // Doesn't change encoding.
 
   MCFixupKindInfo Info = getFixupKindInfo(Fixup.getKind());
 
   // Shift the value into position.
   Value <<= Info.TargetOffset;

llvm/trunk/lib/Target/ARM/MCTargetDesc/ARMAsmBackend.h

 #include "MCTargetDesc/ARMMCTargetDesc.h"
 #include "llvm/MC/MCAsmBackend.h"
 #include "llvm/MC/MCSubtargetInfo.h"
 #include "llvm/Support/TargetRegistry.h"
 
 namespace llvm {
 
 class ARMAsmBackend : public MCAsmBackend {
+  // The STI from the target triple the MCAsmBackend was instantiated with
+  // note that MCFragments may have a different local STI that should be
+  // used in preference.
   const MCSubtargetInfo &STI;
   bool isThumbMode;    // Currently emitting Thumb code.
 public:
   ARMAsmBackend(const Target &T, const MCSubtargetInfo &STI,
                 support::endianness Endian)
       : MCAsmBackend(Endian), STI(STI),
         isThumbMode(STI.getTargetTriple().isThumb()) {}
 
   unsigned getNumFixupKinds() const override {
     return ARM::NumTargetFixupKinds;
   }
 
+  // FIXME: this should be calculated per fragment as the STI may be
+  // different.
   bool hasNOP() const { return STI.getFeatureBits()[ARM::HasV6T2Ops]; }
 
   const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const override;
 
   bool shouldForceRelocation(const MCAssembler &Asm, const MCFixup &Fixup,
                              const MCValue &Target) override;
 
   unsigned adjustFixupValue(const MCAssembler &Asm, const MCFixup &Fixup,
                             const MCValue &Target, uint64_t Value,
-                            bool IsResolved, MCContext &Ctx) const;
+                            bool IsResolved, MCContext &Ctx,
+                            const MCSubtargetInfo *STI) const;
 
   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                   const MCValue &Target, MutableArrayRef<char> Data,
-                  uint64_t Value, bool IsResolved) const override;
+                  uint64_t Value, bool IsResolved,
+                  const MCSubtargetInfo *STI) const override;
 
-  unsigned getRelaxedOpcode(unsigned Op) const;
+  unsigned getRelaxedOpcode(unsigned Op, const MCSubtargetInfo &STI) const;
 
-  bool mayNeedRelaxation(const MCInst &Inst) const override;
+  bool mayNeedRelaxation(const MCInst &Inst,
+                         const MCSubtargetInfo &STI) const override;
 
   const char *reasonForFixupRelaxation(const MCFixup &Fixup,
                                        uint64_t Value) const;
 
   bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
                             const MCRelaxableFragment *DF,
                             const MCAsmLayout &Layout) const override;
 

llvm/trunk/lib/Target/ARM/MCTargetDesc/ARMAsmBackend.cpp

case MCAF_Code16:
     setIsThumb(true);
     break;
   case MCAF_Code32:
     setIsThumb(false);
     break;
   }
 }
 
-unsigned ARMAsmBackend::getRelaxedOpcode(unsigned Op) const {
+unsigned ARMAsmBackend::getRelaxedOpcode(unsigned Op,
+                                         const MCSubtargetInfo &STI) const {
   bool HasThumb2 = STI.getFeatureBits()[ARM::FeatureThumb2];
   bool HasV8MBaselineOps = STI.getFeatureBits()[ARM::HasV8MBaselineOps];
 
   switch (Op) {
   default:
     return Op;
   case ARM::tBcc:
     return HasThumb2 ? (unsigned)ARM::t2Bcc : Op;
   case ARM::tLDRpci:
     return HasThumb2 ? (unsigned)ARM::t2LDRpci : Op;
   case ARM::tADR:
     return HasThumb2 ? (unsigned)ARM::t2ADR : Op;
   case ARM::tB:
     return HasV8MBaselineOps ? (unsigned)ARM::t2B : Op;
   case ARM::tCBZ:
     return ARM::tHINT;
   case ARM::tCBNZ:
     return ARM::tHINT;
   }
 }
 
-bool ARMAsmBackend::mayNeedRelaxation(const MCInst &Inst) const {
-  if (getRelaxedOpcode(Inst.getOpcode()) != Inst.getOpcode())
+bool ARMAsmBackend::mayNeedRelaxation(const MCInst &Inst,
+                                      const MCSubtargetInfo &STI) const {
+  if (getRelaxedOpcode(Inst.getOpcode(), STI) != Inst.getOpcode())
     return true;
   return false;
 }
 
 const char *ARMAsmBackend::reasonForFixupRelaxation(const MCFixup &Fixup,
                                                     uint64_t Value) const {
   switch ((unsigned)Fixup.getKind()) {
   case ARM::fixup_arm_thumb_br: {
bool ARMAsmBackend::fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
                                          const MCRelaxableFragment *DF,
                                          const MCAsmLayout &Layout) const {
   return reasonForFixupRelaxation(Fixup, Value);
 }
 
 void ARMAsmBackend::relaxInstruction(const MCInst &Inst,
                                      const MCSubtargetInfo &STI,
                                      MCInst &Res) const {
-  unsigned RelaxedOp = getRelaxedOpcode(Inst.getOpcode());
+  unsigned RelaxedOp = getRelaxedOpcode(Inst.getOpcode(), STI);
 
   // Sanity check w/ diagnostic if we get here w/ a bogus instruction.
   if (RelaxedOp == Inst.getOpcode()) {
     SmallString<256> Tmp;
     raw_svector_ostream OS(Tmp);
     Inst.dump_pretty(OS);
     OS << "\n";
     report_fatal_error("unexpected instruction to relax: " + OS.str());
static uint32_t joinHalfWords(uint32_t FirstHalf, uint32_t SecondHalf,
   }
 
   return Value;
 }
 
 unsigned ARMAsmBackend::adjustFixupValue(const MCAssembler &Asm,
                                          const MCFixup &Fixup,
                                          const MCValue &Target, uint64_t Value,
-                                         bool IsResolved, MCContext &Ctx) const {
+                                         bool IsResolved, MCContext &Ctx,
+                                         const MCSubtargetInfo* STI) const {
   unsigned Kind = Fixup.getKind();
 
   // MachO tries to make .o files that look vaguely pre-linked, so for MOVW/MOVT
   // and .word relocations they put the Thumb bit into the addend if possible.
   // Other relocation types don't want this bit though (branches couldn't encode
   // it if it *was* present, and no other relocations exist) and it can
   // interfere with checking valid expressions.
   if (const MCSymbolRefExpr *A = Target.getSymA()) {
unsigned ARMAsmBackend::adjustFixupValue(const MCAssembler &Asm,
   case FK_Data_2:
   case FK_Data_4:
     return Value;
   case FK_SecRel_2:
     return Value;
   case FK_SecRel_4:
     return Value;
   case ARM::fixup_arm_movt_hi16:
-    if (IsResolved || !STI.getTargetTriple().isOSBinFormatELF())
+    assert(STI != nullptr);
+    if (IsResolved || !STI->getTargetTriple().isOSBinFormatELF())
       Value >>= 16;
     LLVM_FALLTHROUGH;
   case ARM::fixup_arm_movw_lo16: {
     unsigned Hi4 = (Value & 0xF000) >> 12;
     unsigned Lo12 = Value & 0x0FFF;
     // inst{19-16} = Hi4;
     // inst{11-0} = Lo12;
     Value = (Hi4 << 16) | (Lo12);
     return Value;
   }
   case ARM::fixup_t2_movt_hi16:
-    if (IsResolved || !STI.getTargetTriple().isOSBinFormatELF())
+    assert(STI != nullptr);
+    if (IsResolved || !STI->getTargetTriple().isOSBinFormatELF())
       Value >>= 16;
     LLVM_FALLTHROUGH;
   case ARM::fixup_t2_movw_lo16: {
     unsigned Hi4 = (Value & 0xF000) >> 12;
     unsigned i = (Value & 0x800) >> 11;
     unsigned Mid3 = (Value & 0x700) >> 8;
     unsigned Lo8 = Value & 0x0FF;
     // inst{19-16} = Hi4;
case ARM::fixup_t2_condbranch: {
     out |= (Value & 0x20000) >> 4; // J1 bit
     out |= (Value & 0x1F800) << 5; // imm6 field
     out |= (Value & 0x007FF);      // imm11 field
 
     return swapHalfWords(out, Endian == support::little);
   }
   case ARM::fixup_arm_thumb_bl: {
     if (!isInt<25>(Value - 4) ||
-        (!STI.getFeatureBits()[ARM::FeatureThumb2] &&
-         !STI.getFeatureBits()[ARM::HasV8MBaselineOps] &&
-         !STI.getFeatureBits()[ARM::HasV6MOps] &&
+        (!STI->getFeatureBits()[ARM::FeatureThumb2] &&
+         !STI->getFeatureBits()[ARM::HasV8MBaselineOps] &&
+         !STI->getFeatureBits()[ARM::HasV6MOps] &&
          !isInt<23>(Value - 4))) {
       Ctx.reportError(Fixup.getLoc(), "Relocation out of range");
       return 0;
     }
 
     // The value doesn't encode the low bit (always zero) and is offset by
     // four. The 32-bit immediate value is encoded as
     //   imm32 = SignExtend(S:I1:I2:imm10:imm11:0)
case ARM::fixup_arm_thumb_blx: {
     uint32_t SecondHalf = (((uint16_t)J1Bit << 13) | ((uint16_t)J2Bit << 11) |
                            ((uint16_t)imm10LBits) << 1);
     return joinHalfWords(FirstHalf, SecondHalf, Endian == support::little);
   }
   case ARM::fixup_thumb_adr_pcrel_10:
   case ARM::fixup_arm_thumb_cp:
     // On CPUs supporting Thumb2, this will be relaxed to an ldr.w, otherwise we
     // could have an error on our hands.
-    if (!STI.getFeatureBits()[ARM::FeatureThumb2] && IsResolved) {
+    assert(STI != nullptr);
+    if (!STI->getFeatureBits()[ARM::FeatureThumb2] && IsResolved) {
       const char *FixupDiagnostic = reasonForFixupRelaxation(Fixup, Value);
       if (FixupDiagnostic) {
         Ctx.reportError(Fixup.getLoc(), FixupDiagnostic);
         return 0;
       }
     }
     // Offset by 4, and don't encode the low two bits.
     return ((Value - 4) >> 2) & 0xff;
   case ARM::fixup_arm_thumb_cb: {
     // CB instructions can only branch to offsets in [4, 126] in multiples of 2
     // so ensure that the raw value LSB is zero and it lies in [2, 130].
     // An offset of 2 will be relaxed to a NOP.
     if ((int64_t)Value < 2 || Value > 0x82 || Value & 1) {
       Ctx.reportError(Fixup.getLoc(), "out of range pc-relative fixup value");
       return 0;
     }
     // Offset by 4 and don't encode the lower bit, which is always 0.
     // FIXME: diagnose if no Thumb2
     uint32_t Binary = (Value - 4) >> 1;
     return ((Binary & 0x20) << 4) | ((Binary & 0x1f) << 3);
   }
   case ARM::fixup_arm_thumb_br:
     // Offset by 4 and don't encode the lower bit, which is always 0.
-    if (!STI.getFeatureBits()[ARM::FeatureThumb2] &&
-        !STI.getFeatureBits()[ARM::HasV8MBaselineOps]) {
+    assert(STI != nullptr);
+    if (!STI->getFeatureBits()[ARM::FeatureThumb2] &&
+        !STI->getFeatureBits()[ARM::HasV8MBaselineOps]) {
       const char *FixupDiagnostic = reasonForFixupRelaxation(Fixup, Value);
       if (FixupDiagnostic) {
         Ctx.reportError(Fixup.getLoc(), FixupDiagnostic);
         return 0;
       }
     }
     return ((Value - 4) >> 1) & 0x7ff;
   case ARM::fixup_arm_thumb_bcc:
     // Offset by 4 and don't encode the lower bit, which is always 0.
-    if (!STI.getFeatureBits()[ARM::FeatureThumb2]) {
+    assert(STI != nullptr);
+    if (!STI->getFeatureBits()[ARM::FeatureThumb2]) {
       const char *FixupDiagnostic = reasonForFixupRelaxation(Fixup, Value);
       if (FixupDiagnostic) {
         Ctx.reportError(Fixup.getLoc(), FixupDiagnostic);
         return 0;
       }
     }
     return ((Value - 4) >> 1) & 0xff;
   case ARM::fixup_arm_pcrel_10_unscaled: {
case ARM::fixup_t2_so_imm:
     // Instruction size is 4 bytes.
     return 4;
   }
 }
 
 void ARMAsmBackend::applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                                const MCValue &Target,
                                MutableArrayRef<char> Data, uint64_t Value,
-                               bool IsResolved) const {
+                               bool IsResolved,
+                               const MCSubtargetInfo* STI) const {
   unsigned NumBytes = getFixupKindNumBytes(Fixup.getKind());
   MCContext &Ctx = Asm.getContext();
-  Value = adjustFixupValue(Asm, Fixup, Target, Value, IsResolved, Ctx);
+  Value = adjustFixupValue(Asm, Fixup, Target, Value, IsResolved, Ctx, STI);
   if (!Value)
     return; // Doesn't change encoding.
 
   unsigned Offset = Fixup.getOffset();
   assert(Offset + NumBytes <= Data.size() && "Invalid fixup offset!");
 
   // Used to point to big endian bytes.
   unsigned FullSizeBytes;

llvm/trunk/lib/Target/BPF/MCTargetDesc/BPFAsmBackend.cpp

 
 class BPFAsmBackend : public MCAsmBackend {
 public:
   BPFAsmBackend(support::endianness Endian) : MCAsmBackend(Endian) {}
   ~BPFAsmBackend() override = default;
 
   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                   const MCValue &Target, MutableArrayRef<char> Data,
-                  uint64_t Value, bool IsResolved) const override;
+                  uint64_t Value, bool IsResolved,
+                  const MCSubtargetInfo *STI) const override;
 
   std::unique_ptr<MCObjectTargetWriter>
   createObjectTargetWriter() const override;
 
   // No instruction requires relaxation
   bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
                             const MCRelaxableFragment *DF,
                             const MCAsmLayout &Layout) const override {
     return false;
   }
 
   unsigned getNumFixupKinds() const override { return 1; }
 
-  bool mayNeedRelaxation(const MCInst &Inst) const override { return false; }
+  bool mayNeedRelaxation(const MCInst &Inst,
+                         const MCSubtargetInfo &STI) const override {
+    return false;
+  }
 
   void relaxInstruction(const MCInst &Inst, const MCSubtargetInfo &STI,
                         MCInst &Res) const override {}
 
   bool writeNopData(raw_ostream &OS, uint64_t Count) const override;
 };
 
 } // end anonymous namespace
 
 bool BPFAsmBackend::writeNopData(raw_ostream &OS, uint64_t Count) const {
   if ((Count % 8) != 0)
     return false;
 
   for (uint64_t i = 0; i < Count; i += 8)
     support::endian::write<uint64_t>(OS, 0x15000000, Endian);
 
   return true;
 }
 
 void BPFAsmBackend::applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                                const MCValue &Target,
                                MutableArrayRef<char> Data, uint64_t Value,
-                               bool IsResolved) const {
+                               bool IsResolved,
+                               const MCSubtargetInfo *STI) const {
   if (Fixup.getKind() == FK_SecRel_4 || Fixup.getKind() == FK_SecRel_8) {
     assert(Value == 0);
   } else if (Fixup.getKind() == FK_Data_4) {
     support::endian::write<uint32_t>(&Data[Fixup.getOffset()], Value, Endian);
   } else if (Fixup.getKind() == FK_Data_8) {
     support::endian::write<uint64_t>(&Data[Fixup.getOffset()], Value, Endian);
   } else if (Fixup.getKind() == FK_PCRel_4) {
     Value = (uint32_t)((Value - 8) / 8);

llvm/trunk/lib/Target/Hexagon/MCTargetDesc/HexagonAsmBackend.cpp

class HexagonAsmBackend : public MCAsmBackend {
   std::unique_ptr <MCInst *> RelaxTarget;
   MCInst * Extender;
 
   void ReplaceInstruction(MCCodeEmitter &E, MCRelaxableFragment &RF,
                           MCInst &HMB) const {
     SmallVector<MCFixup, 4> Fixups;
     SmallString<256> Code;
     raw_svector_ostream VecOS(Code);
-    E.encodeInstruction(HMB, VecOS, Fixups, RF.getSubtargetInfo());
+    E.encodeInstruction(HMB, VecOS, Fixups, *RF.getSubtargetInfo());
 
     // Update the fragment.
     RF.setInst(HMB);
     RF.getContents() = Code;
     RF.getFixups() = Fixups;
   }
 
 public:
void HandleFixupError(const int bits, const int align_bits,
     llvm_unreachable(errStr.str().c_str());
   }
 
   /// ApplyFixup - Apply the \arg Value for given \arg Fixup into the provided
   /// data fragment, at the offset specified by the fixup and following the
   /// fixup kind as appropriate.
   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                   const MCValue &Target, MutableArrayRef<char> Data,
-                  uint64_t FixupValue, bool IsResolved) const override {
+                  uint64_t FixupValue, bool IsResolved,
+                  const MCSubtargetInfo *STI) const override {
 
     // When FixupValue is 0 the relocation is external and there
     // is nothing for us to do.
     if (!FixupValue) return;
 
     MCFixupKind Kind = Fixup.getKind();
     uint64_t Value;
     uint32_t InstMask;
bool isInstRelaxable(MCInst const &HMI) const {
 
     return Relaxable;
   }
 
   /// MayNeedRelaxation - Check whether the given instruction may need
   /// relaxation.
   ///
   /// \param Inst - The instruction to test.
-  bool mayNeedRelaxation(MCInst const &Inst) const override {
+  bool mayNeedRelaxation(MCInst const &Inst,
+                         const MCSubtargetInfo &STI) const override {
     return true;
   }
 
   /// fixupNeedsRelaxation - Target specific predicate for whether a given
   /// fixup requires the associated instruction to be relaxed.
   bool fixupNeedsRelaxationAdvanced(const MCFixup &Fixup, bool Resolved,
                                     uint64_t Value,
                                     const MCRelaxableFragment *DF,
for (auto I : Layout.getSectionOrder()) {
               auto &RF = cast<MCRelaxableFragment>(*K);
               auto &Inst = const_cast<MCInst &>(RF.getInst());
               while (Size > 0 && HexagonMCInstrInfo::bundleSize(Inst) < 4) {
                 MCInst *Nop = new (Context) MCInst;
                 Nop->setOpcode(Hexagon::A2_nop);
                 Inst.addOperand(MCOperand::createInst(Nop));
                 Size -= 4;
                 if (!HexagonMCChecker(
-                         Context, *MCII, RF.getSubtargetInfo(), Inst,
+                         Context, *MCII, *RF.getSubtargetInfo(), Inst,
                          *Context.getRegisterInfo(), false)
                          .check()) {
                   Inst.erase(Inst.end() - 1);
                   Size = 0;
                 }
               }
               bool Error = HexagonMCShuffle(Context, true, *MCII,
-                                            RF.getSubtargetInfo(), Inst);
+                                            *RF.getSubtargetInfo(), Inst);
               //assert(!Error);
               (void)Error;
               ReplaceInstruction(Asm.getEmitter(), RF, Inst);
               Layout.invalidateFragmentsFrom(&RF);
               Size = 0; // Only look back one instruction
               break;
             }
             }

llvm/trunk/lib/Target/Lanai/MCTargetDesc/LanaiAsmBackend.cpp

class LanaiAsmBackend : public MCAsmBackend {
   Triple::OSType OSType;
 
 public:
   LanaiAsmBackend(const Target &T, Triple::OSType OST)
       : MCAsmBackend(support::big), OSType(OST) {}
 
   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                   const MCValue &Target, MutableArrayRef<char> Data,
-                  uint64_t Value, bool IsResolved) const override;
+                  uint64_t Value, bool IsResolved,
+                  const MCSubtargetInfo *STI) const override;
 
   std::unique_ptr<MCObjectTargetWriter>
   createObjectTargetWriter() const override;
 
   // No instruction requires relaxation
   bool fixupNeedsRelaxation(const MCFixup & /*Fixup*/, uint64_t /*Value*/,
                             const MCRelaxableFragment * /*DF*/,
                             const MCAsmLayout & /*Layout*/) const override {
     return false;
   }
 
   const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const override;
 
   unsigned getNumFixupKinds() const override {
     return Lanai::NumTargetFixupKinds;
   }
 
-  bool mayNeedRelaxation(const MCInst & /*Inst*/) const override {
+  bool mayNeedRelaxation(const MCInst & /*Inst*/,
+                         const MCSubtargetInfo &STI) const override {
     return false;
   }
 
   void relaxInstruction(const MCInst & /*Inst*/,
                         const MCSubtargetInfo & /*STI*/,
                         MCInst & /*Res*/) const override {}
 
   bool writeNopData(raw_ostream &OS, uint64_t Count) const override;
 };
 
 bool LanaiAsmBackend::writeNopData(raw_ostream &OS, uint64_t Count) const {
   if ((Count % 4) != 0)
     return false;
 
   for (uint64_t i = 0; i < Count; i += 4)
     OS.write("\x15\0\0\0", 4);
 
   return true;
 }
 
 void LanaiAsmBackend::applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                                  const MCValue &Target,
                                  MutableArrayRef<char> Data, uint64_t Value,
-                                 bool /*IsResolved*/) const {
+                                 bool /*IsResolved*/,
+                                 const MCSubtargetInfo */*STI*/) const {
   MCFixupKind Kind = Fixup.getKind();
   Value = adjustFixupValue(static_cast<unsigned>(Kind), Value);
 
   if (!Value)
     return; // This value doesn't change the encoding
 
   // Where in the object and where the number of bytes that need
   // fixing up

llvm/trunk/lib/Target/Mips/MCTargetDesc/MipsAsmBackend.h

MipsAsmBackend(const Target &T, const MCRegisterInfo &MRI, const Triple &TT,
       : MCAsmBackend(TT.isLittleEndian() ? support::little : support::big),
         TheTriple(TT), IsN32(N32) {}
 
   std::unique_ptr<MCObjectTargetWriter>
   createObjectTargetWriter() const override;
 
   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                   const MCValue &Target, MutableArrayRef<char> Data,
-                  uint64_t Value, bool IsResolved) const override;
+                  uint64_t Value, bool IsResolved,
+                  const MCSubtargetInfo *STI) const override;
 
   Optional<MCFixupKind> getFixupKind(StringRef Name) const override;
   const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const override;
 
   unsigned getNumFixupKinds() const override {
     return Mips::NumTargetFixupKinds;
   }
 
   /// @name Target Relaxation Interfaces
   /// @{
 
   /// MayNeedRelaxation - Check whether the given instruction may need
   /// relaxation.
   ///
   /// \param Inst - The instruction to test.
-  bool mayNeedRelaxation(const MCInst &Inst) const override {
+  bool mayNeedRelaxation(const MCInst &Inst,
+                         const MCSubtargetInfo &STI) const override {
     return false;
   }
 
   /// fixupNeedsRelaxation - Target specific predicate for whether a given
   /// fixup requires the associated instruction to be relaxed.
    bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
                              const MCRelaxableFragment *DF,
                              const MCAsmLayout &Layout) const override {

llvm/trunk/lib/Target/Mips/MCTargetDesc/MipsAsmBackend.cpp

 }
 
 /// ApplyFixup - Apply the \p Value for given \p Fixup into the provided
 /// data fragment, at the offset specified by the fixup and following the
 /// fixup kind as appropriate.
 void MipsAsmBackend::applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                                 const MCValue &Target,
                                 MutableArrayRef<char> Data, uint64_t Value,
-                                bool IsResolved) const {
+                                bool IsResolved,
+                                const MCSubtargetInfo *STI) const {
   MCFixupKind Kind = Fixup.getKind();
   MCContext &Ctx = Asm.getContext();
   Value = adjustFixupValue(Fixup, Value, Ctx);
 
   if (!Value)
     return; // Doesn't change encoding.
 
   // Where do we start in the object

llvm/trunk/lib/Target/PowerPC/MCTargetDesc/PPCAsmBackend.cpp

assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
            "Invalid kind!");
     return (Endian == support::little
                 ? InfosLE
                 : InfosBE)[Kind - FirstTargetFixupKind];
   }
 
   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                   const MCValue &Target, MutableArrayRef<char> Data,
-                  uint64_t Value, bool IsResolved) const override {
+                  uint64_t Value, bool IsResolved,
+                  const MCSubtargetInfo *STI) const override {
     Value = adjustFixupValue(Fixup.getKind(), Value);
     if (!Value) return;           // Doesn't change encoding.
 
     unsigned Offset = Fixup.getOffset();
     unsigned NumBytes = getFixupKindNumBytes(Fixup.getKind());
 
     // For each byte of the fragment that the fixup touches, mask in the bits
     // from the fixup value. The Value has been "split up" into the appropriate
case PPC::fixup_ppc_br24abs:
           if ((Other & ELF::STO_PPC64_LOCAL_MASK) != 0)
             return true;
         }
       }
       return false;
     }
   }
 
-  bool mayNeedRelaxation(const MCInst &Inst) const override {
+  bool mayNeedRelaxation(const MCInst &Inst,
+                         const MCSubtargetInfo &STI) const override {
     // FIXME.
     return false;
   }
 
   bool fixupNeedsRelaxation(const MCFixup &Fixup,
                             uint64_t Value,
                             const MCRelaxableFragment *DF,
                             const MCAsmLayout &Layout) const override {

llvm/trunk/lib/Target/Sparc/MCTargetDesc/SparcAsmBackend.cpp

bool shouldForceRelocation(const MCAssembler &Asm, const MCFixup &Fixup,
       case Sparc::fixup_sparc_tls_ie_ldx:
       case Sparc::fixup_sparc_tls_ie_add:
       case Sparc::fixup_sparc_tls_le_hix22:
       case Sparc::fixup_sparc_tls_le_lox10:
         return true;
       }
     }
 
-    bool mayNeedRelaxation(const MCInst &Inst) const override {
+    bool mayNeedRelaxation(const MCInst &Inst,
+                           const MCSubtargetInfo &STI) const override {
       // FIXME.
       return false;
     }
 
     /// fixupNeedsRelaxation - Target specific predicate for whether a given
     /// fixup requires the associated instruction to be relaxed.
     bool fixupNeedsRelaxation(const MCFixup &Fixup,
                               uint64_t Value,
namespace {
   class ELFSparcAsmBackend : public SparcAsmBackend {
     Triple::OSType OSType;
   public:
     ELFSparcAsmBackend(const Target &T, Triple::OSType OSType) :
       SparcAsmBackend(T), OSType(OSType) { }
 
     void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                     const MCValue &Target, MutableArrayRef<char> Data,
-                    uint64_t Value, bool IsResolved) const override {
+                    uint64_t Value, bool IsResolved,
+                    const MCSubtargetInfo *STI) const override {
 
       Value = adjustFixupValue(Fixup.getKind(), Value);
       if (!Value) return;           // Doesn't change encoding.
 
       unsigned Offset = Fixup.getOffset();
 
       // For each byte of the fragment that the fixup touches, mask in the bits
       // from the fixup value. The Value has been "split up" into the

llvm/trunk/lib/Target/SystemZ/MCTargetDesc/SystemZMCAsmBackend.cpp

public:
 
   // Override MCAsmBackend
   unsigned getNumFixupKinds() const override {
     return SystemZ::NumTargetFixupKinds;
   }
   const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const override;
   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                   const MCValue &Target, MutableArrayRef<char> Data,
-                  uint64_t Value, bool IsResolved) const override;
-  bool mayNeedRelaxation(const MCInst &Inst) const override {
+                  uint64_t Value, bool IsResolved,
+                  const MCSubtargetInfo *STI) const override;
+  bool mayNeedRelaxation(const MCInst &Inst,
+                         const MCSubtargetInfo &STI) const override {
     return false;
   }
   bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
                             const MCRelaxableFragment *Fragment,
                             const MCAsmLayout &Layout) const override {
     return false;
   }
   void relaxInstruction(const MCInst &Inst, const MCSubtargetInfo &STI,
assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
          "Invalid kind!");
   return Infos[Kind - FirstTargetFixupKind];
 }
 
 void SystemZMCAsmBackend::applyFixup(const MCAssembler &Asm,
                                      const MCFixup &Fixup,
                                      const MCValue &Target,
                                      MutableArrayRef<char> Data, uint64_t Value,
-                                     bool IsResolved) const {
+                                     bool IsResolved,
+                                     const MCSubtargetInfo *STI) const {
   MCFixupKind Kind = Fixup.getKind();
   unsigned Offset = Fixup.getOffset();
   unsigned BitSize = getFixupKindInfo(Kind).TargetSize;
   unsigned Size = (BitSize + 7) / 8;
 
   assert(Offset + Size <= Data.size() && "Invalid fixup offset!");
 
   // Big-endian insertion of Size bytes.

llvm/trunk/lib/Target/X86/MCTargetDesc/X86AsmBackend.cpp

const MCFixupKindInfo &getFixupKindInfo(MCFixupKind Kind) const override {
     assert(unsigned(Kind - FirstTargetFixupKind) < getNumFixupKinds() &&
            "Invalid kind!");
     assert(Infos[Kind - FirstTargetFixupKind].Name && "Empty fixup name!");
     return Infos[Kind - FirstTargetFixupKind];
   }
 
   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                   const MCValue &Target, MutableArrayRef<char> Data,
-                  uint64_t Value, bool IsResolved) const override {
+                  uint64_t Value, bool IsResolved,
+                  const MCSubtargetInfo *STI) const override {
     unsigned Size = 1 << getFixupKindLog2Size(Fixup.getKind());
 
     assert(Fixup.getOffset() + Size <= Data.size() && "Invalid fixup offset!");
 
     // Check that uppper bits are either all zeros or all ones.
     // Specifically ignore overflow/underflow as long as the leakage is
     // limited to the lower bits. This is to remain compatible with
     // other assemblers.
     assert(isIntN(Size * 8 + 1, Value) &&
            "Value does not fit in the Fixup field");
 
     for (unsigned i = 0; i != Size; ++i)
       Data[Fixup.getOffset() + i] = uint8_t(Value >> (i * 8));
   }
 
-  bool mayNeedRelaxation(const MCInst &Inst) const override;
+  bool mayNeedRelaxation(const MCInst &Inst,
+                         const MCSubtargetInfo &STI) const override;
 
   bool fixupNeedsRelaxation(const MCFixup &Fixup, uint64_t Value,
                             const MCRelaxableFragment *DF,
                             const MCAsmLayout &Layout) const override;
 
   void relaxInstruction(const MCInst &Inst, const MCSubtargetInfo &STI,
                         MCInst &Res) const override;
 

 
 static unsigned getRelaxedOpcode(const MCInst &Inst, bool is16BitMode) {
   unsigned R = getRelaxedOpcodeArith(Inst);
   if (R != Inst.getOpcode())
     return R;
   return getRelaxedOpcodeBranch(Inst, is16BitMode);
 }
 
-bool X86AsmBackend::mayNeedRelaxation(const MCInst &Inst) const {
+bool X86AsmBackend::mayNeedRelaxation(const MCInst &Inst,
+                                      const MCSubtargetInfo &STI) const {
   // Branches can always be relaxed in either mode.
   if (getRelaxedOpcodeBranch(Inst, false) != Inst.getOpcode())
     return true;
 
   // Check if this instruction is ever relaxable.
   if (getRelaxedOpcodeArith(Inst) == Inst.getOpcode())
     return false;
 

llvm/trunk/test/CodeGen/ARM/relax-per-target-feature.ll

+; RUN: llc -mtriple=thumbv4t-linux-gnueabi -o - %s | FileCheck %s
+
+; Functions may have more features than the base triple; code generation and
+; instruction selection may be performed based on this information. This test
+; makes sure that the MC layer performs instruction relaxation based on the
+; target-features of the function. The relaxation for tail call is particularly
+; important on Thumb2 as the 16-bit Thumb branch instruction has an extremely
+; short range.
+
+declare dso_local void @g(...) local_unnamed_addr #2
+
+define dso_local void @f() local_unnamed_addr #0 {
+entry:
+  tail call void bitcast (void (...)* @g to void ()*)() #3
+  ret void
+}
+; Function has thumb2 target-feature, tail call is allowed and must be widened.
+; CHECK: f:
+; CHECK: b g
+
+define dso_local void @h() local_unnamed_addr #2 {
+entry:
+  tail call void bitcast (void (...)* @g to void ()*)() #3
+  ret void
+}
+; Function does not have thumb2 target-feature, tail call should not be
+; generated as it cannot be widened.
+; CHECK: h:
+; CHECK: bl g
+
+attributes #0 = { nounwind  "disable-tail-calls"="false" "target-cpu"="cortex-a53" "target-features"="+crypto,+fp-armv8,+neon,+soft-float-abi,+strict-align,+thumb-mode,-crc,-dotprod,-dsp,-hwdiv,-hwdiv-arm,-ras" "use-soft-float"="true" }
+
+attributes #2 = { nounwind  "disable-tail-calls"="false" "target-cpu"="arm7tdmi" "target-features"="+strict-align,+thumb-mode,-crc,-dotprod,-dsp,-hwdiv,-hwdiv-arm,-ras" "unsafe-fp-math"="false" "use-soft-float"="true" }
+attributes #3 = { nounwind }

llvm/trunk/test/MC/ARM/AlignedBundling/illegal-subtarget-change.s

+# RUN: not llvm-mc -filetype=obj -triple armv7-linux-gnueabi %s -o - 2>&1 | FileCheck %s
+
+        # We cannot switch subtargets mid-bundle
+        .syntax unified
+        .text
+        .bundle_align_mode 4
+        .arch armv4t
+        bx lr
+        .bundle_lock
+        bx lr
+        .arch armv7a
+        movt r0, #0xffff
+        movw r0, #0xffff
+        .bundle_unlock
+        bx lr
+# CHECK: LLVM ERROR: A Bundle can only have one Subtarget.

llvm/trunk/test/MC/ARM/AlignedBundling/subtarget-change.s

+# RUN: llvm-mc -filetype=obj -triple armv7-linux-gnueabi %s -o - \
+# RUN:   | llvm-objdump -no-show-raw-insn -triple armv7 -disassemble - | FileCheck %s
+
+        # We can switch subtargets with .arch outside of a bundle
+        .syntax unified
+        .text
+        .bundle_align_mode 4
+        .arch armv4t
+        bx lr
+        .bundle_lock
+        and r1, r1, r1
+        and r1, r1, r1
+        .bundle_unlock
+        bx lr
+
+        # We can switch subtargets at the start of a bundle
+        bx lr
+        .bundle_lock align_to_end
+        .arch armv7a
+        movt r0, #0xffff
+        movw r0, #0xffff
+        .bundle_unlock
+        bx lr
+
+# CHECK:      0: bx    lr
+# CHECK-NEXT: 4: and   r1, r1, r1
+# CHECK-NEXT: 8: and   r1, r1, r1
+# CHECK-NEXT: c: bx    lr
+# CHECK-NEXT: 10: bx    lr
+# CHECK-NEXT: 14: nop
+# CHECK-NEXT: 18: movt  r0, #65535
+# CHECK-NEXT: 1c: movw  r0, #65535
+# CHECK-NEXT: 20: bx    lr

llvm/trunk/test/MC/ARM/fixup-per-fragment.s

+@ RUN: not llvm-mc -triple armv7a-linux-gnueabihf %s -filetype=obj -o %s.o 2>&1 | FileCheck %s
+
+@ The relaxations should be applied using the subtarget from the fragment
+@ containing the fixup and not the per module subtarget.
+
+        .syntax unified
+        .thumb
+        @ Place a literal pool out of range of the 16-bit ldr but within
+        @ range of the 32-bit ldr.w
+        .text
+        @ Relaxation to ldr.w as target triple supports Thumb2
+        ldr r0,=0x12345678
+        .arch armv4t
+        @ No relaxation as v4t does not support Thumb
+        @ expect out of range error message
+        ldr r0,=0x87654321
+        .space 1024
+
+@ CHECK: error: out of range pc-relative fixup value
+@ CHECK-NEXT: ldr r0,=0x87654321

llvm/trunk/test/MC/X86/AlignedBundling/bundle-subtarget-change-error.s

+# RUN: not llvm-mc -filetype=obj -triple x86_64-pc-linux-gnu -mcpu=pentiumpro %s -o - 2>&1 | FileCheck %s
+# RUN: not llvm-mc -filetype=obj -triple x86_64-pc-linux-gnu -mcpu=pentiumpro -mc-relax-all %s -o - 2>&1 | FileCheck %s
+
+# Switching mode will change subtarget, which we can't do within a bundle
+  .text
+  .code64
+  .bundle_align_mode 4
+foo:
+  pushq   %rbp
+  .bundle_lock
+  addl    %ebp, %eax
+  .code32
+  movb  $0x0, (%si)
+  .bundle_unlock
+
+CHECK:  LLVM ERROR: A Bundle can only have one Subtarget.