[X86] Convert esp-relative movs of function arguments to pushes, step 2

[mkuper]

Summary

This is a first stab at the next step of the mov-to-push transformation.

It moves the transformation earlier in the pass order so that it can do load-folding, and prepares the required infrastructure.
It is still enabled only in cases where it should be a clear win - when we don't expect to have a reserved call frame, or when optimizing for size.
The next step will be a heuristic that makes a smarter decision on when this should be enabled.

As a side note - I've done some internal testing for effects on the code size, but I'd like to do some testing for things other people care about as well. So, if you have a x86-32 code-base where you care about the code size, and is publicly available, let me know.

[mkuper]

Removed a horrible hack that was, in addition to being horrible, completely wrong, and added a test-case to cover the issue.

Also, ping?

[delena]

I suggest to check also varargs and stdcall functions, were the callee clears the stack.

lib/Target/X86/CMakeLists.txt

17

Can you add this code to X86FrameLowering.cpp ?

lib/Target/X86/X86.h

70

I suggest to choose another name, something like optimizeCallFrameForSize

lib/Target/X86/X86ConvertMovsToPushes.cpp

40

I don't think that we really need this knob.

114

If you change instructions inside bb, your iterator may be broken.

213

It should be immediate, right? Can we have a relocation here?

225

SlowPush should be a property of the target, like slowLea

227

The comment is missing here.

[mkuper]

Thanks, Elena!
Will upload a new version.

lib/Target/X86/X86.h

70

I wasn't happy with the name either, but didn't have any good ideas at the time.
Will do.

lib/Target/X86/X86ConvertMovsToPushes.cpp

40

I'd rather keep this knob, it's fairly useful for debugging.
Of course, it's internal only, not exposed to clang.

114

As far as I know, MBB iterators aren't invalidated by removing other instructions, and we don't remove the FrameSetup itself.
But it's probably better to keep going from the FrameDestroy instead of the next instruction.
Will change that.

213

It can be a relocation, but in that case, isImm() will fail.
Will document that more clearly

225

I agree. Unfortunately, I've run out of bits.
The Subtarget features are 64-bit bitfield, and they're all taken.

lib/Target/X86/X86RegisterInfo.cpp

508

And, apparently, this is still wrong, because eliminateCallFramePseudoInstr() may actually adjust the SP by a different amount than what PEI passes as the SPAdj, e.g. due to stack alignment concerns.

[mkuper]

lib/Target/X86/X86ConvertMovsToPushes.cpp

286

Argh. This is nonsense. Commented one thing, coded another...
(mayStore() is extremely far from being a strong enough condition to allow this.)

[rnk]

lib/Target/X86/X86ConvertMovsToPushes.cpp

40

I would also like this as a temporary testing knob so that I can evaluate this across a large codebase.

[mkuper]

So, this version should actually work (e.g. it can self-host and past check-llvm. Without the stackalign restriction of course, since that currently makes it a nop except on windows).
Unfortunately, it has several big warts, so I'm not planning to commit it as is. This is more of a request for ideas on how to improve the code.

So, any ideas on how to make this sane, especially X86InstrInfo::getSPAdjust(), are welcome.

[rnk]

I haven't finished reviewing yet, but I've got to run and handle something personal.

At a high level, is there any reason we shouldn't commit to push/pop earlier to allow for better ISel, rather than trying to transform call sequences later? Specifically, I'm thinking about adding an X86ISD::PUSH DAG node and changing X86TargetLowering::LowerCall() to use it.

lib/CodeGen/PrologEpilogInserter.cpp

856–857

This seems like an x86-specific quirk, right? Given "push [esp + 8]", x86 chips will load [esp + 8] before adjusting esp, and I think this code motion accomplishes that.

I'm OK with that motion so long as there are no other upstream LLVM backends with CISC-y instructions like "push [SP-mem]". :)

lib/Target/X86/X86ConvertMovsToPushes.cpp

12

s/stck/stack/

83–84

I think it's important to at least support __thiscall eventually, since that's a very common convention with one regparm.

85–87

I guess I would justify this more in terms of reducing the extra CFI that we would have to emit to describe the SP adjustments. Converting a few movs to pushes isn't worth the complexity.

144

Can you explain why this is unprofitable? I guess if we get here we are in dyanamic alloca plus stack realignment land, i.e. the worst thing that could possibly happen. Is this about extra code for preserving the outgoing stack alignment then? Like on Linux, where we provide 16 byte stack alignment?

[mkuper]

Thanks, Reid!

Waiting for the second part, you didn't get to the really horrible stuff yet...

Regarding the high level, two reasons:

1. It seemed like it was going to be simpler. I'm not so sure anymore, but I still think it is. (Note that we'll still need to fix all of the code that tracks SP adjustment, that's not going away in either case).
2. The main problem is that next step after this is going to be a function-scope heuristic. To use this transformation for even one call-site, I have to disable the reserved frame for the whole function. So, I need to try to approximate the impact on the whole function (which contains some calls that will be converted to use pushes, and some calls that won't be). I don't see how this can be done on the DAG level.

lib/CodeGen/PrologEpilogInserter.cpp

856–857

This call to SPAdjust() always returns 0 right now (barring the code in this patch), it was added as part of my refactoring in D6863, and I added it in the wrong place.

The motivation here wasn't a push, actually, since I try to never generate push [esp + 8], that's filtered out by the code in the optimization. Although I can probably start generating them - I was trying to filter them out precisely because I didn't want all of this complexity at the first stage, but apparently it's necessary.

The problem is that once we don't have a reserved call frame (regardless of the push transformation), you can have things like
CALL32r <fi#1>, where the call is callee-pop.
So you need to resolve the indirect call using the stack-pointer from before the call.

lib/Target/X86/X86ConvertMovsToPushes.cpp

83–84

Yes, and maybe even for _fastcall (It looks like gcc will do this for fastcall, icc won't).
But I am still trying to do this gradually, to the extent that I can. :-)

85–87

You're right, that too.

144

If we get here, we're in opt-for-size + stack-realignment land.

And, yes, that's exactly what it is is about.
If you are passing only one parameter, the original code would be:

mov %eax, 128(%esp)
call $foo

Without re-alignment, you have

push %eax
call $foo
add $4, %esp

which is still a win in terms of code-size

With re-alignment, you get:

sub $16, %esp
push %eax
call $foo
add $12, %esp

Which is... questionable. The code size for the sequence is the same (in this case, 7 bytes for both, not including the call), but if you have other call sites which you didn't convert, you may actually lose. And, of course, you lose performance (3 instructions instead of 1) without anything to show for it.

Once there is a heuristic that tries to estimate the overhead, we can address this on a case-by-case basis (e.g. if we have 16-byte stack re-alignment, but most call-sites have a lot of parameters, then it's still worth it.)

[rnk]

lib/Target/X86/X86ConvertMovsToPushes.cpp

129–131

I think I misinterpreted this on the first pass. We always expect this to be profitable if we know we *can't* reserve space for the call frame. Maybe rename the bool to CannotReserveFrame to match the sense?

144

Based on my misinterpretation, I think I understand why you get this code. SP is assumed to be aligned coming into the sequence. We realign SP after dynamic allocas. The sequence is probably more like:

sub $12, %esp
push %eax
call $foo
add $16, %esp

I can see why this is less profitable.

209

std::map is really malloc heavy. This can probably be a SmallVector<MachineInstr*, 8> or something, mapping slot index to the MI that fills it. The frame setup opcode should tell you how much stack space to allocate up front, and you can index into the vector by StackOffset / 4.

221–223

This seems worth tackling, given that you had to handle the call <fi> case. :)

365–369

It's not clear to me that same BB is sufficient, consider this potential BB:

movl (%edi), %eax
movl $42, (%edi)
<call setup>
movl %eax, (%esp)
calll foo
<call end>

We can't move the load if there is a potentially aliasing store in the way. There might be a utility to help with the aliasing query, or you can assume that any stores other than arg stores might alias it and bail on that.

lib/Target/X86/X86InstrInfo.cpp

1717–1718

This is the best thing I can think of at the moment. =/

test/CodeGen/X86/movtopush.ll

212

Test case suggestions:

; Where the callee is indirect via the stack, `call <fi>`
define void @test10() optsize {
  %stack_fptr = alloca void (i32, i32, i32, i32)*
  store void (i32, i32, i32, i32)* @good, void (i32, i32, i32, i32)** %stack_fptr
  %good_ptr = load void (i32, i32, i32, i32)** %stack_fptr
  call void (i32, i32, i32, i32)* %good_ptr(i32 1, i32 2, i32 3, i32 4)
  ret void
}

; We can't fold the load into the push here, skipping the store.
@the_global = global i32
define void @test11() optsize {
  %myload = load i32* @the_global
  store i32 42, i32* @the_global
  call void @good(i32 %myload, i32 2, i32 3, i32 4)
  ret void
}

[mkuper]

Thanks, Reid!

lib/Target/X86/X86ConvertMovsToPushes.cpp

129–131

Err, yes, you're right, sorry about that... got distracted while naming the variable, I guess, I meant the opposite.
Thanks!

144

Yes, that sequence. :-)

It doesn't depend on dynamic allocas, though.
If you don't have a reserved frame (for whatever reason - for x86 after this patch, it's either dynamic allocas, or because we forced it not to reserve by using pushes), then you need this re-alignment.

209

That can work. Thanks, I'll try.

221–223

Yes, definitely. :-)
It may even work out of the box now. But I think I still want to split it into a separate commit.

365–369

Right now I'm way more conservative than even that - I'm checking below that everything between this mov and the call setup is a MOV32rm. The "same basic block" check here is just a way to short-circuit the obviously wrong cases.

This catches some common cases like the one in the comment above, but of course misses other opportunities.
I could check for a mayStore() instead, but I'm not sure that's safe enough. I'd like to relax the condition - but again, I think that ought to be a separate commit.

lib/Target/X86/X86InstrInfo.cpp

1717–1718

Too bad. :-\

So you think I should commit with this code as is?
This shouldn't be a huge problem in terms of compile-time (since I'm looking only until the next call, it can't go quadratic), but it's insanely ugly.

[mkuper]

• Applied review comments
• Fixed another bug in the way PEI was handling push sequences (argh) - this required adding a target query.
• Made the tests check a bit more (which would have exposed the bug above earlier).

[rnk]

lib/Target/X86/X86InstrInfo.cpp

1717–1718

Yeah, if we go with this MI pass approach to mov -> push conversion, then we'll have to keep this ADJCALLSTACKUP scan. We aren't going to move the callee cleanup stack adjustment onto the CALL instr without major changes.

1745

I wonder if it's possible for __readeflags() (pushf ; pop %reg) or others to get folded into a call sequence. Probably not.

[mkuper]

lib/Target/X86/X86InstrInfo.cpp

1717–1718

This will have to happen regardless of the MI pass vs. DAG approach.

I mean, I still think doing it on the DAG is unfeasible, but even if we could do that, it wouldn't help.
This code is used for the case where fi resolution needs to handle a a sequence where there is a fi reference between the call and the adjcallstackup, with callee cleanup for the call.
This is just a side effect of making canSimplifyCallFramePseudos return false.

1745

I don't see how it could happen.
In any case, we won't match either the pushf or the pop, so it should be ok.

[rnk]

lgtm

I still think forming pushes prior to isel is the way to go long term. It's a lot easier to convert pushes to 'load, SP adjust, store' than it is to go the other way.

include/llvm/Target/TargetFrameLowering.h

196

"- Do" uppercase

lib/Target/X86/X86ConvertMovsToPushes.cpp

100

Can this be for (MachineBasicBlock &BB : *MF) {?

102

Ditto, for (MachineInstr &MI : BB) { ?

lib/Target/X86/X86InstrInfo.cpp

1713–1725

I would shorten this to just something like "look for the ADJCALLSTACKUP instr that follows the call".

1717–1718

I was imagining in the DAG LowerCall implementation we emit FrameIndex operands with some kind of SP offset to indicate the current stack level. We'd end up with MI looking like this:

ADJCALLSTACKDOWN32 <N> ; N is <size-of-args> % <stack-alignment>, which is usually zero
PUSH32rmm <fi> <sp offset, N>
PUSH32rmm <fi> <sp offset, N + 4>
PUSH32rmm <fi> <sp offset, N + 8>
CALL32rm <fi> <sp offset, N + 12>
ADJCALLSTACKUP32 <N + 12>

The main thing is that if we commit to pushes instead of movs at DAG time, it's impossible for the push conversion to fail for hard to diagnose reasons.

It looks like the frame index MachineOperand type has an unused offset field.

[mkuper]

Hi Chandler,

This is something that Reid and I talked about on IRC, but I don’t think we came to a conclusion both of us were happy with (hence Reid’s “lgtm with reservations”, I guess :-) )

First, I don’t think the decision on whether to use movs or pushes belongs in the DAG.
The decision on whether a call-site should use movs or pushes needs to be aware of its context, because having even one call-site use pushes means we will not have a reserved call frame, which affects the way all other call sites are treated as well. This patch makes the decision based on global attributes only (opt for size vs. speed, stack alignment), but the next step will be to make it based on an analysis of the call-sites – e.g. even with stack alignment of 16, it can still often be a win, depending on just how many of the function calls we can actually transform, and how many memory arguments each call has.

So the way I envision the next step is that the pass will:

a) Collect the necessary information from all call sites in the function.

b) Make a judgment on whether the transformation is worth it – in terms of size for Os/Oz, in terms of performance for other opt levels.

c) Perform the transformation.
I don’t see how we can do this on the DAG.

If I understand Reid’s last suggestion, he proposed to flip the default – that is, emit pushes in the DAG, and have an MI pass that does the opposite (push -> mov) transformation if necessary.

I don’t believe that removes a lot of complexity or would improve performance.
The code in PEI, InstrInfo and FrameLowering is just a side effect on not being able to rely on a 0 SPAdj in PEI anymore (that is, canSimplifyCallFramePseudos() can now return false), and is needed regardless of how the transformation is performed. And we will still need the heuristic decision.
Some of the logic in looking for sequences where the conversion is possible will disappear, but I think a lot of it will remain as conditions on the incoming operand DAG nodes. And since we don’t want to transform each push into a “adjust esp, mov” but rather want to group all the esp adjustments back into the ADJCALLSTACKs, we will still need to have code in the pass that make sure this is safe w.r.t to the final sequence.
The main benefit I see is that we will no longer need to have the folding code – rather, we will have to unfold PUSH32rmm, which is simpler. However, I hope I can eventually get rid of the folding here by teaching PeepholeOptimizer to be smarter about this.

On the other hand, X86TargetLowering::LowerCall() is already, IMHO, a fairly complex piece of code, and I’d rather avoid making it even more complex.
Conceptually, I’d prefer that LowerCall() did standard mov-based lowering in all cases like it does now (we aren’t always going to lower to pushes anyway – it doesn’t really make sense for x864-64) and treat pushes as an optimization where available.

What do you think?

Michael

From: Chandler Carruth [mailto:[e-mail]]
Sent: Thursday, January 22, 2015 23:07
To: [e-mail]
Cc: Kuperstein, Michael M; Nadav Rotem; Demikhovsky, Elena; Commit Messages and Patches for LLVM
Subject: Re: [PATCH] [X86] Convert esp-relative movs of function arguments to pushes, step 2

include/llvm/Target/TargetFrameLowering.h

public:
   /// need adjusting for the call frame adjustments. Normally, that's true
   /// if the function has a reserved call frame or a frame pointer. Some
   /// targets (Thumb2, for example) may have more complicated criteria,
   /// however, and can override this behavior.
   virtual bool canSimplifyCallFramePseudos(const MachineFunction &MF) const {
     return hasReservedCallFrame(MF) || hasFP(MF);
   }
 
+  // needsFrameIndexResolution - do we need to perform FI resolution for
+  // this function. Normally, this is required only when the function
+  // has any stack objects. However, targets may want to override this.
+  virtual bool needsFrameIndexResolution(const MachineFunction &MF) const;
+
   /// getFrameIndexOffset - Returns the displacement from the frame register to
   /// the stack frame of the specified index.
   virtual int getFrameIndexOffset(const MachineFunction &MF, int FI) const;
 
   /// getFrameIndexReference - This method should return the base register
   /// and offset used to reference a frame index location. The offset is
   /// returned directly, and the base register is returned via FrameReg.
   virtual int getFrameIndexReference(const MachineFunction &MF, int FI,

lib/CodeGen/PrologEpilogInserter.cpp

void PEI::insertPrologEpilogCode(MachineFunction &Fn) {
   if (Fn.getFunction()->getCallingConv() == CallingConv::HiPE)
     TFI.adjustForHiPEPrologue(Fn);
 }
 
 /// replaceFrameIndices - Replace all MO_FrameIndex operands with physical
 /// register references and actual offsets.
 ///
 void PEI::replaceFrameIndices(MachineFunction &Fn) {
-  if (!Fn.getFrameInfo()->hasStackObjects()) return; // Nothing to do?
+  const TargetFrameLowering &TFI = *Fn.getSubtarget().getFrameLowering();
+  if (!TFI.needsFrameIndexResolution(Fn)) return;
 
   // Store SPAdj at exit of a basic block.
   SmallVector<int, 8> SPState;
   SPState.resize(Fn.getNumBlockIDs());
   SmallPtrSet<MachineBasicBlock*, 8> Reachable;
 
   // Iterate over the reachable blocks in DFS order.
   for (auto DFI = df_ext_begin(&Fn, Reachable), DFE = df_ext_end(&Fn, Reachable);
if (I->getOpcode() == FrameSetupOpcode ||
       // Visit the instructions created by eliminateCallFramePseudoInstr().
       if (PrevI == BB->end())
         I = BB->begin();     // The replaced instr was the first in the block.
       else
         I = std::next(PrevI);
       continue;
     }
 
-    // If we are looking at a call sequence, we need to keep track of
-    // the SP adjustment made by each instruction in the sequence.
-    // This includes both the frame setup/destroy pseudos (handled above),
-    // as well as other instructions that have side effects w.r.t the SP.
-    if (InsideCallSequence)
-      SPAdj += TII.getSPAdjust(I);
-
     MachineInstr *MI = I;
     bool DoIncr = true;
     for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
       if (!MI->getOperand(i).isFI())
         continue;
 
       // Frame indicies in debug values are encoded in a target independent
       // way with simply the frame index and offset rather than any
for (unsigned i = 0, e = MI->getNumOperands(); i != e; ++i) {
         I = BB->begin();
         DoIncr = false;
       }
 
       MI = nullptr;
       break;
     }
 
+    // If we are looking at a call sequence, we need to keep track of
+    // the SP adjustment made by each instruction in the sequence.
+    // This includes both the frame setup/destroy pseudos (handled above),
+    // as well as other instructions that have side effects w.r.t the SP.
+    // Note that this must come after eliminateFrameIndex, because 
+    // if I itself referred to a frame index, we shouldn't count its own
+    // adjustment.
+    if (MI && InsideCallSequence)
+      SPAdj += TII.getSPAdjust(MI);
+
     if (DoIncr && I != BB->end()) ++I;
 
     // Update register states.
     if (RS && !FrameIndexVirtualScavenging && MI) RS->forward(MI);
   }
 }
 
 /// scavengeFrameVirtualRegs - Replace all frame index virtual registers

lib/CodeGen/TargetFrameLoweringImpl.cpp

int TargetFrameLowering::getFrameIndexReference(const MachineFunction &MF,
   const TargetRegisterInfo *RI = MF.getSubtarget().getRegisterInfo();
 
   // By default, assume all frame indices are referenced via whatever
   // getFrameRegister() says. The target can override this if it's doing
   // something different.
   FrameReg = RI->getFrameRegister(MF);
   return getFrameIndexOffset(MF, FI);
 }
+
+bool TargetFrameLowering::needsFrameIndexResolution(
+    const MachineFunction &MF) const {
+  return MF.getFrameInfo()->hasStackObjects();
+}

lib/Target/X86/CMakeLists.txt

 set(LLVM_TARGET_DEFINITIONS X86.td)
 
 tablegen(LLVM X86GenRegisterInfo.inc -gen-register-info)
 tablegen(LLVM X86GenDisassemblerTables.inc -gen-disassembler)
 tablegen(LLVM X86GenInstrInfo.inc -gen-instr-info)
 tablegen(LLVM X86GenAsmWriter.inc -gen-asm-writer)
 tablegen(LLVM X86GenAsmWriter1.inc -gen-asm-writer -asmwriternum=1)
 tablegen(LLVM X86GenAsmMatcher.inc -gen-asm-matcher)
 tablegen(LLVM X86GenDAGISel.inc -gen-dag-isel)
 tablegen(LLVM X86GenFastISel.inc -gen-fast-isel)
 tablegen(LLVM X86GenCallingConv.inc -gen-callingconv)
 tablegen(LLVM X86GenSubtargetInfo.inc -gen-subtarget)
 add_public_tablegen_target(X86CommonTableGen)
 
 set(sources
   X86AsmPrinter.cpp
+  X86ConvertMovsToPushes.cpp
   X86FastISel.cpp
   X86FloatingPoint.cpp
   X86FrameLowering.cpp
   X86ISelDAGToDAG.cpp
   X86ISelLowering.cpp
   X86InstrInfo.cpp
   X86MCInstLower.cpp
   X86MachineFunctionInfo.cpp

lib/Target/X86/X86.h

 /// with NOOPs. This will prevent a stall when returning on the Atom.
 FunctionPass *createX86PadShortFunctions();
 /// createX86FixupLEAs - Return a a pass that selectively replaces
 /// certain instructions (like add, sub, inc, dec, some shifts,
 /// and some multiplies) by equivalent LEA instructions, in order
 /// to eliminate execution delays in some Atom processors.
 FunctionPass *createX86FixupLEAs();
 
+/// createX86CallFrameOptimization - Return a pass that optimizes
+/// the code-size of x86 call sequences. This is done by replacing
+/// esp-relative movs with pushes.
+FunctionPass *createX86CallFrameOptimization();
+
 } // End llvm namespace
 
 #endif

lib/Target/X86/X86ConvertMovsToPushes.cpp

+//===----- X86CallFrameOptimization.cpp - Optimize x86 call sequences -----===//
+//
+//                     The LLVM Compiler Infrastructure
+//
+// This file is distributed under the University of Illinois Open Source
+// License. See LICENSE.TXT for details.
+//
+//===----------------------------------------------------------------------===//
+//
+// This file defines a pass that optimizes call sequences on x86.
+// Currently, it converts movs of function parameters onto the stack into 
+// pushes. This is beneficial for two main reasons:
+// 1) The push instruction encoding is much smaller than an esp-relative mov
+// 2) It is possible to push memory arguments directly. So, if the
+//    the transformation is preformed pre-reg-alloc, it can help relieve
+//    register pressure.
+//
+//===----------------------------------------------------------------------===//
+
+#include <algorithm>
+
+#include "X86.h"
+#include "X86InstrInfo.h"
+#include "X86Subtarget.h"
+#include "X86MachineFunctionInfo.h"
+#include "llvm/ADT/Statistic.h"
+#include "llvm/CodeGen/MachineFunctionPass.h"
+#include "llvm/CodeGen/MachineInstrBuilder.h"
+#include "llvm/CodeGen/MachineRegisterInfo.h"
+#include "llvm/CodeGen/Passes.h"
+#include "llvm/IR/Function.h"
+#include "llvm/Support/Debug.h"
+#include "llvm/Support/raw_ostream.h"
+#include "llvm/Target/TargetInstrInfo.h"
+
+using namespace llvm;
+
+#define DEBUG_TYPE "x86-cf-opt"
+
+cl::opt<bool> NoX86CFOpt("no-x86-call-frame-opt",
+              cl::desc("Avoid optimizing x86 call frames for size"),
+              cl::init(false), cl::Hidden);
+
+namespace {
+class X86CallFrameOptimization : public MachineFunctionPass {
+public:
+  X86CallFrameOptimization() : MachineFunctionPass(ID) {}
+
+  bool runOnMachineFunction(MachineFunction &MF) override;
+
+private:
+  bool shouldPerformTransformation(MachineFunction &MF);
+
+  bool adjustCallSequence(MachineFunction &MF, MachineBasicBlock &MBB,
+                          MachineBasicBlock::iterator I);
+
+  MachineInstr *canFoldIntoRegPush(MachineBasicBlock::iterator FrameSetup,
+                                   unsigned Reg);
+
+  const char *getPassName() const override {
+    return "X86 Optimize Call Frame";
+  }
+
+  const TargetInstrInfo *TII;
+  const TargetFrameLowering *TFL;
+  const MachineRegisterInfo *MRI;
+  static char ID;
+};
+
+char X86CallFrameOptimization::ID = 0;
+}
+
+FunctionPass *llvm::createX86CallFrameOptimization() {
+  return new X86CallFrameOptimization();
+}
+
+// This checks whether the transformation is legal and profitable
+bool X86CallFrameOptimization::shouldPerformTransformation(MachineFunction &MF) {
+  if (NoX86CFOpt.getValue())
+    return false;
+
+  // We currently only support call sequences where *all* parameters.
+  // are passed on the stack.
+  // No point in running this in 64-bit mode, since some arguments are
+  // passed in-register in all common calling conventions, so the pattern
+  // we're looking for will never match.
+  const X86Subtarget &STI = MF.getTarget().getSubtarget<X86Subtarget>();
+  if (STI.is64Bit())
+    return false;
+
+  // You would expect straight-line code between call-frame setup and
+  // call-frame destroy. You would be wrong. There are circumstances (e.g.
+  // CMOV_GR8 expansion of a select that feeds a function call!) where we can
+  // end up with the setup and the destroy in different basic blocks.
+  // This is bad, and breaks SP adjustment.
+  // So, check that all of the frames in the function are closed inside
+  // the same block, and, for good measure, that there are no nested frames.
+  int FrameSetupOpcode = TII->getCallFrameSetupOpcode();
+  int FrameDestroyOpcode = TII->getCallFrameDestroyOpcode();
+  for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); BB != E; ++BB) {
+    bool InsideFrameSequence = false;
+    for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I) {
+      if (I->getOpcode() == FrameSetupOpcode) {
+        if (InsideFrameSequence)
+          return false;
+        InsideFrameSequence = true;
+      }
+      else if (I->getOpcode() == FrameDestroyOpcode) {
+        if (!InsideFrameSequence)
+          return false;
+        InsideFrameSequence = false;
+      }
+    }
+
+    if (InsideFrameSequence)
+      return false;
+  }
+
+  // Now that we know the transformation is legal, check if it is
+  // profitable.
+  // TODO: Add a heuristic that actually looks at the function,
+  //       and enable this for more cases.
+
+  // This transformation is always a win when we expected to have
+  // a reserved call frame. Under other circumstances, it may be either 
+  // a win or a loss, and requires a heuristic.
+  // For now, enable it only for the relatively clear win cases.
+  bool CannotReservedFrame = MF.getFrameInfo()->hasVarSizedObjects();
+  if (CannotReservedFrame)
+    return true;
+
+  // For now, don't even try to evaluate the profitability when
+  // not optimizing for size.
+  AttributeSet FnAttrs = MF.getFunction()->getAttributes();
+  bool OptForSize =
+    FnAttrs.hasAttribute(AttributeSet::FunctionIndex,
+    Attribute::OptimizeForSize) ||
+    FnAttrs.hasAttribute(AttributeSet::FunctionIndex, Attribute::MinSize);
+
+  if (!OptForSize)
+    return false;
+
+  // Stack re-alignment can make this unprofitable even in terms of size.
+  // As mentioned above, a better heuristic is needed. For now, don't do this
+  // when the required alignment is above 8. (4 would be the safe choice, but
+  // some experimentation showed 8 is generally good).
+  if (TFL->getStackAlignment() > 8)
+    return false;
+
+  return true;
+}
+
+bool X86CallFrameOptimization::runOnMachineFunction(MachineFunction &MF) {
+  TII = MF.getSubtarget().getInstrInfo();
+  TFL = MF.getSubtarget().getFrameLowering();
+  MRI = &MF.getRegInfo();
+
+  if (!shouldPerformTransformation(MF))
+    return false;
+
+  int FrameSetupOpcode = TII->getCallFrameSetupOpcode();
+
+  bool Changed = false;
+
+  for (MachineFunction::iterator BB = MF.begin(), E = MF.end(); BB != E; ++BB)
+    for (MachineBasicBlock::iterator I = BB->begin(); I != BB->end(); ++I)
+      if (I->getOpcode() == FrameSetupOpcode)
+        Changed |= adjustCallSequence(MF, *BB, I);
+
+  return Changed;
+}
+
+bool X86CallFrameOptimization::adjustCallSequence(MachineFunction &MF,
+                                                MachineBasicBlock &MBB,
+                                                MachineBasicBlock::iterator I) {
+
+  // Check that this particular call sequence is amenable to the
+  // transformation.
+  const X86RegisterInfo &RegInfo = *static_cast<const X86RegisterInfo *>(
+                                       MF.getSubtarget().getRegisterInfo());
+  unsigned StackPtr = RegInfo.getStackRegister();
+  int FrameDestroyOpcode = TII->getCallFrameDestroyOpcode();
+
+  // We expect to enter this at the beginning of a call sequence
+  assert(I->getOpcode() == TII->getCallFrameSetupOpcode());
+  MachineBasicBlock::iterator FrameSetup = I++;
+
+  
+  // For globals in PIC mode, we can have some LEAs here.
+  // Ignore them, they don't bother us.
+  // TODO: Extend this to something that covers more cases.
+  while (I->getOpcode() == X86::LEA32r)
+    ++I;
+  
+  // We expect a copy instruction here.
+  // TODO: The copy instruction is a lowering artifact.
+  //       We should also support a copy-less version, where the stack
+  //       pointer is used directly.
+  if (!I->isCopy() || !I->getOperand(0).isReg())
+    return false;
+  MachineBasicBlock::iterator SPCopy = I++;
+  StackPtr = SPCopy->getOperand(0).getReg();
+
+  // Scan the call setup sequence for the pattern we're looking for.
+  // We only handle a simple case - a sequence of MOV32mi or MOV32mr
+  // instructions, that push a sequence of 32-bit values onto the stack, with
+  // no gaps between them.
+  SmallVector<MachineInstr*, 4> MovVector(4, nullptr);
+  unsigned int MaxAdjust = FrameSetup->getOperand(0).getImm() / 4;
+  if (MaxAdjust > 4)
+    MovVector.resize(MaxAdjust, nullptr);
+
+  do {
+    int Opcode = I->getOpcode();
+    if (Opcode != X86::MOV32mi && Opcode != X86::MOV32mr)
+      break;
+
+    // We only want movs of the form:
+    // movl imm/r32, k(%esp)
+    // If we run into something else, bail.
+    // Note that AddrBaseReg may, counter to its name, not be a register,
+    // but rather a frame index.
+    // TODO: Support the fi case. This should probably work now that we
+    // have the infrastructure to track the stack pointer within a call
+    // sequence.
+    if (!I->getOperand(X86::AddrBaseReg).isReg() ||
+        (I->getOperand(X86::AddrBaseReg).getReg() != StackPtr) ||
+        !I->getOperand(X86::AddrScaleAmt).isImm() ||
+        (I->getOperand(X86::AddrScaleAmt).getImm() != 1) ||
+        (I->getOperand(X86::AddrIndexReg).getReg() != X86::NoRegister) ||
+        (I->getOperand(X86::AddrSegmentReg).getReg() != X86::NoRegister) ||
+        !I->getOperand(X86::AddrDisp).isImm())
+      return false;
+
+    int64_t StackDisp = I->getOperand(X86::AddrDisp).getImm();
+    assert(StackDisp >= 0 && "Negative stack displacement when passing parameters");
+
+    // We really don't want to consider the unaligned case.
+    if (StackDisp % 4)
+      return false;
+    StackDisp /= 4;
+
+    assert((size_t)StackDisp < MovVector.size() &&
+      "Function call has more parameters than the stack is adjusted for.");
+
+    // If the same stack slot is being filled twice, something's fishy.
+    if (MovVector[StackDisp] != nullptr)
+      return false;
+    MovVector[StackDisp] = I;
+
+    ++I;
+  } while (I != MBB.end());
+
+  // We now expect the end of the sequence - a call and a stack adjust.
+  if (I == MBB.end())
+    return false;
+
+  // For PCrel calls, we expect an additional COPY of the basereg.
+  // If we find one, skip it.
+  if (I->isCopy()) {
+    if (I->getOperand(1).getReg() ==
+      MF.getInfo<X86MachineFunctionInfo>()->getGlobalBaseReg())
+      ++I;
+    else
+      return false;
+  }
+
+  if (!I->isCall())
+    return false;
+  MachineBasicBlock::iterator Call = I;
+  if ((++I)->getOpcode() != FrameDestroyOpcode)
+    return false;
+
+  // Now, go through the vector, and see that we don't have any gaps,
+  // but only a series of 32-bit MOVs.
+  
+  int64_t ExpectedDist = 0;
+  auto MMI = MovVector.begin(), MME = MovVector.end();
+  for (; MMI != MME; ++MMI, ExpectedDist += 4)
+    if (*MMI == nullptr)
+      break;
+  
+  // If the call had no parameters, do nothing
+  if (!ExpectedDist)
+    return false;
+
+  // We are either at the last parameter, or a gap. 
+  // Make sure it's not a gap
+  for (; MMI != MME; ++MMI)
+    if (*MMI != nullptr)
+      return false;
+
+  // Ok, we can in fact do the transformation for this call.
+  // Do not remove the FrameSetup instruction, but adjust the parameters.
+  // PEI will end up finalizing the handling of this.
+  FrameSetup->getOperand(1).setImm(ExpectedDist);
+
+  DebugLoc DL = I->getDebugLoc();
+  // Now, iterate through the vector in reverse order, and replace the movs
+  // with pushes. MOVmi/MOVmr doesn't have any defs, so no need to 
+  // replace uses.
+  for (int Idx = (ExpectedDist / 4) - 1; Idx >= 0; --Idx) {
+    MachineBasicBlock::iterator MOV = *MovVector[Idx];
+    MachineOperand PushOp = MOV->getOperand(X86::AddrNumOperands);
+    if (MOV->getOpcode() == X86::MOV32mi) {
+      unsigned PushOpcode = X86::PUSHi32;
+      // If the operand is a small (8-bit) immediate, we can use a
+      // PUSH instruction with a shorter encoding.
+      // Note that isImm() may fail even though this is a MOVmi, because
+      // the operand can also be a symbol.
+      if (PushOp.isImm()) {
+        int64_t Val = PushOp.getImm();
+        if (isInt<8>(Val))
+          PushOpcode = X86::PUSH32i8;
+      }
+      BuildMI(MBB, Call, DL, TII->get(PushOpcode)).addOperand(PushOp);
+    } else {
+      unsigned int Reg = PushOp.getReg();
+
+      // If PUSHrmm is not slow on this target, try to fold the source of the
+      // push into the instruction.
+      const X86Subtarget &ST = MF.getTarget().getSubtarget<X86Subtarget>();
+      bool SlowPUSHrmm = ST.isAtom() || ST.isSLM();
+
+      // Check that this is legal to fold. Right now, we're extremely
+      // conservative about that.
+      MachineInstr *DefMov = nullptr;
+      if (!SlowPUSHrmm && (DefMov = canFoldIntoRegPush(FrameSetup, Reg))) {
+        MachineInstr *Push = BuildMI(MBB, Call, DL, TII->get(X86::PUSH32rmm));
+
+        unsigned NumOps = DefMov->getDesc().getNumOperands();
+        for (unsigned i = NumOps - X86::AddrNumOperands; i != NumOps; ++i)
+          Push->addOperand(DefMov->getOperand(i));
+
+        DefMov->eraseFromParent();
+      } else {
+        BuildMI(MBB, Call, DL, TII->get(X86::PUSH32r)).addReg(Reg).getInstr();
+      }
+    }
+
+    MBB.erase(MOV);
+  }
+
+  // The stack-pointer copy is no longer used in the call sequences.
+  // There should not be any other users, but we can't commit to that, so:
+  if (MRI->use_empty(SPCopy->getOperand(0).getReg()))
+    SPCopy->eraseFromParent();
+
+  // Once we've done this, we need to make sure PEI doesn't assume a reserved
+  // frame.
+  X86MachineFunctionInfo *FuncInfo = MF.getInfo<X86MachineFunctionInfo>();
+  FuncInfo->setHasPushSequences(true);
+
+  return true;
+}
+
+MachineInstr *X86CallFrameOptimization::canFoldIntoRegPush(
+    MachineBasicBlock::iterator FrameSetup, unsigned Reg) {
+  // Do an extremely restricted form of load folding.
+  // ISel will often create patterns like:
+  // movl    4(%edi), %eax
+  // movl    8(%edi), %ecx
+  // movl    12(%edi), %edx
+  // movl    %edx, 8(%esp)
+  // movl    %ecx, 4(%esp)
+  // movl    %eax, (%esp)
+  // call
+  // Get rid of those with prejudice.
+  if (!TargetRegisterInfo::isVirtualRegister(Reg))
+    return nullptr;
+
+  // Make sure this is the only use of Reg.
+  if (!MRI->hasOneNonDBGUse(Reg))
+    return nullptr;
+
+  MachineBasicBlock::iterator DefMI = MRI->getVRegDef(Reg);
+
+  // Make sure the def is a MOV from memory.
+  // If the def is an another block, give up.
+  if (DefMI->getOpcode() != X86::MOV32rm ||
+      DefMI->getParent() != FrameSetup->getParent())
+    return nullptr;
+
+  // Be careful with movs that load from a stack slot, since it may get
+  // resolved incorrectly.
+  // TODO: Again, we already have the infrastructure, so this should work.
+  if (!DefMI->getOperand(1).isReg())
+    return nullptr;
+
+  // Now, make sure everything else up until the ADJCALLSTACK is a sequence
+  // of MOVs. To be less conservative would require duplicating a lot of the
+  // logic from PeepholeOptimizer.
+  // FIXME: A possibly better approach would be to teach the PeepholeOptimizer
+  // to be smarter about folding into pushes. 
+  for (auto I = DefMI; I != FrameSetup; ++I)
+    if (I->getOpcode() != X86::MOV32rm)
+      return nullptr;
+
+  return DefMI;
+}

lib/Target/X86/X86FastISel.cpp

bool X86FastISel::fastLowerCall(CallLoweringInfo &CLI) {
   CCInfo.AnalyzeCallOperands(OutVTs, OutFlags, CC_X86);
 
   // Get a count of how many bytes are to be pushed on the stack.
   unsigned NumBytes = CCInfo.getNextStackOffset();
 
   // Issue CALLSEQ_START
   unsigned AdjStackDown = TII.getCallFrameSetupOpcode();
   BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc, TII.get(AdjStackDown))
-    .addImm(NumBytes);
+    .addImm(NumBytes).addImm(0);
 
   // Walk the register/memloc assignments, inserting copies/loads.
   const X86RegisterInfo *RegInfo = static_cast<const X86RegisterInfo *>(
       TM.getSubtargetImpl()->getRegisterInfo());
   for (unsigned i = 0, e = ArgLocs.size(); i != e; ++i) {
     CCValAssign const &VA = ArgLocs[i];
     const Value *ArgVal = OutVals[VA.getValNo()];
     MVT ArgVT = OutVTs[VA.getValNo()];

lib/Target/X86/X86FrameLowering.h

public:
 
   bool restoreCalleeSavedRegisters(MachineBasicBlock &MBB,
                                   MachineBasicBlock::iterator MI,
                                   const std::vector<CalleeSavedInfo> &CSI,
                                   const TargetRegisterInfo *TRI) const override;
 
   bool hasFP(const MachineFunction &MF) const override;
   bool hasReservedCallFrame(const MachineFunction &MF) const override;
+  bool canSimplifyCallFramePseudos(const MachineFunction &MF) const override;
+  bool needsFrameIndexResolution(const MachineFunction &MF) const override;
 
   int getFrameIndexOffset(const MachineFunction &MF, int FI) const override;
   int getFrameIndexReference(const MachineFunction &MF, int FI,
                              unsigned &FrameReg) const override;
 
   int getFrameIndexOffsetFromSP(const MachineFunction &MF, int FI) const;
   int getFrameIndexReferenceFromSP(const MachineFunction &MF, int FI,
                                    unsigned &FrameReg) const override;

lib/Target/X86/X86FrameLowering.cpp

 #include <cstdlib>
 
 using namespace llvm;
 
 // FIXME: completely move here.
 extern cl::opt<bool> ForceStackAlign;
 
 bool X86FrameLowering::hasReservedCallFrame(const MachineFunction &MF) const {
-  return !MF.getFrameInfo()->hasVarSizedObjects();
+  return !MF.getFrameInfo()->hasVarSizedObjects() &&
+         !MF.getInfo<X86MachineFunctionInfo>()->getHasPushSequences();
+}
+
+/// canSimplifyCallFramePseudos - If there is a reserved call frame, the
+/// call frame pseudos can be simplified.  Having a FP, as in the default
+/// implementation, is not sufficient here since we can't always use it.
+/// Use a more nuanced condition.
+bool
+X86FrameLowering::canSimplifyCallFramePseudos(const MachineFunction &MF) const {
+  const X86RegisterInfo *TRI = static_cast<const X86RegisterInfo *>
+                               (MF.getSubtarget().getRegisterInfo());
+  return hasReservedCallFrame(MF) ||
+         (hasFP(MF) && !TRI->needsStackRealignment(MF))
+         || TRI->hasBasePointer(MF);
+}
+
+// needsFrameIndexResolution - do we need to perform FI resolution for
+// this function. Normally, this is required only when the function
+// has any stack objects. However, FI resolution actually has another job,
+// not apparent from the title - it resolves callframesetup/destroy 
+// that were not simplified earlier.
+// So, this is required for x86 functions that have push sequences even
+// when there are no stack objects.
+bool
+X86FrameLowering::needsFrameIndexResolution(const MachineFunction &MF) const {
+  return MF.getFrameInfo()->hasStackObjects() ||
+         MF.getInfo<X86MachineFunctionInfo>()->getHasPushSequences();
 }
 
 /// hasFP - Return true if the specified function should have a dedicated frame
 /// pointer register.  This is true if the function has variable sized allocas
 /// or if frame pointer elimination is disabled.
 bool X86FrameLowering::hasFP(const MachineFunction &MF) const {
   const MachineFrameInfo *MFI = MF.getFrameInfo();
   const MachineModuleInfo &MMI = MF.getMMI();
if (isInt<8>(Imm))
       return X86::AND64ri8;
     return X86::AND64ri32;
   }
   if (isInt<8>(Imm))
     return X86::AND32ri8;
   return X86::AND32ri;
 }
 
-static unsigned getPUSHiOpcode(bool IsLP64, MachineOperand MO) {
-  // We don't support LP64 for now.
-  assert(!IsLP64);
-
-  if (MO.isImm() && isInt<8>(MO.getImm()))
-    return X86::PUSH32i8;
-
-  return X86::PUSHi32;;
-}
-
 static unsigned getLEArOpcode(unsigned IsLP64) {
   return IsLP64 ? X86::LEA64r : X86::LEA32r;
 }
 
 /// findDeadCallerSavedReg - Return a caller-saved register that isn't live
 /// when it reaches the "return" instruction. We can then pop a stack object
 /// to this register without worry about clobbering it.
 static unsigned findDeadCallerSavedReg(MachineBasicBlock &MBB,
if (MaxStack > Guaranteed) {
     incStackMBB->addSuccessor(&prologueMBB, 99);
     incStackMBB->addSuccessor(incStackMBB, 1);
   }
 #ifdef XDEBUG
   MF.verify();
 #endif
 }
 
-bool X86FrameLowering::
-convertArgMovsToPushes(MachineFunction &MF, MachineBasicBlock &MBB,
-                       MachineBasicBlock::iterator I, uint64_t Amount) const {
-  const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
-  const X86RegisterInfo &RegInfo = *static_cast<const X86RegisterInfo *>(
-    MF.getSubtarget().getRegisterInfo());
-  unsigned StackPtr = RegInfo.getStackRegister();
-
-  // Scan the call setup sequence for the pattern we're looking for.
-  // We only handle a simple case now - a sequence of MOV32mi or MOV32mr
-  // instructions, that push a sequence of 32-bit values onto the stack, with
-  // no gaps.  
-  std::map<int64_t, MachineBasicBlock::iterator> MovMap;
-  do {
-    int Opcode = I->getOpcode();
-    if (Opcode != X86::MOV32mi && Opcode != X86::MOV32mr)
-      break;
- 
-    // We only want movs of the form:
-    // movl imm/r32, k(%ecx)
-    // If we run into something else, bail
-    // Note that AddrBaseReg may, counterintuitively, not be a register...
-    if (!I->getOperand(X86::AddrBaseReg).isReg() || 
-        (I->getOperand(X86::AddrBaseReg).getReg() != StackPtr) ||
-        !I->getOperand(X86::AddrScaleAmt).isImm() ||
-        (I->getOperand(X86::AddrScaleAmt).getImm() != 1) ||
-        (I->getOperand(X86::AddrIndexReg).getReg() != X86::NoRegister) ||
-        (I->getOperand(X86::AddrSegmentReg).getReg() != X86::NoRegister) ||
-        !I->getOperand(X86::AddrDisp).isImm())
-      return false;
-
-    int64_t StackDisp = I->getOperand(X86::AddrDisp).getImm();
-    
-    // We don't want to consider the unaligned case.
-    if (StackDisp % 4)
-      return false;
-
-    // If the same stack slot is being filled twice, something's fishy.
-    if (!MovMap.insert(std::pair<int64_t, MachineInstr*>(StackDisp, I)).second)
-      return false;
-
-    ++I;
-  } while (I != MBB.end());
-
-  // We now expect the end of the sequence - a call and a stack adjust.
-  if (I == MBB.end())
-    return false;
-  if (!I->isCall())
-    return false;
-  MachineBasicBlock::iterator Call = I;
-  if ((++I)->getOpcode() != TII.getCallFrameDestroyOpcode())
-    return false;
-
-  // Now, go through the map, and see that we don't have any gaps,
-  // but only a series of 32-bit MOVs.
-  // Since std::map provides ordered iteration, the original order
-  // of the MOVs doesn't matter.
-  int64_t ExpectedDist = 0;
-  for (auto MMI = MovMap.begin(), MME = MovMap.end(); MMI != MME; 
-       ++MMI, ExpectedDist += 4)
-    if (MMI->first != ExpectedDist)
-      return false;
-
-  // Ok, everything looks fine. Do the transformation.
-  DebugLoc DL = I->getDebugLoc();
-
-  // It's possible the original stack adjustment amount was larger than
-  // that done by the pushes. If so, we still need a SUB.
-  Amount -= ExpectedDist;
-  if (Amount) {
-    MachineInstr* Sub = BuildMI(MBB, Call, DL,
-                          TII.get(getSUBriOpcode(false, Amount)), StackPtr)
-                  .addReg(StackPtr).addImm(Amount);
-    Sub->getOperand(3).setIsDead();
-  }
-
-  // Now, iterate through the map in reverse order, and replace the movs
-  // with pushes. MOVmi/MOVmr doesn't have any defs, so need to replace uses.
-  for (auto MMI = MovMap.rbegin(), MME = MovMap.rend(); MMI != MME; ++MMI) {
-    MachineBasicBlock::iterator MOV = MMI->second;
-    MachineOperand PushOp = MOV->getOperand(X86::AddrNumOperands);
-
-    // Replace MOVmr with PUSH32r, and MOVmi with PUSHi of appropriate size
-    int PushOpcode = X86::PUSH32r;
-    if (MOV->getOpcode() == X86::MOV32mi)
-      PushOpcode = getPUSHiOpcode(false, PushOp);
-
-    BuildMI(MBB, Call, DL, TII.get(PushOpcode)).addOperand(PushOp);
-    MBB.erase(MOV);
-  }
-
-  return true;
-}
-
 void X86FrameLowering::
 eliminateCallFramePseudoInstr(MachineFunction &MF, MachineBasicBlock &MBB,
                               MachineBasicBlock::iterator I) const {
   const TargetInstrInfo &TII = *MF.getSubtarget().getInstrInfo();
   const X86RegisterInfo &RegInfo = *static_cast<const X86RegisterInfo *>(
                                        MF.getSubtarget().getRegisterInfo());
   unsigned StackPtr = RegInfo.getStackRegister();
   bool reserveCallFrame = hasReservedCallFrame(MF);
   int Opcode = I->getOpcode();
   bool isDestroy = Opcode == TII.getCallFrameDestroyOpcode();
   const X86Subtarget &STI = MF.getTarget().getSubtarget<X86Subtarget>();
   bool IsLP64 = STI.isTarget64BitLP64();
   DebugLoc DL = I->getDebugLoc();
   uint64_t Amount = !reserveCallFrame ? I->getOperand(0).getImm() : 0;
-  uint64_t CalleeAmt = isDestroy ? I->getOperand(1).getImm() : 0;
+  uint64_t InternalAmt = (isDestroy || Amount) ? I->getOperand(1).getImm() : 0;
   I = MBB.erase(I);
 
   if (!reserveCallFrame) {
     // If the stack pointer can be changed after prologue, turn the
     // adjcallstackup instruction into a 'sub ESP, <amt>' and the
     // adjcallstackdown instruction into 'add ESP, <amt>'
     if (Amount == 0)
       return;
 
     // We need to keep the stack aligned properly.  To do this, we round the
     // amount of space needed for the outgoing arguments up to the next
     // alignment boundary.
     unsigned StackAlign = MF.getTarget()
                               .getSubtargetImpl()
                               ->getFrameLowering()
                               ->getStackAlignment();
     Amount = (Amount + StackAlign - 1) / StackAlign * StackAlign;
 
     MachineInstr *New = nullptr;
-    if (Opcode == TII.getCallFrameSetupOpcode()) {
-      // Try to convert movs to the stack into pushes.
-      // We currently only look for a pattern that appears in 32-bit
-      // calling conventions.
-      if (!IsLP64 && convertArgMovsToPushes(MF, MBB, I, Amount))
-        return;
 
-      New = BuildMI(MF, DL, TII.get(getSUBriOpcode(IsLP64, Amount)),
-                    StackPtr)
-        .addReg(StackPtr)
-        .addImm(Amount);
+    // Factor out the amount that gets handled inside the sequence
+    // (Pushes of argument for frame setup, callee pops for frame destroy)
+    Amount -= InternalAmt;
+
+    if (Amount) {
+      if (Opcode == TII.getCallFrameSetupOpcode()) {
+        New = BuildMI(MF, DL, TII.get(getSUBriOpcode(IsLP64, Amount)), StackPtr)
+          .addReg(StackPtr).addImm(Amount);
       } else {
         assert(Opcode == TII.getCallFrameDestroyOpcode());
 
-      // Factor out the amount the callee already popped.
-      Amount -= CalleeAmt;
-
-      if (Amount) {
         unsigned Opc = getADDriOpcode(IsLP64, Amount);
         New = BuildMI(MF, DL, TII.get(Opc), StackPtr)
           .addReg(StackPtr).addImm(Amount);
       }
     }
 
     if (New) {
       // The EFLAGS implicit def is dead.
       New->getOperand(3).setIsDead();
 
       // Replace the pseudo instruction with a new instruction.
       MBB.insert(I, New);
     }
 
     return;
   }
 
-  if (Opcode == TII.getCallFrameDestroyOpcode() && CalleeAmt) {
+  if (Opcode == TII.getCallFrameDestroyOpcode() && InternalAmt) {
     // If we are performing frame pointer elimination and if the callee pops
     // something off the stack pointer, add it back.  We do this until we have
     // more advanced stack pointer tracking ability.
-    unsigned Opc = getSUBriOpcode(IsLP64, CalleeAmt);
+    unsigned Opc = getSUBriOpcode(IsLP64, InternalAmt);
     MachineInstr *New = BuildMI(MF, DL, TII.get(Opc), StackPtr)
-      .addReg(StackPtr).addImm(CalleeAmt);
+      .addReg(StackPtr).addImm(InternalAmt);
 
     // The EFLAGS implicit def is dead.
     New->getOperand(3).setIsDead();
 
     // We are not tracking the stack pointer adjustment by the callee, so make
     // sure we restore the stack pointer immediately after the call, there may
     // be spill code inserted between the CALL and ADJCALLSTACKUP instructions.
     MachineBasicBlock::iterator B = MBB.begin();
     while (I != B && !std::prev(I)->isCall())
       --I;
     MBB.insert(I, New);
   }
 }
 

lib/Target/X86/X86InstrCompiler.td

 
 
 // ADJCALLSTACKDOWN/UP implicitly use/def ESP because they may be expanded into
 // a stack adjustment and the codegen must know that they may modify the stack
 // pointer before prolog-epilog rewriting occurs.
 // Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
 // sub / add which can clobber EFLAGS.
 let Defs = [ESP, EFLAGS], Uses = [ESP] in {
-def ADJCALLSTACKDOWN32 : I<0, Pseudo, (outs), (ins i32imm:$amt),
+def ADJCALLSTACKDOWN32 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
                            "#ADJCALLSTACKDOWN",
-                           [(X86callseq_start timm:$amt)]>,
+                           []>,
                           Requires<[NotLP64]>;
 def ADJCALLSTACKUP32   : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
                            "#ADJCALLSTACKUP",
                            [(X86callseq_end timm:$amt1, timm:$amt2)]>,
                           Requires<[NotLP64]>;
 }
+def : Pat<(X86callseq_start timm:$amt1),
+          (ADJCALLSTACKDOWN32 i32imm:$amt1, 0)>, Requires<[NotLP64]>;
+
 
 // ADJCALLSTACKDOWN/UP implicitly use/def RSP because they may be expanded into
 // a stack adjustment and the codegen must know that they may modify the stack
 // pointer before prolog-epilog rewriting occurs.
 // Pessimistically assume ADJCALLSTACKDOWN / ADJCALLSTACKUP will become
 // sub / add which can clobber EFLAGS.
 let Defs = [RSP, EFLAGS], Uses = [RSP] in {
-def ADJCALLSTACKDOWN64 : I<0, Pseudo, (outs), (ins i32imm:$amt),
+def ADJCALLSTACKDOWN64 : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
                            "#ADJCALLSTACKDOWN",
-                           [(X86callseq_start timm:$amt)]>,
+                           []>,
                           Requires<[IsLP64]>;
 def ADJCALLSTACKUP64   : I<0, Pseudo, (outs), (ins i32imm:$amt1, i32imm:$amt2),
                            "#ADJCALLSTACKUP",
                            [(X86callseq_end timm:$amt1, timm:$amt2)]>,
                           Requires<[IsLP64]>;
 }
-
+def : Pat<(X86callseq_start timm:$amt1),
+          (ADJCALLSTACKDOWN64 i32imm:$amt1, 0)>, Requires<[IsLP64]>;
 
 
 // x86-64 va_start lowering magic.
 let usesCustomInserter = 1, Defs = [EFLAGS] in {
 def VASTART_SAVE_XMM_REGS : I<0, Pseudo,
                               (outs),
                               (ins GR8:$al,
                                    i64imm:$regsavefi, i64imm:$offset,

lib/Target/X86/X86InstrInfo.h

public:
   explicit X86InstrInfo(X86Subtarget &STI);
 
   /// getRegisterInfo - TargetInstrInfo is a superset of MRegister info.  As
   /// such, whenever a client has an instance of instruction info, it should
   /// always be able to get register info as well (through this method).
   ///
   const X86RegisterInfo &getRegisterInfo() const { return RI; }
 
+  /// getSPAdjust - This returns the stack pointer adjustment made by
+  /// this instruction. For x86, we need to handle more complex call
+  /// sequences involving PUSHes.
+  int getSPAdjust(const MachineInstr *MI) const override;
+
   /// isCoalescableExtInstr - Return true if the instruction is a "coalescable"
   /// extension instruction. That is, it's like a copy where it's legal for the
   /// source to overlap the destination. e.g. X86::MOVSX64rr32. If this returns
   /// true, then it's expected the pre-extension value is available as a subreg
   /// of the result register. This also returns the sub-register index in
   /// SubIdx.
   bool isCoalescableExtInstr(const MachineInstr &MI,
                              unsigned &SrcReg, unsigned &DstReg,

lib/Target/X86/X86InstrInfo.cpp

case X86::MOVSX64rr32:
       break;
     }
     return true;
   }
   }
   return false;
 }
 
+int X86InstrInfo::getSPAdjust(const MachineInstr *MI) const {
+  const MachineFunction *MF = MI->getParent()->getParent();
+  const TargetFrameLowering *TFI = MF->getSubtarget().getFrameLowering();
+
+  if (MI->getOpcode() == getCallFrameSetupOpcode() ||
+      MI->getOpcode() == getCallFrameDestroyOpcode()) {
+    unsigned StackAlign = TFI->getStackAlignment();
+    int SPAdj = (MI->getOperand(0).getImm() + StackAlign - 1) / StackAlign * 
+                 StackAlign;
+
+    SPAdj -= MI->getOperand(1).getImm();
+
+    if (MI->getOpcode() == getCallFrameSetupOpcode())
+      return SPAdj;
+    else
+      return -SPAdj;
+  }
+  
+  // FIXME: This is horrible.
+  // The problem is that to know whether a call adjusts the stack, we
+  // need information that is bound to the following ADJCALLSTACKUP
+  // pseudo. This leaves us with a choice:
+  // 1) Look for the next ADJCALLSTACKUP, like the code below. This
+  // works, but certainly doesn't belong here.
+  // 2) Factor out the code that makes the decision on whether a function
+  // has callee-pop, and call it here as well as when creating the 
+  // ADJCALLSTACKUP. This is also insanely ugly.
+  // 3) Don't use SP adjustments (in practice, disable mov -> push 
+  // tranformation) in the presence of callee-pop. This would be 
+  // unfortunate, but may be a good temporary solution.
+  // 4) ?
+  if (MI->isCall()) {
+    const MachineBasicBlock* MBB = MI->getParent();
+    auto I = ++MachineBasicBlock::const_iterator(MI);
+    for (auto E = MBB->end(); I != E; ++I) {
+      if (I->getOpcode() == getCallFrameDestroyOpcode() ||
+          I->isCall())
+        break;
+    }
+
+    // If we could not find a frame destroy opcode, then it has already
+    // been simplified, so we don't care.
+    if (I->getOpcode() != getCallFrameDestroyOpcode())
+      return 0;
+
+    return -(I->getOperand(1).getImm());
+  }
+
+  // Currently handle only PUSHes we can reasonably expect to see
+  // in call sequences
+  switch (MI->getOpcode()) {
+  default: 
+    return 0;
+  case X86::PUSH32i8:
+  case X86::PUSH32r:
+  case X86::PUSH32rmm:
+  case X86::PUSH32rmr:
+  case X86::PUSHi32:
+    return 4;
+  }
+}
+
 /// isFrameOperand - Return true and the FrameIndex if the specified
 /// operand and follow operands form a reference to the stack frame.
 bool X86InstrInfo::isFrameOperand(const MachineInstr *MI, unsigned int Op,
                                   int &FrameIndex) const {
   if (MI->getOperand(Op+X86::AddrBaseReg).isFI() &&
       MI->getOperand(Op+X86::AddrScaleAmt).isImm() &&
       MI->getOperand(Op+X86::AddrIndexReg).isReg() &&
       MI->getOperand(Op+X86::AddrDisp).isImm() &&

lib/Target/X86/X86MachineFunctionInfo.h

class X86MachineFunctionInfo : public MachineFunctionInfo {
   unsigned VarArgsGPOffset;
   /// VarArgsFPOffset - X86-64 vararg func fp reg offset.
   unsigned VarArgsFPOffset;
   /// ArgumentStackSize - The number of bytes on stack consumed by the arguments
   /// being passed on the stack.
   unsigned ArgumentStackSize;
   /// NumLocalDynamics - Number of local-dynamic TLS accesses.
   unsigned NumLocalDynamics;
+  /// HasPushSequences - Keeps track of whether this function uses sequences
+  /// of pushes to pass function parameters.
+  bool HasPushSequences;
 
 private:
   /// ForwardedMustTailRegParms - A list of virtual and physical registers
   /// that must be forwarded to every musttail call.
   SmallVector<ForwardedRegister, 1> ForwardedMustTailRegParms;
 
 public:
   X86MachineFunctionInfo() : ForceFramePointer(false),
                              RestoreBasePointerOffset(0),
                              CalleeSavedFrameSize(0),
                              BytesToPopOnReturn(0),
                              ReturnAddrIndex(0),
                              TailCallReturnAddrDelta(0),
                              SRetReturnReg(0),
                              GlobalBaseReg(0),
                              VarArgsFrameIndex(0),
                              RegSaveFrameIndex(0),
                              VarArgsGPOffset(0),
                              VarArgsFPOffset(0),
                              ArgumentStackSize(0),
-                             NumLocalDynamics(0) {}
+                             NumLocalDynamics(0),
+                             HasPushSequences(false) {}
 
   explicit X86MachineFunctionInfo(MachineFunction &MF)
     : ForceFramePointer(false),
       RestoreBasePointerOffset(0),
       CalleeSavedFrameSize(0),
       BytesToPopOnReturn(0),
       ReturnAddrIndex(0),
       TailCallReturnAddrDelta(0),
       SRetReturnReg(0),
       GlobalBaseReg(0),
       VarArgsFrameIndex(0),
       RegSaveFrameIndex(0),
       VarArgsGPOffset(0),
       VarArgsFPOffset(0),
       ArgumentStackSize(0),
-      NumLocalDynamics(0) {}
+      NumLocalDynamics(0),
+      HasPushSequences(false) {}
 
   bool getForceFramePointer() const { return ForceFramePointer;}
   void setForceFramePointer(bool forceFP) { ForceFramePointer = forceFP; }
 
+  bool getHasPushSequences() const { return HasPushSequences; }
+  void setHasPushSequences(bool HasPush) { HasPushSequences = HasPush; }
+
   bool getRestoreBasePointer() const { return RestoreBasePointerOffset!=0; }
   void setRestoreBasePointer(const MachineFunction *MF);
   int getRestoreBasePointerOffset() const {return RestoreBasePointerOffset; }
 
   unsigned getCalleeSavedFrameSize() const { return CalleeSavedFrameSize; }
   void setCalleeSavedFrameSize(unsigned bytes) { CalleeSavedFrameSize = bytes; }
 
   unsigned getBytesToPopOnReturn() const { return BytesToPopOnReturn; }

lib/Target/X86/X86RegisterInfo.cpp

bool X86RegisterInfo::hasReservedSpillSlot(const MachineFunction &MF,
   // this function neither used nor tested.
   llvm_unreachable("Unused function on X86. Otherwise need a test case.");
 }
 
 void
 X86RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
                                      int SPAdj, unsigned FIOperandNum,
                                      RegScavenger *RS) const {
-  assert(SPAdj == 0 && "Unexpected");
-
   MachineInstr &MI = *II;
   MachineFunction &MF = *MI.getParent()->getParent();
   const TargetFrameLowering *TFI = MF.getSubtarget().getFrameLowering();
   int FrameIndex = MI.getOperand(FIOperandNum).getIndex();
   unsigned BasePtr;
 
   unsigned Opc = MI.getOpcode();
   bool AfterFPPop = Opc == X86::TAILJMPm64 || Opc == X86::TAILJMPm;
X86RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
   int FIOffset;
   if (AfterFPPop) {
     // Tail call jmp happens after FP is popped.
     const MachineFrameInfo *MFI = MF.getFrameInfo();
     FIOffset = MFI->getObjectOffset(FrameIndex) - TFI->getOffsetOfLocalArea();
   } else
     FIOffset = TFI->getFrameIndexOffset(MF, FrameIndex);
 
+  if (BasePtr == StackPtr)
+    FIOffset += SPAdj;
+
   // The frame index format for stackmaps and patchpoints is different from the
   // X86 format. It only has a FI and an offset.
   if (Opc == TargetOpcode::STACKMAP || Opc == TargetOpcode::PATCHPOINT) {
     assert(BasePtr == FramePtr && "Expected the FP as base register");
     int64_t Offset = MI.getOperand(FIOperandNum + 1).getImm() + FIOffset;
     MI.getOperand(FIOperandNum + 1).ChangeToImmediate(Offset);
     return;
   }

lib/Target/X86/X86TargetMachine.cpp

public:
 
   const X86Subtarget &getX86Subtarget() const {
     return *getX86TargetMachine().getSubtargetImpl();
   }
 
   void addIRPasses() override;
   bool addInstSelector() override;
   bool addILPOpts() override;
+  void addPreRegAlloc() override;
   void addPostRegAlloc() override;
   void addPreEmitPass() override;
 };
 } // namespace
 
 TargetPassConfig *X86TargetMachine::createPassConfig(PassManagerBase &PM) {
   return new X86PassConfig(this, PM);
 }
bool X86PassConfig::addInstSelector() {
   return false;
 }
 
 bool X86PassConfig::addILPOpts() {
   addPass(&EarlyIfConverterID);
   return true;
 }
 
+void X86PassConfig::addPreRegAlloc() {
+  addPass(createX86CallFrameOptimization());
+}
+
 void X86PassConfig::addPostRegAlloc() {
   addPass(createX86FloatingPointStackifierPass());
 }
 
 void X86PassConfig::addPreEmitPass() {
   if (getOptLevel() != CodeGenOpt::None && getX86Subtarget().hasSSE2())
     addPass(createExecutionDependencyFixPass(&X86::VR128RegClass));
 
   if (UseVZeroUpper)
     addPass(createX86IssueVZeroUpperPass());
 
   if (getOptLevel() != CodeGenOpt::None) {
     addPass(createX86PadShortFunctions());
     addPass(createX86FixupLEAs());
   }
 }

test/CodeGen/X86/inalloca-invoke.ll

; CHECK:  leal    12(%[[beg]]), %[[end:[^ ]*]]
 
   call void @begin(%Iter* sret %temp.lvalue)
 ; CHECK:  calll _begin
 
   invoke void @plus(%Iter* sret %end, %Iter* %temp.lvalue, i32 4)
           to label %invoke.cont unwind label %lpad
 
 ;  Uses end as sret param.
-; CHECK:  movl %[[end]], (%esp)
+; CHECK:  pushl %[[end]]
 ; CHECK:  calll _plus
 
 invoke.cont:
   call void @begin(%Iter* sret %beg)
 
 ; CHECK:  pushl %[[beg]]
 ; CHECK:  calll _begin
 

test/CodeGen/X86/movtopush.ll

 ; RUN: llc < %s -mtriple=i686-windows | FileCheck %s -check-prefix=NORMAL
+; RUN: llc < %s -mtriple=x86_64-windows | FileCheck %s -check-prefix=X64
 ; RUN: llc < %s -mtriple=i686-windows -force-align-stack -stack-alignment=32 | FileCheck %s -check-prefix=ALIGNED 
+
 declare void @good(i32 %a, i32 %b, i32 %c, i32 %d)
 declare void @inreg(i32 %a, i32 inreg %b, i32 %c, i32 %d)
 
 ; Here, we should have a reserved frame, so we don't expect pushes
-; NORMAL-LABEL: test1
+; NORMAL-LABEL: test1:
 ; NORMAL: subl    $16, %esp
 ; NORMAL-NEXT: movl    $4, 12(%esp)
 ; NORMAL-NEXT: movl    $3, 8(%esp)
 ; NORMAL-NEXT: movl    $2, 4(%esp)
 ; NORMAL-NEXT: movl    $1, (%esp)
 ; NORMAL-NEXT: call
+; NORMAL-NEXT: addl $16, %esp
 define void @test1() {
 entry:
   call void @good(i32 1, i32 2, i32 3, i32 4)
   ret void
 }
 
-; Here, we expect a sequence of 4 immediate pushes
-; NORMAL-LABEL: test2
+; We're optimizing for code size, so we should get pushes for x86,
+; even though there is a reserved call frame.
+; Make sure we don't touch x86-64
+; NORMAL-LABEL: test1b:
+; NORMAL-NOT: subl {{.*}} %esp
+; NORMAL: pushl   $4
+; NORMAL-NEXT: pushl   $3
+; NORMAL-NEXT: pushl   $2
+; NORMAL-NEXT: pushl   $1
+; NORMAL-NEXT: call
+; NORMAL-NEXT: addl $16, %esp
+; X64-LABEL: test1b:
+; X64: movl    $1, %ecx
+; X64-NEXT: movl    $2, %edx
+; X64-NEXT: movl    $3, %r8d
+; X64-NEXT: movl    $4, %r9d
+; X64-NEXT: callq   good
+define void @test1b() optsize {
+entry:
+  call void @good(i32 1, i32 2, i32 3, i32 4)
+  ret void
+}
+
+; Same as above, but for minsize
+; NORMAL-LABEL: test1c:
+; NORMAL-NOT: subl {{.*}} %esp
+; NORMAL: pushl   $4
+; NORMAL-NEXT: pushl   $3
+; NORMAL-NEXT: pushl   $2
+; NORMAL-NEXT: pushl   $1
+; NORMAL-NEXT: call
+; NORMAL-NEXT: addl $16, %esp
+define void @test1c() minsize {
+entry:
+  call void @good(i32 1, i32 2, i32 3, i32 4)
+  ret void
+}
+
+; If we have a reserved frame, we should have pushes
+; NORMAL-LABEL: test2:
 ; NORMAL-NOT: subl {{.*}} %esp
 ; NORMAL: pushl   $4
 ; NORMAL-NEXT: pushl   $3
 ; NORMAL-NEXT: pushl   $2
 ; NORMAL-NEXT: pushl   $1
 ; NORMAL-NEXT: call
 define void @test2(i32 %k) {
 entry:
   %a = alloca i32, i32 %k
   call void @good(i32 1, i32 2, i32 3, i32 4)
   ret void
 }
 
 ; Again, we expect a sequence of 4 immediate pushes
 ; Checks that we generate the right pushes for >8bit immediates
-; NORMAL-LABEL: test2b
+; NORMAL-LABEL: test2b:
 ; NORMAL-NOT: subl {{.*}} %esp
 ; NORMAL: pushl   $4096
 ; NORMAL-NEXT: pushl   $3072
 ; NORMAL-NEXT: pushl   $2048
 ; NORMAL-NEXT: pushl   $1024
 ; NORMAL-NEXT: call
-define void @test2b(i32 %k) {
+; NORMAL-NEXT: addl $16, %esp
+define void @test2b() optsize {
 entry:
-  %a = alloca i32, i32 %k
   call void @good(i32 1024, i32 2048, i32 3072, i32 4096)
   ret void
 }
 
 ; The first push should push a register
-; NORMAL-LABEL: test3
+; NORMAL-LABEL: test3:
 ; NORMAL-NOT: subl {{.*}} %esp
 ; NORMAL: pushl   $4
 ; NORMAL-NEXT: pushl   $3
 ; NORMAL-NEXT: pushl   $2
 ; NORMAL-NEXT: pushl   %e{{..}}
 ; NORMAL-NEXT: call
-define void @test3(i32 %k) {
+; NORMAL-NEXT: addl $16, %esp
+define void @test3(i32 %k) optsize {
 entry:
-  %a = alloca i32, i32 %k
   call void @good(i32 %k, i32 2, i32 3, i32 4)
   ret void
 }
 
 ; We don't support weird calling conventions
-; NORMAL-LABEL: test4
+; NORMAL-LABEL: test4:
 ; NORMAL: subl    $12, %esp
 ; NORMAL-NEXT: movl    $4, 8(%esp)
 ; NORMAL-NEXT: movl    $3, 4(%esp)
 ; NORMAL-NEXT: movl    $1, (%esp)
 ; NORMAL-NEXT: movl    $2, %eax
 ; NORMAL-NEXT: call
-define void @test4(i32 %k) {
+; NORMAL-NEXT: addl $12, %esp
+define void @test4() optsize {
 entry:
-  %a = alloca i32, i32 %k
   call void @inreg(i32 1, i32 2, i32 3, i32 4)
   ret void
 }
 
-; Check that additional alignment is added when the pushes
-; don't add up to the required alignment.
-; ALIGNED-LABEL: test5
+; When there is no reserved call frame, check that additional alignment
+; is added when the pushes don't add up to the required alignment.
+; ALIGNED-LABEL: test5:
 ; ALIGNED: subl    $16, %esp
 ; ALIGNED-NEXT: pushl   $4
 ; ALIGNED-NEXT: pushl   $3
 ; ALIGNED-NEXT: pushl   $2
 ; ALIGNED-NEXT: pushl   $1
 ; ALIGNED-NEXT: call
 define void @test5(i32 %k) {
 entry:
   %a = alloca i32, i32 %k
   call void @good(i32 1, i32 2, i32 3, i32 4)
   ret void
 }
 
 ; Check that pushing the addresses of globals (Or generally, things that 
 ; aren't exactly immediates) isn't broken.
 ; Fixes PR21878.
-; NORMAL-LABEL: test6
+; NORMAL-LABEL: test6:
 ; NORMAL: pushl    $_ext
 ; NORMAL-NEXT: call
 declare void @f(i8*)
 @ext = external constant i8
 
 define void @test6() {
   call void @f(i8* @ext)
   br label %bb
 bb:
   alloca i32
   ret void
 }
+
+; Check that we fold simple cases into the push
+; NORMAL-LABEL: test7:
+; NORMAL-NOT: subl {{.*}} %esp
+; NORMAL: movl 4(%esp), [[EAX:%e..]]
+; NORMAL-NEXT: pushl   $4
+; NORMAL-NEXT: pushl   ([[EAX]])
+; NORMAL-NEXT: pushl   $2
+; NORMAL-NEXT: pushl   $1
+; NORMAL-NEXT: call
+; NORMAL-NEXT: addl $16, %esp
+define void @test7(i32* %ptr) optsize {
+entry:
+  %val = load i32* %ptr
+  call void @good(i32 1, i32 2, i32 %val, i32 4)
+  ret void
+}
+
+; But we don't want to fold stack-relative loads into the push,
+; because the offset will be wrong
+; NORMAL-LABEL: test8:
+; NORMAL-NOT: subl {{.*}} %esp
+; NORMAL: movl 4(%esp), [[EAX:%e..]]
+; NORMAL-NEXT: pushl   $4
+; NORMAL-NEXT: pushl   [[EAX]]
+; NORMAL-NEXT: pushl   $2
+; NORMAL-NEXT: pushl   $1
+; NORMAL-NEXT: call
+; NORMAL-NEXT: addl $16, %esp
+define void @test8(i32* %ptr) optsize {
+entry:
+  %val = ptrtoint i32* %ptr to i32
+  call void @good(i32 1, i32 2, i32 %val, i32 4)
+  ret void
+}
+
+; If one function is using push instructions, and the other isn't
+; (because it has frame-index references), then we must resolve
+; these references correctly.
+; NORMAL-LABEL: test9:
+; NORMAL-NOT: leal (%esp), 
+; NORMAL: pushl $4
+; NORMAL-NEXT: pushl $3
+; NORMAL-NEXT: pushl $2
+; NORMAL-NEXT: pushl $1
+; NORMAL-NEXT: call
+; NORMAL-NEXT: addl $16, %esp
+; NORMAL-NEXT: subl $16, %esp
+; NORMAL-NEXT: leal 16(%esp), [[EAX:%e..]]
+; NORMAL-NEXT: movl    [[EAX]], 12(%esp)
+; NORMAL-NEXT: movl    $7, 8(%esp)
+; NORMAL-NEXT: movl    $6, 4(%esp)
+; NORMAL-NEXT: movl    $5, (%esp)
+; NORMAL-NEXT: call
+; NORMAL-NEXT: addl $16, %esp
+define void @test9() optsize {
+entry:
+  %p = alloca i32, align 4
+  call void @good(i32 1, i32 2, i32 3, i32 4)
+  %0 = ptrtoint i32* %p to i32
+  call void @good(i32 5, i32 6, i32 7, i32 %0)
+  ret void
+}
+
+; We can end up with an indirect call which gets reloaded on the spot.
+; Make sure we reference the correct stack slot - we spill into (%esp)
+; and reload from 16(%esp) due to the pushes.
+; NORMAL-LABEL: test10:
+; NORMAL: movl $_good, [[ALLOC:.*]]
+; NORMAL-NEXT: movl [[ALLOC]], [[EAX:%e..]]
+; NORMAL-NEXT: movl [[EAX]], (%esp) # 4-byte Spill
+; NORMAL: nop
+; NORMAL: pushl $4
+; NORMAL-NEXT: pushl $3
+; NORMAL-NEXT: pushl $2
+; NORMAL-NEXT: pushl $1
+; NORMAL-NEXT: calll *16(%esp)
+; NORMAL-NEXT: addl $16, %esp
+define void @test10() optsize {
+  %stack_fptr = alloca void (i32, i32, i32, i32)*
+  store void (i32, i32, i32, i32)* @good, void (i32, i32, i32, i32)** %stack_fptr
+  %good_ptr = load volatile void (i32, i32, i32, i32)** %stack_fptr
+  call void asm sideeffect "nop", "~{ax},~{bx},~{cx},~{dx},~{bp},~{si},~{di}"()
+  call void (i32, i32, i32, i32)* %good_ptr(i32 1, i32 2, i32 3, i32 4)
+  ret void
+}
+
+; We can't fold the load from the global into the push because of 
+; interference from the store
+; NORMAL-LABEL: test11:
+; NORMAL: movl    _the_global, [[EAX:%e..]]
+; NORMAL-NEXT: movl    $42, _the_global
+; NORMAL-NEXT: pushl $4
+; NORMAL-NEXT: pushl $3
+; NORMAL-NEXT: pushl $2
+; NORMAL-NEXT: pushl [[EAX]]
+; NORMAL-NEXT: call
+; NORMAL-NEXT: addl $16, %esp
+@the_global = external global i32
+define void @test11() optsize {
+  %myload = load i32* @the_global
+  store i32 42, i32* @the_global
+  call void @good(i32 %myload, i32 2, i32 3, i32 4)
+  ret void
+}
 No newline at end of file