[ARM] Align stack objects passed to memory intrinsics

[john.brawn]

Summary

Memcpy, and other memory intrinsics, typically tries to use LDM/STM if the source and target addresses are 4-byte aligned. In CodeGenPrepare look for calls to memory intrinsics and, if the object is on the stack, 4-byte align it if it's large enough that we expect that memcpy would want to use LDM/STM to copy it.

[rengolin]

Hi John,

Have you done some stack consumption analysis? I remember Arnaud trying to reduce the size of the stack and this patch seems to go against it.

I agree it's a good idea to align it at 4, but we need to be careful, especially in -Os/z.

cheers,
--renato

[john.brawn]

Zero change in stack consumption in any EEMBC benchmark (I haven't checked SPEC). Also by CodeGen time we don't know if we're compiling -Os/-Oz as we just have CodeGenOpt::Level.

[hfinkel]

In D7908#130548, @john.brawn wrote:

Zero change in stack consumption in any EEMBC benchmark (I haven't checked SPEC). Also by CodeGen time we don't know if we're compiling -Os/-Oz as we just have CodeGenOpt::Level.

No, the size optimization attribute is independent of CodeGenOpt::Level. You can get at it from CodeGen by querying:

MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);

[hfinkel]

In D7908#130569, @hfinkel wrote:

In D7908#130548, @john.brawn wrote:

Zero change in stack consumption in any EEMBC benchmark (I haven't checked SPEC). Also by CodeGen time we don't know if we're compiling -Os/-Oz as we just have CodeGenOpt::Level.

No, the size optimization attribute is independent of CodeGenOpt::Level. You can get at it from CodeGen by querying:

MF.getFunction()->hasFnAttribute(Attribute::OptimizeForSize);

That having been said, code size and stack size, in this context, might be inversely related.

It might be better to implement this in CodeGenPrep. You could look directly for allocas that are used by memcpy intrinsics or as byval function argument parameters, and increase the alignment of only those directly.

[hfinkel]

Also worth noting, is that in getMemcpyLoadsAndStores in lib/CodeGen/SelectionDAG/SelectionDAG.cpp, there already is code that attempts to reset the alignment of a destination stack object when possible. Perhaps this should just be extended to do the same for the source.

[john.brawn]

In D7908#131187, @hfinkel wrote:

Also worth noting, is that in getMemcpyLoadsAndStores in lib/CodeGen/SelectionDAG/SelectionDAG.cpp, there already is code that attempts to reset the alignment of a destination stack object when possible. Perhaps this should just be extended to do the same for the source.

The goal is to align not just on direct calls to memcpy, but also calls to functions that then may call memcpy. It looks like doing it as a CodeGenPrep pass can do both - we first align arguments to a call, then check if the call is actually a memcpy and up its alignment if possible.

[john.brawn]

New patch attached that does this instead as a CodeGenPrepare pass.

[hfinkel]

Can you please put this into CodeGenPrep proper (with some target hook to enable it). I likely want to use this for PowerPC too, and the actual logic is fairly simple (so far).

lib/Target/ARM/ARMAlignAllocaPass.cpp

32 ↗

(On Diff #21012)

Line too long?

62 ↗

(On Diff #21012)

What are you actually trying to check for here? I assume you're mostly interested in looking for GEPs here (and not, for example, ptrtoint). Also, for GEPs, do you want to check that the offet to the underlying alloca is divisible by the alignment of interest. Maybe you should use something like Val->stripAndAccumulateInBoundsConstantOffsets() -- true, you'd miss the dynamic offset case, but I don't know if you have examples where that's important.

[john.brawn]

New patch attached that moves this into CodeGenPrepare, enabled by shouldAlignAllocaArgs in TargetLowering.

[rengolin]

lib/CodeGen/CodeGenPrepare.cpp

1236 ↗

(On Diff #21355)

I'm a bit averse of setting parameters for the computation inside a conditional and having to assert right after every call. I'd rather have (or reuse) some default size/alignment parameters from elsewhere.

[john.brawn]

lib/CodeGen/CodeGenPrepare.cpp

1236 ↗

(On Diff #21355)

If an implementation of shouldAlignAllocaArgs were to not set AllocaAlign then that indicates that it's not doing what it should, so failing an assert is better than hiding that by silently using a default value.

[rengolin]

lib/CodeGen/CodeGenPrepare.cpp

1236 ↗

(On Diff #21355)

I'm not against the assert, I'm against passing a reference for AllocaSize and AllocaAlign on a check call. The call is not explicit that it's setting the two values here, nor is it on TargetLowering's default implementation.

A more clear way would be to make a function whose only job is to set those variables and make it explicit on its name, like "setAlignAllocaArgs(CI, Size, Align)" and add "Size > 0" to the conditional. You can avoid the assert in that case, too.

Also, your check for TLI is redundant, since you already check for TD.

[john.brawn]

lib/CodeGen/CodeGenPrepare.cpp

1236 ↗

(On Diff #21355)

Doing it this way (return a bool, set a reference/pointer argument only if returning true) seems to be the standard way these queries are done in TargetInfo - see allowsMisalignedMemoryAccesses, getStackCookieLocation, GetAddrModeArguments.

[hfinkel]

include/llvm/Target/TargetLowering.h

980 ↗

(On Diff #21355)

Please name AllocaSize -> MinAllocaSize to make its purpose clearer.

lib/CodeGen/CodeGenPrepare.cpp

1235 ↗

(On Diff #21355)

Don't initialize these here; that makes it clear that they must be set by the TLI call, and allows mistakes to be caught by tools that detect use of undefined values (valgrind, etc.).

1236 ↗

(On Diff #21355)

I'm not against the assert, I'm against passing a reference for AllocaSize and AllocaAlign on a check call. The call is not explicit that it's setting the two values here, nor is it on TargetLowering's default implementation.

A more clear way would be to make a function whose only job is to set those variables and make it explicit on its name, like "setAlignAllocaArgs(CI, Size, Align)" and add "Size > 0" to the conditional. You can avoid the assert in that case, too.

I disagree. So long as the values are not initialized above the call, it is clear that the call must be setting them (and returning true if it has done so). This pattern is used by a number of other callbacks, even some in this file (getStackCookieLocation, getPreIndexedAddressParts, getPostIndexedAddressParts, ShrinkDemandedConstant, etc.), and I think it is appropriate here.

[rengolin]

ok

[john.brawn]

I may be reworking this patch to align not just allocas, but also global variables defined in this translation unit. The ultimate thing I'm trying to optimize is cases like

void fn() {
  char arr[31];
  strcpy(arr, "some string");
  // do somthing with arr
}

This patch aligns arr, but the variable generated for the string needs to be aligned as well. The plan was to do that in clang, but I'm thinking that maybe doing it here would be better.

Or maybe it would be better to get this checked in then do that, but the name shouldAlignAllocaArgs would have to be changed (maybe shouldAlignPointerArgs).

[john.brawn]

New patch that addresses review comments. It also renames things slightly in expectation that we'll be also aligning GlobalVariables. I'll be doing that in an upcoming patch.

[hfinkel]

You're currently doing this only for pointers that are captured as call arguments. Do you care about pointers captured in any other way (by having their address stored somewhere, or returned, for example)?

lib/CodeGen/CodeGenPrepare.cpp

1256 ↗

(On Diff #21488)

may have improved -> may be able to improve

[john.brawn]

In D7908#136838, @hfinkel wrote:

You're currently doing this only for pointers that are captured as call arguments. Do you care about pointers captured in any other way (by having their address stored somewhere, or returned, for example)?

Maybe? The original patch did that as it just aligned everything that was used in a FrameIndex instruction. I guess the question is, how much do we care if we over-align things pointlessly? When it's an argument to a memcpy call it's definitely a good idea, but in an argument to some other call or when the address is stored somewhere it may turn out to be pointless (which is possibly an argument that I should be restricting this to only memcpy/move etc. calls).

[hfinkel]

In D7908#136944, @john.brawn wrote:

In D7908#136838, @hfinkel wrote:

You're currently doing this only for pointers that are captured as call arguments. Do you care about pointers captured in any other way (by having their address stored somewhere, or returned, for example)?

Maybe? The original patch did that as it just aligned everything that was used in a FrameIndex instruction. I guess the question is, how much do we care if we over-align things pointlessly? When it's an argument to a memcpy call it's definitely a good idea, but in an argument to some other call or when the address is stored somewhere it may turn out to be pointless (which is possibly an argument that I should be restricting this to only memcpy/move etc. calls).

Agreed.

I don't have a strong opinion here. I suppose I'm inclined to go for other kinds of captures as well, or to stick with just memcpy/move/set. Picking only function calls, which amounts to some, but not all, captures, seems like an inconsistency likely to cause different kinds of abstractions to get different optimizations, and that's likely not good. Let's go to one extreme or the other ;)

[john.brawn]

New patch attached that goes for the approach of just aligning memory intrinsics.

[hfinkel]

lib/CodeGen/CodeGenPrepare.cpp

1252 ↗

(On Diff #21573)

If this is not an inbounds GEP, then this subtraction could underflow (it could overflow otherwise too, but without inbounds it could underflow even in a case where the code does not have undefined behavior). I recommend making sure the value being subtracted is not greater than the alloc size.

Please add a test case for this.

[john.brawn]

lib/CodeGen/CodeGenPrepare.cpp

1252 ↗

(On Diff #21573)

stripAndAccumulateInBoundsConstantOffsets will stop if it hits a non-inbound GEP, so we wouldn't get here in that case. Adding a test sound like a good idea though, I'll do that.

[hfinkel]

lib/CodeGen/CodeGenPrepare.cpp

1252 ↗

(On Diff #21573)

Yea, but even for an inbounds gep, we should do the right thing (the code may have undefined behavior, but having that reset the alignment of some objects seems a bit silly).

[john.brawn]

Attached patch handles size<offset by doing the comparison in a different way (by checking size >= minsize+offset), and adds a bunch of tests of offset handling.

[john.brawn]

While working on the next part of this (aligning global variables) I realised that we want to 8-byte align on some targets as it's faster, so the attached patch does that.

[hfinkel]

Looks good as far as I'm concerned. You might want to do the global variable handling as a separate patch.

llvm/trunk/include/llvm/Target/TargetLowering.h

virtual bool getStackCookieLocation(unsigned &/*AddressSpace*/,
     return false;
   }
 
   /// Returns true if a cast between SrcAS and DestAS is a noop.
   virtual bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const {
     return false;
   }
 
+  /// Return true if the pointer arguments to CI should be aligned by aligning
+  /// the object whose address is being passed. If so then MinSize is set to the
+  /// minimum size the object must be to be aligned and PrefAlign is set to the
+  /// preferred alignment.
+  virtual bool shouldAlignPointerArgs(CallInst */*CI*/, unsigned &/*MinSize*/,
+                                      unsigned &/*PrefAlign*/) const {
+    return false;
+  }
+
   //===--------------------------------------------------------------------===//
   /// \name Helpers for TargetTransformInfo implementations
   /// @{
 
   /// Get the ISD node that corresponds to the Instruction class opcode.
   int InstructionOpcodeToISD(unsigned Opcode) const;
 
   /// Estimate the cost of type-legalization and the legalized type.

llvm/trunk/lib/CodeGen/CodeGenPrepare.cpp

if (TLI->ExpandInlineAsm(CI)) {
       SunkAddrs.clear();
       return true;
     }
     // Sink address computing for memory operands into the block.
     if (OptimizeInlineAsmInst(CI))
       return true;
   }
 
+  const DataLayout *TD = TLI ? TLI->getDataLayout() : nullptr;
+
+  // Align the pointer arguments to this call if the target thinks it's a good
+  // idea
+  unsigned MinSize, PrefAlign;
+  if (TLI && TD && TLI->shouldAlignPointerArgs(CI, MinSize, PrefAlign)) {
+    for (auto &Arg : CI->arg_operands()) {
+      // We want to align both objects whose address is used directly and
+      // objects whose address is used in casts and GEPs, though it only makes
+      // sense for GEPs if the offset is a multiple of the desired alignment and
+      // if size - offset meets the size threshold.
+      if (!Arg->getType()->isPointerTy())
+        continue;
+      APInt Offset(TD->getPointerSizeInBits(
+                     cast<PointerType>(Arg->getType())->getAddressSpace()), 0);
+      Value *Val = Arg->stripAndAccumulateInBoundsConstantOffsets(*TD, Offset);
+      uint64_t Offset2 = Offset.getLimitedValue();
+      AllocaInst *AI;
+      if ((Offset2 & (PrefAlign-1)) == 0 &&
+          (AI = dyn_cast<AllocaInst>(Val)) &&
+          AI->getAlignment() < PrefAlign &&
+          TD->getTypeAllocSize(AI->getAllocatedType()) >= MinSize + Offset2)
+        AI->setAlignment(PrefAlign);
+      // TODO: Also align GlobalVariables
+    }
+    // If this is a memcpy (or similar) then we may be able to improve the
+    // alignment
+    if (MemIntrinsic *MI = dyn_cast<MemIntrinsic>(CI)) {
+      unsigned Align = getKnownAlignment(MI->getDest(), *TD);
+      if (MemTransferInst *MTI = dyn_cast<MemTransferInst>(MI))
+        Align = std::min(Align, getKnownAlignment(MTI->getSource(), *TD));
+      if (Align > MI->getAlignment())
+        MI->setAlignment(ConstantInt::get(MI->getAlignmentType(), Align));
+    }
+  }
+
   IntrinsicInst *II = dyn_cast<IntrinsicInst>(CI);
   if (II) {
     switch (II->getIntrinsicID()) {
     default: break;
     case Intrinsic::objectsize: {
       // Lower all uses of llvm.objectsize.*
       bool Min = (cast<ConstantInt>(II->getArgOperand(1))->getZExtValue() == 1);
       Type *ReturnTy = CI->getType();

llvm/trunk/lib/Target/ARM/ARMISelLowering.h

public:
     const TargetRegisterClass *getRegClassFor(MVT VT) const override;
 
     /// Returns true if a cast between SrcAS and DestAS is a noop.
     bool isNoopAddrSpaceCast(unsigned SrcAS, unsigned DestAS) const override {
       // Addrspacecasts are always noops.
       return true;
     }
 
+    bool shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
+                                unsigned &PrefAlign) const override;
+
     /// createFastISel - This method returns a target specific FastISel object,
     /// or null if the target does not support "fast" ISel.
     FastISel *createFastISel(FunctionLoweringInfo &funcInfo,
                              const TargetLibraryInfo *libInfo) const override;
 
     Sched::Preference getSchedulingPreference(SDNode *N) const override;
 
     bool

llvm/trunk/lib/Target/ARM/ARMISelLowering.cpp

 #include "llvm/IR/CallingConv.h"
 #include "llvm/IR/Constants.h"
 #include "llvm/IR/Function.h"
 #include "llvm/IR/GlobalValue.h"
 #include "llvm/IR/IRBuilder.h"
 #include "llvm/IR/Instruction.h"
 #include "llvm/IR/Instructions.h"
 #include "llvm/IR/Intrinsics.h"
+#include "llvm/IR/IntrinsicInst.h"
 #include "llvm/IR/Type.h"
 #include "llvm/MC/MCSectionMachO.h"
 #include "llvm/Support/CommandLine.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/ErrorHandling.h"
 #include "llvm/Support/MathExtras.h"
 #include "llvm/Target/TargetOptions.h"
 #include <utility>
if (Subtarget->hasNEON()) {
     if (VT == MVT::v4i64)
       return &ARM::QQPRRegClass;
     if (VT == MVT::v8i64)
       return &ARM::QQQQPRRegClass;
   }
   return TargetLowering::getRegClassFor(VT);
 }
 
+// memcpy, and other memory intrinsics, typically tries to use LDM/STM if the
+// source/dest is aligned and the copy size is large enough. We therefore want
+// to align such objects passed to memory intrinsics.
+bool ARMTargetLowering::shouldAlignPointerArgs(CallInst *CI, unsigned &MinSize,
+                                               unsigned &PrefAlign) const {
+  if (!isa<MemIntrinsic>(CI))
+    return false;
+  MinSize = 8;
+  // On ARM11 onwards (excluding M class) 8-byte aligned LDM is typically 1
+  // cycle faster than 4-byte aligned LDM.
+  PrefAlign = (Subtarget->hasV6Ops() && !Subtarget->isMClass() ? 8 : 4);
+  return true;
+}
+
 // Create a fast isel object.
 FastISel *
 ARMTargetLowering::createFastISel(FunctionLoweringInfo &funcInfo,
                                   const TargetLibraryInfo *libInfo) const {
   return ARM::createFastISel(funcInfo, libInfo);
 }
 
 Sched::Preference ARMTargetLowering::getSchedulingPreference(SDNode *N) const {

llvm/trunk/test/CodeGen/ARM/memfunc.ll

-; RUN: llc < %s -mtriple=armv7-apple-ios -o - | FileCheck %s
-; RUN: llc < %s -mtriple=thumbv7m-none-macho -o - | FileCheck %s --check-prefix=DARWIN
-; RUN: llc < %s -mtriple=arm-none-eabi -o - | FileCheck --check-prefix=EABI %s
-; RUN: llc < %s -mtriple=arm-none-eabihf -o - | FileCheck --check-prefix=EABI %s
+; RUN: llc < %s -mtriple=armv7-apple-ios -disable-post-ra -o - | FileCheck %s --check-prefix=CHECK-IOS --check-prefix=CHECK
+; RUN: llc < %s -mtriple=thumbv7m-none-macho -disable-post-ra -o - | FileCheck %s --check-prefix=CHECK-DARWIN --check-prefix=CHECK
+; RUN: llc < %s -mtriple=arm-none-eabi -disable-post-ra -o - | FileCheck %s --check-prefix=CHECK-EABI --check-prefix=CHECK
+; RUN: llc < %s -mtriple=arm-none-eabihf -disable-post-ra -o - | FileCheck %s --check-prefix=CHECK-EABI --check-prefix=CHECK
 
 @from = common global [500 x i32] zeroinitializer, align 4
 @to = common global [500 x i32] zeroinitializer, align 4
 
-define void @f() {
+define void @f1() {
 entry:
+  ; CHECK-LABEL: f1
 
-        ; CHECK: memmove
-        ; EABI: __aeabi_memmove
-        call void @llvm.memmove.p0i8.p0i8.i32(i8* bitcast ([500 x i32]* @from to i8*), i8* bitcast ([500 x i32]* @to to i8*), i32 500, i32 0, i1 false)
+  ; CHECK-IOS: memmove
+  ; CHECK-DARWIN: memmove
+  ; CHECK-EABI: __aeabi_memmove
 
-        ; CHECK: memcpy
-        ; EABI: __aeabi_memcpy
-        call void @llvm.memcpy.p0i8.p0i8.i32(i8* bitcast ([500 x i32]* @from to i8*), i8* bitcast ([500 x i32]* @to to i8*), i32 500, i32 0, i1 false)
+  ; CHECK-IOS: memcpy
+  ; CHECK-DARWIN: memcpy
+  ; CHECK-EABI: __aeabi_memcpy
 
-        ; EABI memset swaps arguments
-        ; CHECK: mov r1, #0
-        ; CHECK: memset
-        ; DARWIN: movs r1, #0
-        ; DARWIN: memset
-        ; EABI: mov r2, #0
-        ; EABI: __aeabi_memset
-        call void @llvm.memset.p0i8.i32(i8* bitcast ([500 x i32]* @from to i8*), i8 0, i32 500, i32 0, i1 false)
-        unreachable
+  ; CHECK-IOS: mov r1, #0
+  ; CHECK-IOS: memset
+  ; CHECK-DARWIN: movs r1, #0
+  ; CHECK-DARWIN: memset
+  ; CHECK-EABI: mov r2, #0
+  ; CHECK-EABI: __aeabi_memset
 }
 
+; Check that alloca arguments to memory intrinsics are automatically aligned if at least 8 bytes in size
+define void @f2(i8* %dest, i32 %n) {
+entry:
+  ; CHECK-LABEL: f2
+
+  ; IOS (ARMv7) should 8-byte align, others should 4-byte align
+  ; CHECK-IOS: add r1, sp, #32
+  ; CHECK-IOS: memmove
+  ; CHECK-DARWIN: add r1, sp, #28
+  ; CHECK-DARWIN: memmove
+  ; CHECK-EABI: add r1, sp, #28
+  ; CHECK-EABI: __aeabi_memmove
+  %arr0 = alloca [9 x i8], align 1
+  %0 = bitcast [9 x i8]* %arr0 to i8*
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dest, i8* %0, i32 %n, i32 0, i1 false)
+
+  ; CHECK: add r1, sp, #16
+  ; CHECK-IOS: memcpy
+  ; CHECK-DARWIN: memcpy
+  ; CHECK-EABI: __aeabi_memcpy
+  %arr1 = alloca [9 x i8], align 1
+  %1 = bitcast [9 x i8]* %arr1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %dest, i8* %1, i32 %n, i32 0, i1 false)
+
+  ; CHECK-IOS: mov r0, sp
+  ; CHECK-IOS: mov r1, #0
+  ; CHECK-IOS: memset
+  ; CHECK-DARINW: add r0, sp, #4
+  ; CHECK-DARWIN: movs r1, #0
+  ; CHECK-DARWIN: memset
+  ; CHECK-EABI: add r0, sp, #4
+  ; CHECK-EABI: mov r2, #0
+  ; CHECK-EABI: __aeabi_memset
+  %arr2 = alloca [9 x i8], align 1
+  %2 = bitcast [9 x i8]* %arr2 to i8*
+  call void @llvm.memset.p0i8.i32(i8* %2, i8 0, i32 %n, i32 0, i1 false)
+
+  unreachable
+}
+
+; Check that alloca arguments are not aligned if less than 8 bytes in size
+define void @f3(i8* %dest, i32 %n) {
+entry:
+  ; CHECK-LABEL: f3
+
+  ; CHECK: {{add(.w)? r1, sp, #17|sub(.w)? r1, r7, #15}}
+  ; CHECK-IOS: memmove
+  ; CHECK-DARWIN: memmove
+  ; CHECK-EABI: __aeabi_memmove
+  %arr0 = alloca [7 x i8], align 1
+  %0 = bitcast [7 x i8]* %arr0 to i8*
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dest, i8* %0, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r1, sp, #10}}
+  ; CHECK-IOS: memcpy
+  ; CHECK-DARWIN: memcpy
+  ; CHECK-EABI: __aeabi_memcpy
+  %arr1 = alloca [7 x i8], align 1
+  %1 = bitcast [7 x i8]* %arr1 to i8*
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %dest, i8* %1, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r0, sp, #3}}
+  ; CHECK-IOS: mov r1, #0
+  ; CHECK-IOS: memset
+  ; CHECK-DARWIN: movs r1, #0
+  ; CHECK-DARWIN: memset
+  ; CHECK-EABI: mov r2, #0
+  ; CHECK-EABI: __aeabi_memset
+  %arr2 = alloca [7 x i8], align 1
+  %2 = bitcast [7 x i8]* %arr2 to i8*
+  call void @llvm.memset.p0i8.i32(i8* %2, i8 0, i32 %n, i32 0, i1 false)
+
+  unreachable
+}
+
+; Check that alloca arguments are not aligned if size+offset is less than 8 bytes
+define void @f4(i8* %dest, i32 %n) {
+entry:
+  ; CHECK-LABEL: f4
+
+  ; CHECK: {{add(.w)? r., sp, #23|sub(.w)? r., r7, #17}}
+  ; CHECK-IOS: memmove
+  ; CHECK-DARWIN: memmove
+  ; CHECK-EABI: __aeabi_memmove
+  %arr0 = alloca [9 x i8], align 1
+  %0 = getelementptr inbounds [9 x i8], [9 x i8]* %arr0, i32 0, i32 4
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dest, i8* %0, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r., sp, #(10|14)}}
+  ; CHECK-IOS: memcpy
+  ; CHECK-DARWIN: memcpy
+  ; CHECK-EABI: __aeabi_memcpy
+  %arr1 = alloca [9 x i8], align 1
+  %1 = getelementptr inbounds [9 x i8], [9 x i8]* %arr1, i32 0, i32 4
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %dest, i8* %1, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r., sp, #(1|5)}}
+  ; CHECK-IOS: mov r1, #0
+  ; CHECK-IOS: memset
+  ; CHECK-DARWIN: movs r1, #0
+  ; CHECK-DARWIN: memset
+  ; CHECK-EABI: mov r2, #0
+  ; CHECK-EABI: __aeabi_memset
+  %arr2 = alloca [9 x i8], align 1
+  %2 = getelementptr inbounds [9 x i8], [9 x i8]* %arr2, i32 0, i32 4
+  call void @llvm.memset.p0i8.i32(i8* %2, i8 0, i32 %n, i32 0, i1 false)
+
+  unreachable
+}
+
+; Check that alloca arguments are not aligned if the offset is not a multiple of 4
+define void @f5(i8* %dest, i32 %n) {
+entry:
+  ; CHECK-LABEL: f5
+
+  ; CHECK: {{add(.w)? r., sp, #27|sub(.w)? r., r7, #21}}
+  ; CHECK-IOS: memmove
+  ; CHECK-DARWIN: memmove
+  ; CHECK-EABI: __aeabi_memmove
+  %arr0 = alloca [13 x i8], align 1
+  %0 = getelementptr inbounds [13 x i8], [13 x i8]* %arr0, i32 0, i32 1
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dest, i8* %0, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r., sp, #(10|14)}}
+  ; CHECK-IOS: memcpy
+  ; CHECK-DARWIN: memcpy
+  ; CHECK-EABI: __aeabi_memcpy
+  %arr1 = alloca [13 x i8], align 1
+  %1 = getelementptr inbounds [13 x i8], [13 x i8]* %arr1, i32 0, i32 1
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %dest, i8* %1, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r., sp, #(1|5)}}
+  ; CHECK-IOS: mov r1, #0
+  ; CHECK-IOS: memset
+  ; CHECK-DARWIN: movs r1, #0
+  ; CHECK-DARWIN: memset
+  ; CHECK-EABI: mov r2, #0
+  ; CHECK-EABI: __aeabi_memset
+  %arr2 = alloca [13 x i8], align 1
+  %2 = getelementptr inbounds [13 x i8], [13 x i8]* %arr2, i32 0, i32 1
+  call void @llvm.memset.p0i8.i32(i8* %2, i8 0, i32 %n, i32 0, i1 false)
+
+  unreachable
+}
+
+; Check that alloca arguments are not aligned if the offset is unknown
+define void @f6(i8* %dest, i32 %n, i32 %i) {
+entry:
+  ; CHECK-LABEL: f6
+
+  ; CHECK: {{add(.w)? r., sp, #27|sub(.w)? r., r7, #25}}
+  ; CHECK-IOS: memmove
+  ; CHECK-DARWIN: memmove
+  ; CHECK-EABI: __aeabi_memmove
+  %arr0 = alloca [13 x i8], align 1
+  %0 = getelementptr inbounds [13 x i8], [13 x i8]* %arr0, i32 0, i32 %i
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dest, i8* %0, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r., sp, #(10|14)}}
+  ; CHECK-IOS: memcpy
+  ; CHECK-DARWIN: memcpy
+  ; CHECK-EABI: __aeabi_memcpy
+  %arr1 = alloca [13 x i8], align 1
+  %1 = getelementptr inbounds [13 x i8], [13 x i8]* %arr1, i32 0, i32 %i
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %dest, i8* %1, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r., sp, #(1|5)}}
+  ; CHECK-IOS: mov r1, #0
+  ; CHECK-IOS: memset
+  ; CHECK-DARWIN: movs r1, #0
+  ; CHECK-DARWIN: memset
+  ; CHECK-EABI: mov r2, #0
+  ; CHECK-EABI: __aeabi_memset
+  %arr2 = alloca [13 x i8], align 1
+  %2 = getelementptr inbounds [13 x i8], [13 x i8]* %arr2, i32 0, i32 %i
+  call void @llvm.memset.p0i8.i32(i8* %2, i8 0, i32 %n, i32 0, i1 false)
+
+  unreachable
+}
+
+; Check that alloca arguments are not aligned if the GEP is not inbounds
+define void @f7(i8* %dest, i32 %n) {
+entry:
+  ; CHECK-LABEL: f7
+
+  ; CHECK: {{add(.w)? r., sp, #27|sub(.w)? r., r7, #21}}
+  ; CHECK-IOS: memmove
+  ; CHECK-DARWIN: memmove
+  ; CHECK-EABI: __aeabi_memmove
+  %arr0 = alloca [13 x i8], align 1
+  %0 = getelementptr [13 x i8], [13 x i8]* %arr0, i32 0, i32 4
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dest, i8* %0, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r., sp, #(10|14)}}
+  ; CHECK-IOS: memcpy
+  ; CHECK-DARWIN: memcpy
+  ; CHECK-EABI: __aeabi_memcpy
+  %arr1 = alloca [13 x i8], align 1
+  %1 = getelementptr [13 x i8], [13 x i8]* %arr1, i32 0, i32 4
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %dest, i8* %1, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r., sp, #(1|5)}}
+  ; CHECK-IOS: mov r1, #0
+  ; CHECK-IOS: memset
+  ; CHECK-DARWIN: movs r1, #0
+  ; CHECK-DARWIN: memset
+  ; CHECK-EABI: mov r2, #0
+  ; CHECK-EABI: __aeabi_memset
+  %arr2 = alloca [13 x i8], align 1
+  %2 = getelementptr [13 x i8], [13 x i8]* %arr2, i32 0, i32 4
+  call void @llvm.memset.p0i8.i32(i8* %2, i8 0, i32 %n, i32 0, i1 false)
+
+  unreachable
+}
+
+; Check that alloca arguments are not aligned when the offset is past the end of the allocation
+define void @f8(i8* %dest, i32 %n) {
+entry:
+  ; CHECK-LABEL: f8
+
+  ; CHECK: {{add(.w)? r., sp, #27|sub(.w)? r., r7, #21}}
+  ; CHECK-IOS: memmove
+  ; CHECK-DARWIN: memmove
+  ; CHECK-EABI: __aeabi_memmove
+  %arr0 = alloca [13 x i8], align 1
+  %0 = getelementptr inbounds [13 x i8], [13 x i8]* %arr0, i32 0, i32 16
+  call void @llvm.memmove.p0i8.p0i8.i32(i8* %dest, i8* %0, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r., sp, #(10|14)}}
+  ; CHECK-IOS: memcpy
+  ; CHECK-DARWIN: memcpy
+  ; CHECK-EABI: __aeabi_memcpy
+  %arr1 = alloca [13 x i8], align 1
+  %1 = getelementptr inbounds [13 x i8], [13 x i8]* %arr1, i32 0, i32 16
+  call void @llvm.memcpy.p0i8.p0i8.i32(i8* %dest, i8* %1, i32 %n, i32 0, i1 false)
+
+  ; CHECK: {{add(.w)? r., sp, #(1|5)}}
+  ; CHECK-IOS: mov r1, #0
+  ; CHECK-IOS: memset
+  ; CHECK-DARWIN: movs r1, #0
+  ; CHECK-DARWIN: memset
+  ; CHECK-EABI: mov r2, #0
+  ; CHECK-EABI: __aeabi_memset
+  %arr2 = alloca [13 x i8], align 1
+  %2 = getelementptr inbounds [13 x i8], [13 x i8]* %arr2, i32 0, i32 16
+  call void @llvm.memset.p0i8.i32(i8* %2, i8 0, i32 %n, i32 0, i1 false)
+
+  unreachable
+}
+
 declare void @llvm.memmove.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1) nounwind
 declare void @llvm.memcpy.p0i8.p0i8.i32(i8* nocapture, i8* nocapture, i32, i32, i1) nounwind
 declare void @llvm.memset.p0i8.i32(i8* nocapture, i8, i32, i32, i1) nounwind

llvm/trunk/test/CodeGen/ARM/memset-inline.ll

 }
 
 define void @t2() nounwind ssp {
 entry:
 ; CHECK-LABEL: t2:
 ; CHECK: add.w r1, r0, #10
 ; CHECK: vmov.i32 {{q[0-9]+}}, #0x0
 ; CHECK: vst1.16 {d{{[0-9]+}}, d{{[0-9]+}}}, [r1]
-; CHECK: vst1.32 {d{{[0-9]+}}, d{{[0-9]+}}}, [r0]
+; CHECK: vst1.64 {d{{[0-9]+}}, d{{[0-9]+}}}, [r0]
   %buf = alloca [26 x i8], align 1
   %0 = getelementptr inbounds [26 x i8], [26 x i8]* %buf, i32 0, i32 0
   call void @llvm.memset.p0i8.i32(i8* %0, i8 0, i32 26, i32 1, i1 false)
   call void @something(i8* %0) nounwind
   ret void
 }
 
 declare void @something(i8*) nounwind
 declare void @llvm.memset.p0i8.i32(i8* nocapture, i8, i32, i32, i1) nounwind
 declare void @llvm.memset.p0i8.i64(i8* nocapture, i8, i64, i32, i1) nounwind