[PPC]: Peephole optimize small accesss to aligned globals.

[iteratee]

Summary

[PPC]: Peephole optimize small accesss to aligned globals.

Access to aligned globals gives us a chance to peephole optimize nonzero
offsets. If a struct is 4 byte aligned, then accesses to bytes 0-3 won't
overflow the available displacement. For example:

addis 3, 2, b4v@toc@ha
addi 4, 3, b4v@toc@l
lbz 5, b4v@toc@l(3) ; This is the result of the current peephole
lbz 6, 1(4)         ; optimizer
lbz 7, 2(4)
lbz 8, 3(4)

If b4v is 4-byte aligned, we can skip using register 4 because we know
that b4v@toc@l+{1,2,3} won't overflow 32K, and instead generate:

addis 3, 2, b4v@toc@ha
lbz 4, b4v@toc@l(3)
lbz 5, b4v@toc@l+1(3)
lbz 6, b4v@toc@l+2(3)
lbz 7, b4v@toc@l+3(3)

Saving a register and an addition.
Larger alignments allow larger structures/arrays to be optimized.
This only applies on power8 if there is a single use resulting in a new fusion opportunity, or when optimizing for size.

[echristo]

One small inline comment.

lib/Target/PowerPC/PPCISelDAGToDAG.cpp

4167–4181

Not sure I understand the moves here?

[iteratee]

lib/Target/PowerPC/PPCISelDAGToDAG.cpp

4167–4181

They got sorted according to size when I thought I needed to know the size.
Reverted.

[iteratee]

Ping? Any objections?

[hfinkel]

lib/Target/PowerPC/PPCISelDAGToDAG.cpp

4143

But we should do this when optimizing for code size, even on the P8:

if (PPCSubTarget->hasFusion() && !MF->getFunction()->optForSize())

or, if this really *hurts* performance on the P8, use optForMinSize().

Also, we don't need to turn this off on the P8 when there is only a single (non-debug) user because, as the ELF v2 ABI spec points out:

addis r4, r3, upper
<lbz,lhz,lwz,ld> r4, lower(r4)

is also good.

[iteratee]

Run the optimization on fusion platforms if it results in a new fusion opportunity, or optimizing for size.

[iteratee]

I implemented your suggestion and fired this selectively for power8

lib/Target/PowerPC/PPCISelDAGToDAG.cpp

4143

Nice catch. I've implemented that.

[echristo]

lib/Target/PowerPC/PPCISelDAGToDAG.cpp

4197–4199

Looks weird. Feel like collapsing it to a single if conditional?

[iteratee]

Collapse nested condition

[iteratee]

Adjust formatting.

[hfinkel]

LGTM

lib/Target/PowerPC/PPCISelDAGToDAG.cpp

 void PPCDAGToDAGISel::PeepholePPC64() {
   // These optimizations are currently supported only for 64-bit SVR4.
   if (PPCSubTarget->isDarwin() || !PPCSubTarget->isPPC64())
     return;
 
   SelectionDAG::allnodes_iterator Position(CurDAG->getRoot().getNode());
   ++Position;
 
   while (Position != CurDAG->allnodes_begin()) {
     SDNode *N = &*--Position;
     // Skip dead nodes and any non-machine opcodes.
     if (N->use_empty() || !N->isMachineOpcode())
       continue;
 
     unsigned FirstOp;
     unsigned StorageOpcode = N->getMachineOpcode();
 
     switch (StorageOpcode) {
     default: continue;
 
     case PPC::LBZ:
     case PPC::LBZ8:
     case PPC::LD:
     case PPC::LFD:
     case PPC::LFS:
     case PPC::LHA:
     case PPC::LHA8:
     case PPC::LHZ:
     case PPC::LHZ8:
     case PPC::LWA:
     case PPC::LWZ:
     case PPC::LWZ8:
       FirstOp = 0;
       break;
 
     case PPC::STB:
     case PPC::STB8:
     case PPC::STD:
     case PPC::STFD:
     case PPC::STFS:
     case PPC::STH:
     case PPC::STH8:
     case PPC::STW:
     case PPC::STW8:
       FirstOp = 1;
       break;
     }
 
-    // If this is a load or store with a zero offset, we may be able to
-    // fold an add-immediate into the memory operation.
-    if (!isa<ConstantSDNode>(N->getOperand(FirstOp)) ||
-        N->getConstantOperandVal(FirstOp) != 0)
+    // If this is a load or store with a zero offset, or within the alignment,
+    // we may be able to fold an add-immediate into the memory operation.
+    // The check against alignment is below, as it can't occur until we check
+    // the arguments to N
+    if (!isa<ConstantSDNode>(N->getOperand(FirstOp)))
       continue;
 
     SDValue Base = N->getOperand(FirstOp + 1);
     if (!Base.isMachineOpcode())
       continue;
 
+    // On targets with fusion, we don't want this to fire and remove a fusion
+    // opportunity, unless a) it results in another fusion opportunity or
+    // b) optimizing for size.
+    if (PPCSubTarget->hasFusion() &&
+        (!MF->getFunction()->optForSize() && !Base.hasOneUse())
+      continue;
+
     unsigned Flags = 0;
     bool ReplaceFlags = true;
 
     // When the feeding operation is an add-immediate of some sort,
     // determine whether we need to add relocation information to the
     // target flags on the immediate operand when we fold it into the
     // load instruction.
     //
while (Position != CurDAG->allnodes_begin()) {
     case PPC::ADDItlsldL:
       Flags = PPCII::MO_TLSLD_LO;
       break;
     case PPC::ADDItocL:
       Flags = PPCII::MO_TOC_LO;
       break;
     }
 
+    SDValue ImmOpnd = Base.getOperand(1);
+    int MaxDisplacement = 0;
+    if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) {
+      const GlobalValue *GV = GA->getGlobal();
+      MaxDisplacement = GV->getAlignment() - 1;
+    }
+
+    int Offset = N->getConstantOperandVal(FirstOp);
+    if (Offset < 0 || Offset > MaxDisplacement)
+      continue;
+
     // We found an opportunity.  Reverse the operands from the add
     // immediate and substitute them into the load or store.  If
     // needed, update the target flags for the immediate operand to
     // reflect the necessary relocation information.
     DEBUG(dbgs() << "Folding add-immediate into mem-op:\nBase:    ");
     DEBUG(Base->dump(CurDAG));
     DEBUG(dbgs() << "\nN: ");
     DEBUG(N->dump(CurDAG));
     DEBUG(dbgs() << "\n");
 
-    SDValue ImmOpnd = Base.getOperand(1);
-
     // If the relocation information isn't already present on the
     // immediate operand, add it now.
     if (ReplaceFlags) {
       if (GlobalAddressSDNode *GA = dyn_cast<GlobalAddressSDNode>(ImmOpnd)) {
         SDLoc dl(GA);
         const GlobalValue *GV = GA->getGlobal();
         // We can't perform this optimization for data whose alignment
         // is insufficient for the instruction encoding.
         if (GV->getAlignment() < 4 &&
             (StorageOpcode == PPC::LD || StorageOpcode == PPC::STD ||
-             StorageOpcode == PPC::LWA)) {
+             StorageOpcode == PPC::LWA || (Offset % 4) != 0)) {
           DEBUG(dbgs() << "Rejected this candidate for alignment.\n\n");
           continue;
         }
-        ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, 0, Flags);
+        ImmOpnd = CurDAG->getTargetGlobalAddress(GV, dl, MVT::i64, Offset, Flags);
       } else if (ConstantPoolSDNode *CP =
                  dyn_cast<ConstantPoolSDNode>(ImmOpnd)) {
         const Constant *C = CP->getConstVal();
         ImmOpnd = CurDAG->getTargetConstantPool(C, MVT::i64,
                                                 CP->getAlignment(),
-                                                0, Flags);
+                                                Offset, Flags);
       }
     }
 
     if (FirstOp == 1) // Store
       (void)CurDAG->UpdateNodeOperands(N, N->getOperand(0), ImmOpnd,
                                        Base.getOperand(0), N->getOperand(3));
     else // Load
       (void)CurDAG->UpdateNodeOperands(N, ImmOpnd, Base.getOperand(0),

test/CodeGen/PowerPC/peephole-align.ll

+; RUN: llc -mcpu=pwr7 -O1 -code-model=medium <%s | FileCheck -check-prefix=POWER7 -check-prefix=CHECK %s
+; RUN: llc -mcpu=pwr8 -O1 -code-model=medium <%s | FileCheck -check-prefix=POWER8 -check-prefix=CHECK %s
+
+; Test peephole optimization for medium code model (32-bit TOC offsets)
+; for loading and storing small offsets within aligned values.
+; For power8, verify that the optimization doesn't fire, as it prevents fusion
+; opportunities.
+
+target datalayout = "E-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-f128:128:128-v128:128:128-n32:64"
+target triple = "powerpc64-unknown-linux-gnu"
+
+%struct.b4 = type<{ i8, i8, i8, i8 }>
+%struct.h2 = type<{ i16, i16 }>
+
+%struct.b8 = type<{ i8, i8, i8, i8, i8, i8, i8, i8 }>
+%struct.h4 = type<{ i16, i16, i16, i16 }>
+%struct.w2 = type<{ i32, i32 }>
+
+%struct.d2 = type<{ i64, i64 }>
+%struct.misalign = type<{ i8, i64 }>
+
+@b4v = global %struct.b4 <{ i8 1, i8 2, i8 3, i8 4 }>, align 4
+@h2v = global %struct.h2 <{ i16 1, i16 2 }>, align 4
+
+@b8v = global %struct.b8 <{ i8 1, i8 2, i8 3, i8 4, i8 5, i8 6, i8 7, i8 8 }>, align 8
+@h4v = global %struct.h4 <{ i16 1, i16 2, i16 3, i16 4 }>, align 8
+@w2v = global %struct.w2 <{ i32 1, i32 2 }>, align 8
+
+@d2v = global %struct.d2 <{ i64 1, i64 2 }>, align 16
+@misalign_v = global %struct.misalign <{ i8 1, i64 2 }>, align 16
+
+; CHECK-LABEL: test_b4:
+; POWER7: addis [[REGSTRUCT:[0-9]+]], 2, b4v@toc@ha
+; POWER7-DAG: lbz [[REG0_0:[0-9]+]], b4v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: lbz [[REG1_0:[0-9]+]], b4v@toc@l+1([[REGSTRUCT]])
+; POWER7-DAG: lbz [[REG2_0:[0-9]+]], b4v@toc@l+2([[REGSTRUCT]])
+; POWER7-DAG: lbz [[REG3_0:[0-9]+]], b4v@toc@l+3([[REGSTRUCT]])
+; POWER7-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER7-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER7-DAG: addi [[REG2_1:[0-9]+]], [[REG2_0]], 3
+; POWER7-DAG: addi [[REG3_1:[0-9]+]], [[REG3_0]], 4
+; POWER7-DAG: stb [[REG0_1]], b4v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: stb [[REG1_1]], b4v@toc@l+1([[REGSTRUCT]])
+; POWER7-DAG: stb [[REG2_1]], b4v@toc@l+2([[REGSTRUCT]])
+; POWER7-DAG: stb [[REG3_1]], b4v@toc@l+3([[REGSTRUCT]])
+
+; POWER8: addis [[REGSTRUCT:[0-9]+]], 2, b4v@toc@ha
+; POWER8-NEXT: addi [[REGSTRUCT]], [[REGSTRUCT]], b4v@toc@l
+; POWER8-DAG: lbz [[REG0_0:[0-9]+]], 0([[REGSTRUCT]])
+; POWER8-DAG: lbz [[REG1_0:[0-9]+]], 1([[REGSTRUCT]])
+; POWER8-DAG: lbz [[REG2_0:[0-9]+]], 2([[REGSTRUCT]])
+; POWER8-DAG: lbz [[REG3_0:[0-9]+]], 3([[REGSTRUCT]])
+; POWER8-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER8-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER8-DAG: addi [[REG2_1:[0-9]+]], [[REG2_0]], 3
+; POWER8-DAG: addi [[REG3_1:[0-9]+]], [[REG3_0]], 4
+; POWER8-DAG: stb [[REG0_1]], 0([[REGSTRUCT]])
+; POWER8-DAG: stb [[REG1_1]], 1([[REGSTRUCT]])
+; POWER8-DAG: stb [[REG2_1]], 2([[REGSTRUCT]])
+; POWER8-DAG: stb [[REG3_1]], 3([[REGSTRUCT]])
+define void @test_b4() nounwind {
+entry:
+  %0 = load i8, i8* getelementptr inbounds (%struct.b4, %struct.b4* @b4v, i32 0, i32 0), align 1
+  %inc0 = add nsw i8 %0, 1
+  store i8 %inc0, i8* getelementptr inbounds (%struct.b4, %struct.b4* @b4v, i32 0, i32 0), align 1
+  %1 = load i8, i8* getelementptr inbounds (%struct.b4, %struct.b4* @b4v, i32 0, i32 1), align 1
+  %inc1 = add nsw i8 %1, 2
+  store i8 %inc1, i8* getelementptr inbounds (%struct.b4, %struct.b4* @b4v, i32 0, i32 1), align 1
+  %2 = load i8, i8* getelementptr inbounds (%struct.b4, %struct.b4* @b4v, i32 0, i32 2), align 1
+  %inc2 = add nsw i8 %2, 3
+  store i8 %inc2, i8* getelementptr inbounds (%struct.b4, %struct.b4* @b4v, i32 0, i32 2), align 1
+  %3 = load i8, i8* getelementptr inbounds (%struct.b4, %struct.b4* @b4v, i32 0, i32 3), align 1
+  %inc3 = add nsw i8 %3, 4
+  store i8 %inc3, i8* getelementptr inbounds (%struct.b4, %struct.b4* @b4v, i32 0, i32 3), align 1
+  ret void
+}
+
+; CHECK-LABEL: test_h2:
+; POWER7: addis [[REGSTRUCT:[0-9]+]], 2, h2v@toc@ha
+; POWER7-DAG: lhz [[REG0_0:[0-9]+]], h2v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: lhz [[REG1_0:[0-9]+]], h2v@toc@l+2([[REGSTRUCT]])
+; POWER7-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER7-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER7-DAG: sth [[REG0_1]], h2v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: sth [[REG1_1]], h2v@toc@l+2([[REGSTRUCT]])
+
+; POWER8: addis [[REGSTRUCT:[0-9]+]], 2, h2v@toc@ha
+; POWER8-NEXT: addi [[REGSTRUCT]], [[REGSTRUCT]], h2v@toc@l
+; POWER8-DAG: lhz [[REG0_0:[0-9]+]], 0([[REGSTRUCT]])
+; POWER8-DAG: lhz [[REG1_0:[0-9]+]], 2([[REGSTRUCT]])
+; POWER8-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER8-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER8-DAG: sth [[REG0_1]], 0([[REGSTRUCT]])
+; POWER8-DAG: sth [[REG1_1]], 2([[REGSTRUCT]])
+define void @test_h2() nounwind {
+entry:
+  %0 = load i16, i16* getelementptr inbounds (%struct.h2, %struct.h2* @h2v, i32 0, i32 0), align 2
+  %inc0 = add nsw i16 %0, 1
+  store i16 %inc0, i16* getelementptr inbounds (%struct.h2, %struct.h2* @h2v, i32 0, i32 0), align 2
+  %1 = load i16, i16* getelementptr inbounds (%struct.h2, %struct.h2* @h2v, i32 0, i32 1), align 2
+  %inc1 = add nsw i16 %1, 2
+  store i16 %inc1, i16* getelementptr inbounds (%struct.h2, %struct.h2* @h2v, i32 0, i32 1), align 2
+  ret void
+}
+
+; CHECK-LABEL: test_h2_optsize:
+; CHECK: addis [[REGSTRUCT:[0-9]+]], 2, h2v@toc@ha
+; CHECK-DAG: lhz [[REG0_0:[0-9]+]], h2v@toc@l([[REGSTRUCT]])
+; CHECK-DAG: lhz [[REG1_0:[0-9]+]], h2v@toc@l+2([[REGSTRUCT]])
+; CHECK-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; CHECK-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; CHECK-DAG: sth [[REG0_1]], h2v@toc@l([[REGSTRUCT]])
+; CHECK-DAG: sth [[REG1_1]], h2v@toc@l+2([[REGSTRUCT]])
+define void @test_h2_optsize() optsize nounwind {
+entry:
+  %0 = load i16, i16* getelementptr inbounds (%struct.h2, %struct.h2* @h2v, i32 0, i32 0), align 2
+  %inc0 = add nsw i16 %0, 1
+  store i16 %inc0, i16* getelementptr inbounds (%struct.h2, %struct.h2* @h2v, i32 0, i32 0), align 2
+  %1 = load i16, i16* getelementptr inbounds (%struct.h2, %struct.h2* @h2v, i32 0, i32 1), align 2
+  %inc1 = add nsw i16 %1, 2
+  store i16 %inc1, i16* getelementptr inbounds (%struct.h2, %struct.h2* @h2v, i32 0, i32 1), align 2
+  ret void
+}
+
+; CHECK-LABEL: test_b8:
+; POWER7: addis [[REGSTRUCT:[0-9]+]], 2, b8v@toc@ha
+; POWER7-DAG: lbz [[REG0_0:[0-9]+]], b8v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: lbz [[REG1_0:[0-9]+]], b8v@toc@l+1([[REGSTRUCT]])
+; POWER7-DAG: lbz [[REG2_0:[0-9]+]], b8v@toc@l+2([[REGSTRUCT]])
+; POWER7-DAG: lbz [[REG3_0:[0-9]+]], b8v@toc@l+3([[REGSTRUCT]])
+; POWER7-DAG: lbz [[REG4_0:[0-9]+]], b8v@toc@l+4([[REGSTRUCT]])
+; POWER7-DAG: lbz [[REG5_0:[0-9]+]], b8v@toc@l+5([[REGSTRUCT]])
+; POWER7-DAG: lbz [[REG6_0:[0-9]+]], b8v@toc@l+6([[REGSTRUCT]])
+; POWER7-DAG: lbz [[REG7_0:[0-9]+]], b8v@toc@l+7([[REGSTRUCT]])
+; POWER7-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER7-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER7-DAG: addi [[REG2_1:[0-9]+]], [[REG2_0]], 3
+; POWER7-DAG: addi [[REG3_1:[0-9]+]], [[REG3_0]], 4
+; POWER7-DAG: addi [[REG4_1:[0-9]+]], [[REG4_0]], 5
+; POWER7-DAG: addi [[REG5_1:[0-9]+]], [[REG5_0]], 6
+; POWER7-DAG: addi [[REG6_1:[0-9]+]], [[REG6_0]], 7
+; POWER7-DAG: addi [[REG7_1:[0-9]+]], [[REG7_0]], 8
+; POWER7-DAG: stb [[REG0_1]], b8v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: stb [[REG1_1]], b8v@toc@l+1([[REGSTRUCT]])
+; POWER7-DAG: stb [[REG2_1]], b8v@toc@l+2([[REGSTRUCT]])
+; POWER7-DAG: stb [[REG3_1]], b8v@toc@l+3([[REGSTRUCT]])
+; POWER7-DAG: stb [[REG4_1]], b8v@toc@l+4([[REGSTRUCT]])
+; POWER7-DAG: stb [[REG5_1]], b8v@toc@l+5([[REGSTRUCT]])
+; POWER7-DAG: stb [[REG6_1]], b8v@toc@l+6([[REGSTRUCT]])
+; POWER7-DAG: stb [[REG7_1]], b8v@toc@l+7([[REGSTRUCT]])
+
+; POWER8: addis [[REGSTRUCT:[0-9]+]], 2, b8v@toc@ha
+; POWER8-NEXT: addi [[REGSTRUCT]], [[REGSTRUCT]], b8v@toc@l
+; POWER8-DAG: lbz [[REG0_0:[0-9]+]], 0([[REGSTRUCT]])
+; POWER8-DAG: lbz [[REG1_0:[0-9]+]], 1([[REGSTRUCT]])
+; POWER8-DAG: lbz [[REG2_0:[0-9]+]], 2([[REGSTRUCT]])
+; POWER8-DAG: lbz [[REG3_0:[0-9]+]], 3([[REGSTRUCT]])
+; POWER8-DAG: lbz [[REG4_0:[0-9]+]], 4([[REGSTRUCT]])
+; POWER8-DAG: lbz [[REG5_0:[0-9]+]], 5([[REGSTRUCT]])
+; POWER8-DAG: lbz [[REG6_0:[0-9]+]], 6([[REGSTRUCT]])
+; POWER8-DAG: lbz [[REG7_0:[0-9]+]], 7([[REGSTRUCT]])
+; POWER8-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER8-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER8-DAG: addi [[REG2_1:[0-9]+]], [[REG2_0]], 3
+; POWER8-DAG: addi [[REG3_1:[0-9]+]], [[REG3_0]], 4
+; POWER8-DAG: addi [[REG4_1:[0-9]+]], [[REG4_0]], 5
+; POWER8-DAG: addi [[REG5_1:[0-9]+]], [[REG5_0]], 6
+; POWER8-DAG: addi [[REG6_1:[0-9]+]], [[REG6_0]], 7
+; POWER8-DAG: addi [[REG7_1:[0-9]+]], [[REG7_0]], 8
+; POWER8-DAG: stb [[REG0_1]], 0([[REGSTRUCT]])
+; POWER8-DAG: stb [[REG1_1]], 1([[REGSTRUCT]])
+; POWER8-DAG: stb [[REG2_1]], 2([[REGSTRUCT]])
+; POWER8-DAG: stb [[REG3_1]], 3([[REGSTRUCT]])
+; POWER8-DAG: stb [[REG4_1]], 4([[REGSTRUCT]])
+; POWER8-DAG: stb [[REG5_1]], 5([[REGSTRUCT]])
+; POWER8-DAG: stb [[REG6_1]], 6([[REGSTRUCT]])
+; POWER8-DAG: stb [[REG7_1]], 7([[REGSTRUCT]])
+define void @test_b8() nounwind {
+entry:
+  %0 = load i8, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 0), align 1
+  %inc0 = add nsw i8 %0, 1
+  store i8 %inc0, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 0), align 1
+  %1 = load i8, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 1), align 1
+  %inc1 = add nsw i8 %1, 2
+  store i8 %inc1, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 1), align 1
+  %2 = load i8, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 2), align 1
+  %inc2 = add nsw i8 %2, 3
+  store i8 %inc2, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 2), align 1
+  %3 = load i8, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 3), align 1
+  %inc3 = add nsw i8 %3, 4
+  store i8 %inc3, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 3), align 1
+  %4 = load i8, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 4), align 1
+  %inc4 = add nsw i8 %4, 5
+  store i8 %inc4, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 4), align 1
+  %5 = load i8, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 5), align 1
+  %inc5 = add nsw i8 %5, 6
+  store i8 %inc5, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 5), align 1
+  %6 = load i8, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 6), align 1
+  %inc6 = add nsw i8 %6, 7
+  store i8 %inc6, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 6), align 1
+  %7 = load i8, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 7), align 1
+  %inc7 = add nsw i8 %7, 8
+  store i8 %inc7, i8* getelementptr inbounds (%struct.b8, %struct.b8* @b8v, i32 0, i32 7), align 1
+  ret void
+}
+
+; CHECK-LABEL: test_h4:
+; POWER7: addis [[REGSTRUCT:[0-9]+]], 2, h4v@toc@ha
+; POWER7-DAG: lhz [[REG0_0:[0-9]+]], h4v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: lhz [[REG1_0:[0-9]+]], h4v@toc@l+2([[REGSTRUCT]])
+; POWER7-DAG: lhz [[REG2_0:[0-9]+]], h4v@toc@l+4([[REGSTRUCT]])
+; POWER7-DAG: lhz [[REG3_0:[0-9]+]], h4v@toc@l+6([[REGSTRUCT]])
+; POWER7-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER7-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER7-DAG: addi [[REG2_1:[0-9]+]], [[REG2_0]], 3
+; POWER7-DAG: addi [[REG3_1:[0-9]+]], [[REG3_0]], 4
+; POWER7-DAG: sth [[REG0_1]], h4v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: sth [[REG1_1]], h4v@toc@l+2([[REGSTRUCT]])
+; POWER7-DAG: sth [[REG2_1]], h4v@toc@l+4([[REGSTRUCT]])
+; POWER7-DAG: sth [[REG3_1]], h4v@toc@l+6([[REGSTRUCT]])
+
+; POWER8: addis [[REGSTRUCT:[0-9]+]], 2, h4v@toc@ha
+; POWER8-NEXT: addi [[REGSTRUCT]], [[REGSTRUCT]], h4v@toc@l
+; POWER8-DAG: lhz [[REG0_0:[0-9]+]], 0([[REGSTRUCT]])
+; POWER8-DAG: lhz [[REG1_0:[0-9]+]], 2([[REGSTRUCT]])
+; POWER8-DAG: lhz [[REG2_0:[0-9]+]], 4([[REGSTRUCT]])
+; POWER8-DAG: lhz [[REG3_0:[0-9]+]], 6([[REGSTRUCT]])
+; POWER8-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER8-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER8-DAG: addi [[REG2_1:[0-9]+]], [[REG2_0]], 3
+; POWER8-DAG: addi [[REG3_1:[0-9]+]], [[REG3_0]], 4
+; POWER8-DAG: sth [[REG0_1]], 0([[REGSTRUCT]])
+; POWER8-DAG: sth [[REG1_1]], 2([[REGSTRUCT]])
+; POWER8-DAG: sth [[REG2_1]], 4([[REGSTRUCT]])
+; POWER8-DAG: sth [[REG3_1]], 6([[REGSTRUCT]])
+define void @test_h4() nounwind {
+entry:
+  %0 = load i16, i16* getelementptr inbounds (%struct.h4, %struct.h4* @h4v, i32 0, i32 0), align 2
+  %inc0 = add nsw i16 %0, 1
+  store i16 %inc0, i16* getelementptr inbounds (%struct.h4, %struct.h4* @h4v, i32 0, i32 0), align 2
+  %1 = load i16, i16* getelementptr inbounds (%struct.h4, %struct.h4* @h4v, i32 0, i32 1), align 2
+  %inc1 = add nsw i16 %1, 2
+  store i16 %inc1, i16* getelementptr inbounds (%struct.h4, %struct.h4* @h4v, i32 0, i32 1), align 2
+  %2 = load i16, i16* getelementptr inbounds (%struct.h4, %struct.h4* @h4v, i32 0, i32 2), align 2
+  %inc2 = add nsw i16 %2, 3
+  store i16 %inc2, i16* getelementptr inbounds (%struct.h4, %struct.h4* @h4v, i32 0, i32 2), align 2
+  %3 = load i16, i16* getelementptr inbounds (%struct.h4, %struct.h4* @h4v, i32 0, i32 3), align 2
+  %inc3 = add nsw i16 %3, 4
+  store i16 %inc3, i16* getelementptr inbounds (%struct.h4, %struct.h4* @h4v, i32 0, i32 3), align 2
+  ret void
+}
+
+; CHECK-LABEL: test_w2:
+; POWER7: addis [[REGSTRUCT:[0-9]+]], 2, w2v@toc@ha
+; POWER7-DAG: lwz [[REG0_0:[0-9]+]], w2v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: lwz [[REG1_0:[0-9]+]], w2v@toc@l+4([[REGSTRUCT]])
+; POWER7-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER7-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER7-DAG: stw [[REG0_1]], w2v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: stw [[REG1_1]], w2v@toc@l+4([[REGSTRUCT]])
+
+; POWER8: addis [[REGSTRUCT:[0-9]+]], 2, w2v@toc@ha
+; POWER8-NEXT: addi [[REGSTRUCT]], [[REGSTRUCT]], w2v@toc@l
+; POWER8-DAG: lwz [[REG0_0:[0-9]+]], 0([[REGSTRUCT]])
+; POWER8-DAG: lwz [[REG1_0:[0-9]+]], 4([[REGSTRUCT]])
+; POWER8-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER8-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER8-DAG: stw [[REG0_1]], 0([[REGSTRUCT]])
+; POWER8-DAG: stw [[REG1_1]], 4([[REGSTRUCT]])
+define void @test_w2() nounwind {
+entry:
+  %0 = load i32, i32* getelementptr inbounds (%struct.w2, %struct.w2* @w2v, i32 0, i32 0), align 4
+  %inc0 = add nsw i32 %0, 1
+  store i32 %inc0, i32* getelementptr inbounds (%struct.w2, %struct.w2* @w2v, i32 0, i32 0), align 4
+  %1 = load i32, i32* getelementptr inbounds (%struct.w2, %struct.w2* @w2v, i32 0, i32 1), align 4
+  %inc1 = add nsw i32 %1, 2
+  store i32 %inc1, i32* getelementptr inbounds (%struct.w2, %struct.w2* @w2v, i32 0, i32 1), align 4
+  ret void
+}
+
+; CHECK-LABEL: test_d2:
+; POWER7: addis [[REGSTRUCT:[0-9]+]], 2, d2v@toc@ha
+; POWER7-DAG: ld [[REG0_0:[0-9]+]], d2v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: ld [[REG1_0:[0-9]+]], d2v@toc@l+8([[REGSTRUCT]])
+; POWER7-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER7-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER7-DAG: std [[REG0_1]], d2v@toc@l([[REGSTRUCT]])
+; POWER7-DAG: std [[REG1_1]], d2v@toc@l+8([[REGSTRUCT]])
+
+; POWER8: addis [[REGSTRUCT:[0-9]+]], 2, d2v@toc@ha
+; POWER8-NEXT: addi [[REGSTRUCT]], [[REGSTRUCT]], d2v@toc@l
+; POWER8-DAG: ld [[REG0_0:[0-9]+]], 0([[REGSTRUCT]])
+; POWER8-DAG: ld [[REG1_0:[0-9]+]], 8([[REGSTRUCT]])
+; POWER8-DAG: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER8-DAG: addi [[REG1_1:[0-9]+]], [[REG1_0]], 2
+; POWER8-DAG: std [[REG0_1]], 0([[REGSTRUCT]])
+; POWER8-DAG: std [[REG1_1]], 8([[REGSTRUCT]])
+define void @test_d2() nounwind {
+entry:
+  %0 = load i64, i64* getelementptr inbounds (%struct.d2, %struct.d2* @d2v, i32 0, i32 0), align 8
+  %inc0 = add nsw i64 %0, 1
+  store i64 %inc0, i64* getelementptr inbounds (%struct.d2, %struct.d2* @d2v, i32 0, i32 0), align 8
+  %1 = load i64, i64* getelementptr inbounds (%struct.d2, %struct.d2* @d2v, i32 0, i32 1), align 8
+  %inc1 = add nsw i64 %1, 2
+  store i64 %inc1, i64* getelementptr inbounds (%struct.d2, %struct.d2* @d2v, i32 0, i32 1), align 8
+  ret void
+}
+
+; Make sure the optimization fires on power8 if there is a single use resulting
+; in a better fusion opportunity.
+; register 3 is the return value, so it should be chosen
+; CHECK-LABEL: test_singleuse:
+; CHECK: addis 3, 2, d2v@toc@ha
+; CHECK: ld 3, d2v@toc@l+8(3)
+define i64 @test_singleuse() nounwind {
+entry:
+  %0 = load i64, i64* getelementptr inbounds (%struct.d2, %struct.d2* @d2v, i32 0, i32 1), align 8
+  ret i64 %0
+}
+
+; Make sure the optimization fails to fire if the symbol is aligned, but the offset is not.
+; CHECK-LABEL: test_misalign
+; POWER7: addis [[REGSTRUCT_0:[0-9]+]], 2, misalign_v@toc@ha
+; POWER7: addi [[REGSTRUCT:[0-9]+]], [[REGSTRUCT_0]], misalign_v@toc@l
+; POWER7: li [[OFFSET_REG:[0-9]+]], 1
+; POWER7: ldx [[REG0_0:[0-9]+]], [[REGSTRUCT]], [[OFFSET_REG]]
+; POWER7: addi [[REG0_1:[0-9]+]], [[REG0_0]], 1
+; POWER7: stdx [[REG0_1]], [[REGSTRUCT]], [[OFFSET_REG]]
+define void @test_misalign() nounwind {
+entry:
+  %0 = load i64, i64* getelementptr inbounds (%struct.misalign, %struct.misalign* @misalign_v, i32 0, i32 1), align 1
+  %inc0 = add nsw i64 %0, 1
+  store i64 %inc0, i64* getelementptr inbounds (%struct.misalign, %struct.misalign* @misalign_v, i32 0, i32 1), align 1
+  ret void
+}