[lld-macho] Add LOH_ARM64_ADRP_LDR_GOT_LDR optimization hint support

[BertalanD]

Summary

[lld-macho] Add LOH_ARM64_ADRP_LDR_GOT_LDR optimization hint support

This hint relaxes a load from an address specified in the GOT. If the
symbol is external and the GOT is within a +/- 1 MiB range, the adrp +
add used for loading the GOT is turned into adr + nop. If the symbol is
local, the adrp + add + ldr sequence is turned into a literal ldr or
adr(p) + ldr if possible.

This type accounts for more than half of all LOHs in chromium_framework.

This commit moves the eligibility checks into helper functions to
improve the readability of the LOH processing code. Ho functional
changes are intended to the previously implemented LOH types.

[tschuett]

The conditions in the if and else if hurt my eyes. Maybe comments could help your reviewers and future readers? Maybe hoist the 1 << 20 as a constant up?

[BertalanD]

The conditions in the if and else if hurt my eyes

That's a bit harsh, but true. I moved the range checks into helper functions, and the code looks a lot more readable IMO.

[tschuett]

In D129427#3642664, @BertalanD wrote:

The conditions in the if and else if hurt my eyes

That's a bit harsh, but true. I moved the range checks into helper functions, and the code looks a lot more readable IMO.

Sorry, that was not meant to offensive.

[int3]

lld/MachO/Arch/ARM64.cpp

257–259

MathExtras.h has isInt -- I think we could just call that inline (i.e. don't need this function)

271

I know Ldr is small enough to be copyable but personally I'd just pass by const ref here so we don't have to think about that

272

can this just be an assert?

273–274

MathExtras has isShiftedInt

(also consider calling that inline; I'm not sure the sizeValid check / assert above is really necessary)

281

could use maskTrailingOnes instead of the hex constant

302

so this is checking specifically for the unsigned variant of the instruction, right? I assume we don't care about the signed variants because we generally expect the offset to be positive? either way, would be nice to explain in a comment :)

304

delta3? you mean ldr.offset?

306

might have mentioned it in a previous diff, but I would love if we could store p2Size instead of size (i.e. store the power-of-two exponent value). Would let us skip the relatively expensive div/mod operations + writing in bitshifts makes it clearer what the encoding constraints are.

(not gonna block the diff for this though)

306–307

parens unnecessary (I get that it is hard to remember operator precedence in general, but I think it's easy enough to remember that non-bitwise math operations are always higher prec than equality checks, and that the short-circuiting operators are lower prec than both)

326

I think we could avoid the cast if we declared enum ExtendType : uint8_t

[BertalanD]

Use helper functions from MathExtras.h, store the log2 of load sizes, remove a couple of redundant checks.

[BertalanD]

lld/MachO/Arch/ARM64.cpp

257–259

Thank you for telling me about the MathExtras.h functions, I didn't know they existed. I updated the diff to use those, but I'd like to keep the helper functions for readability.

272

Currently, this function *will* be called with ldr.size == 1 and ldr.size == 2, and we fall back to the ADR+LDR transformation if that happens (see line 150 of the test)

[int3]

lgtm, thanks!

lld/MachO/Arch/ARM64.cpp

281

missed this one

302

Thanks for the comment. Am I right to say that the offsets that come in here are usually positive, though? Or is that not true (just curious)

[BertalanD]

lld/MachO/Arch/ARM64.cpp

302

I think you're right. This might actually *never* be negative in normal circumstances. We're transforming

adrp x0, _foo@PAGE
add  x1, x0, _foo@PAGEOFF
ldr  x2, [x1, #off]

into

adrp x0, _foo@PAGE
nop
ldr x2, [x0, _foo@PAGEOFF + #off]

#off cannot be negative because the third instruction of a well-formed AdrpLdrGotLdr sequence must be an LDRXui or LDRWui which have an unsigned offset, and _foo@PAGEOFF is evidently a positive value.

[BertalanD]

lld/MachO/Arch/ARM64.cpp

281

Would that look something like this?

uint32_t imm19 = (ldr.offset & ~maskTrailingOnes<uint32_t>(2)) << 3;

Honestly, that doesn't look more readable to me.

[int3]

lld/MachO/Arch/ARM64.cpp

281

I think it is

uint32_t imm19 = (ldr.offset & maskTrailingOnes<uint32_t>(19)) << 3;

Having to specify the template param is a bit unfortunate, but mainly I think it's nice to have the 19 explicit there, rather than having to parse that from the hex constant. Alternatively (1 << 20) - 1 would be similarly explicit.

Not a big deal though, feel free to keep the hex constant if you prefer

302

That's what I thought. Thanks for confirming :)

[BertalanD]

lld/MachO/Arch/ARM64.cpp

281

Wouldn't it be this instead?

uint32_t imm19 = (ldr.offset & maskTrailingOnes<uint32_t>(21)) << 3;

The 5 least significant bits encode the destination register and the next 19 bits the offset divided by 4. We assert that the offset is a multiple of 4, so after the shift, the 5 LSB will be all zeros.

[int3]

lld/MachO/Arch/ARM64.cpp

281

d'oh, of course :) thanks

lld/MachO/Arch/ARM64.cpp

 enum ExtendType { ZeroExtend = 1, Sign64 = 2, Sign32 = 3 };
 
 struct Ldr {
   uint8_t destRegister;
   uint8_t baseRegister;
   uint8_t size;
   bool isFloat;
   ExtendType extendType;
-  uint64_t offset;
+  int64_t offset;
 };
 
 struct PerformedReloc {
   const Reloc &rel;
   uint64_t referentVA;
 };
 
 class OptimizationHintContext {
 public:
   OptimizationHintContext(uint8_t *buf, const ConcatInputSection *isec,
                           ArrayRef<uint64_t> relocTargets)
       : buf(buf), isec(isec), relocTargets(relocTargets),
         relocIt(isec->relocs.rbegin()) {}
 
   void applyAdrpAdd(const OptimizationHint &);
   void applyAdrpAdrp(const OptimizationHint &);
   void applyAdrpLdr(const OptimizationHint &);
   void applyAdrpLdrGot(const OptimizationHint &);
+  void applyAdrpLdrGotLdr(const OptimizationHint &);
 
 private:
   uint8_t *buf;
   const ConcatInputSection *isec;
   ArrayRef<uint64_t> relocTargets;
   std::vector<Reloc>::const_reverse_iterator relocIt;
 
   uint64_t getRelocTarget(const Reloc &);
if ((insn & 0x3fc00000) == 0x39400000) {
     // LDRSB (immediate), LDRSH (immediate), LDRSW (immediate)
     ldr.size = 1 << size;
     ldr.extendType = static_cast<ExtendType>(opc);
     ldr.isFloat = false;
   } else if ((insn & 0x3f400000) == 0x3d400000) {
     // LDR (immediate, SIMD&FP)
     ldr.extendType = ZeroExtend;
     ldr.isFloat = true;
-    if (size == 2 && opc == 1)
-      ldr.size = 4;
-    else if (size == 3 && opc == 1)
-      ldr.size = 8;
+    if (opc == 1)
+      ldr.size = 1 << size;
     else if (size == 0 && opc == 3)
       ldr.size = 16;
     else
       return false;
   } else {
     return false;
   }
   ldr.offset = ((insn >> 10) & 0xfff) * ldr.size;
   return true;
 }
 
+static bool isValidAdrOffset(int32_t delta) {
+  return -(1 << 20) <= delta && delta < (1 << 20);
+}
+
 static void writeAdr(void *loc, uint32_t dest, int32_t delta) {
+  assert(isValidAdrOffset(delta));
   uint32_t opcode = 0x10000000;
   uint32_t immHi = (delta & 0x001ffffc) << 3;
   uint32_t immLo = (delta & 0x00000003) << 29;
   write32le(loc, opcode | immHi | immLo | dest);
 }
 
 static void writeNop(void *loc) { write32le(loc, 0xd503201f); }
 
-static void writeLiteralLdr(void *loc, Ldr original, int32_t delta) {
-  uint32_t imm19 = (delta & 0x001ffffc) << 3;
-  uint32_t opcode = 0;
-  switch (original.size) {
+static bool isLiteralLdrEligible(Ldr ldr) {
+  bool sizeValid = ldr.size == 4 || ldr.size == 8 || ldr.size == 16;
+  bool aligned = (ldr.offset & 3) == 0;
+  bool inRange = -(1 << 20) <= ldr.offset && ldr.offset < (1 << 20);
+
+  return sizeValid && aligned && inRange;
+}
+
+static void writeLiteralLdr(void *loc, Ldr ldr) {
+  assert(isLiteralLdrEligible(ldr));
+  uint32_t imm19 = (ldr.offset & 0x001ffffc) << 3;
+  uint32_t opcode;
+  switch (ldr.size) {
   case 4:
-    if (original.isFloat)
+    if (ldr.isFloat)
       opcode = 0x1c000000;
     else
-      opcode = original.extendType == Sign64 ? 0x98000000 : 0x18000000;
+      opcode = ldr.extendType == Sign64 ? 0x98000000 : 0x18000000;
     break;
   case 8:
-    opcode = original.isFloat ? 0x5c000000 : 0x58000000;
+    opcode = ldr.isFloat ? 0x5c000000 : 0x58000000;
     break;
   case 16:
     opcode = 0x9c000000;
     break;
   default:
-    assert(false && "Invalid size for literal ldr");
+    llvm_unreachable("Invalid literal ldr size");
   }
-  write32le(loc, opcode | imm19 | original.destRegister);
+  write32le(loc, opcode | imm19 | ldr.destRegister);
+}
+
+static bool isImmediateLdrEligible(Ldr ldr) {
+  // Note: We deviate from ld64 behavior, which converts to immediate loads
+  // only if delta3 < 4096, even though the offset is divided by the load's size
+  // in the 12-bit immediate operand.
+  bool aligned = (ldr.offset % ldr.size) == 0;
+  bool inRange = (ldr.offset >= 0) && (ldr.offset < 4096 * ldr.size);
+
+  return aligned && inRange;
+}
+
+static void writeImmediateLdr(void *loc, Ldr ldr) {
+  assert(isImmediateLdrEligible(ldr));
+  uint32_t opcode = 0x39000000;
+  if (ldr.isFloat) {
+    opcode |= 0x04000000;
+    assert(ldr.extendType == ZeroExtend);
+  }
+  opcode |= ldr.destRegister;
+  opcode |= ldr.baseRegister << 5;
+  uint8_t size, opc;
+  if (ldr.size == 16) {
+    size = 0;
+    opc = 3;
+  } else {
+    opc = static_cast<uint8_t>(ldr.extendType);
+    switch (ldr.size) {
+    case 1:
+      size = 0;
+      break;
+    case 2:
+      size = 1;
+      break;
+    case 4:
+      size = 2;
+      assert(ldr.extendType != Sign32);
+      break;
+    case 8:
+      size = 3;
+      assert(ldr.extendType == ZeroExtend);
+      break;
+    default:
+      llvm_unreachable("Invalid immediate ldr size");
+    }
+  }
+  uint32_t imm12 = (ldr.offset / ldr.size) << 10;
+  write32le(loc, opcode | imm12 | (opc << 22) | (size << 30));
 }
 
 uint64_t OptimizationHintContext::getRelocTarget(const Reloc &reloc) {
   size_t relocIdx = &reloc - isec->relocs.data();
   return relocTargets[relocIdx];
 }
 
 // Optimization hints are sorted in a monotonically increasing order by their
void OptimizationHintContext::applyAdrpAdd(const OptimizationHint &hint) {
 
   Optional<PerformedReloc> rel1 = findPrimaryReloc(hint.offset0);
   Optional<PerformedReloc> rel2 = findReloc(hint.offset0 + hint.delta[0]);
   if (!rel1 || !rel2)
     return;
   if (rel1->referentVA != rel2->referentVA)
     return;
   int64_t delta = rel1->referentVA - rel1->rel.offset - isec->getVA();
-  if (delta >= (1 << 20) || delta < -(1 << 20))
+  if (!isValidAdrOffset(delta))
     return;
 
   writeAdr(buf + hint.offset0, add.destRegister, delta);
   writeNop(buf + hint.offset0 + hint.delta[0]);
 }
 
 // Transforms two adrp instructions into a single adrp if their referent
 // addresses are located on the same 4096 byte page.
void OptimizationHintContext::applyAdrpLdr(const OptimizationHint &hint) {
   if (!rel1 || !rel2)
     return;
   if (ldr.offset != (rel1->referentVA & 0xfff))
     return;
   if ((rel1->referentVA & 3) != 0)
     return;
   if (ldr.size == 1 || ldr.size == 2)
     return;
-  int64_t delta = rel1->referentVA - rel2->rel.offset - isec->getVA();
-  if (delta >= (1 << 20) || delta < -(1 << 20))
+  ldr.offset = rel1->referentVA - rel2->rel.offset - isec->getVA();
+  if (!isLiteralLdrEligible(ldr))
     return;
 
   writeNop(buf + hint.offset0);
-  writeLiteralLdr(buf + hint.offset0 + hint.delta[0], ldr, delta);
+  writeLiteralLdr(buf + hint.offset0 + hint.delta[0], ldr);
 }
 
 // GOT loads are emitted by the compiler as a pair of adrp and ldr instructions,
 // but they may be changed to adrp+add by relaxGotLoad(). This hint performs
 // the AdrpLdr or AdrpAdd transformation depending on whether it was relaxed.
 void OptimizationHintContext::applyAdrpLdrGot(const OptimizationHint &hint) {
   uint32_t ins2 = read32le(buf + hint.offset0 + hint.delta[0]);
   Add add;
   Ldr ldr;
   if (parseAdd(ins2, add))
     applyAdrpAdd(hint);
   else if (parseLdr(ins2, ldr))
     applyAdrpLdr(hint);
 }
 
+// Relaxes a load from an address specified in the GOT.
+// If the symbol is external and the GOT is within a +/- 1 MiB range, the
+// adrp + add used for loading the GOT is turned into adr + nop.
+// If the symbol is local, the adrp + add + ldr sequence is turned into a
+// literal ldr or adr(p) + ldr if possible.
+void OptimizationHintContext::applyAdrpLdrGotLdr(const OptimizationHint &hint) {
+  uint32_t ins1 = read32le(buf + hint.offset0);
+  Adrp adrp;
+  if (!parseAdrp(ins1, adrp))
+    return;
+  uint32_t ins3 = read32le(buf + hint.offset0 + hint.delta[1]);
+  Ldr ldr3;
+  if (!parseLdr(ins3, ldr3))
+    return;
+  uint32_t ins2 = read32le(buf + hint.offset0 + hint.delta[0]);
+  Ldr ldr2;
+  Add add2;
+
+  Optional<PerformedReloc> rel1 = findPrimaryReloc(hint.offset0);
+  Optional<PerformedReloc> rel2 = findReloc(hint.offset0 + hint.delta[0]);
+  if (!rel1 || !rel2)
+    return;
+
+  if (parseAdd(ins2, add2)) {
+    // adrp x0, _foo@PAGE
+    // add  x1, x0, _foo@PAGEOFF
+    // ldr  x2, [x1, #off]
+
+    if (adrp.destRegister != add2.srcRegister)
+      return;
+    if (add2.destRegister != ldr3.baseRegister)
+      return;
+
+    // Load from the target address directly.
+    //   nop
+    //   nop
+    //   ldr x2, [_foo + #off]
+    uint64_t rel3VA = hint.offset0 + hint.delta[1] + isec->getVA();
+    Ldr literalLdr = ldr3;
+    literalLdr.offset += rel1->referentVA - rel3VA;
+    if (isLiteralLdrEligible(literalLdr)) {
+      writeNop(buf + hint.offset0);
+      writeNop(buf + hint.offset0 + hint.delta[0]);
+      writeLiteralLdr(buf + hint.offset0 + hint.delta[1], literalLdr);
+      return;
+    }
+
+    // Load the target address into a register and load from there indirectly.
+    //   adr x1, _foo
+    //   nop
+    //   ldr x2, [x1, #off]
+    int64_t adrOffset = rel1->referentVA - rel1->rel.offset - isec->getVA();
+    if (isValidAdrOffset(adrOffset)) {
+      writeAdr(buf + hint.offset0, ldr3.baseRegister, adrOffset);
+      writeNop(buf + hint.offset0 + hint.delta[0]);
+      return;
+    }
+
+    // Move adding the target's page offset into the ldr's immediate offset.
+    //   adrp x0, _foo@PAGE
+    //   nop
+    //   ldr x2, [x0, _foo@PAGEOFF + #off]
+    Ldr immediateLdr = ldr3;
+    immediateLdr.baseRegister = adrp.destRegister;
+    immediateLdr.offset += add2.addend;
+    if (isImmediateLdrEligible(immediateLdr)) {
+      writeNop(buf + hint.offset0 + hint.delta[0]);
+      writeImmediateLdr(buf + hint.offset0 + hint.delta[1], immediateLdr);
+      return;
+    }
+  } else if (parseLdr(ins2, ldr2)) {
+    // adrp x1, _foo@GOTPAGE
+    // ldr  x2, [x1, _foo@GOTPAGEOFF]
+    // ldr  x3, [x2, #off]
+    if (ldr2.baseRegister != adrp.destRegister)
+      return;
+    if (ldr3.baseRegister != ldr2.destRegister)
+      return;
+    if (ldr2.size != 8 || ldr2.isFloat)
+      return;
+
+    // Load the GOT entry's address directly.
+    //   nop
+    //   ldr x2, _foo@GOTPAGE + _foo@GOTPAGEOFF
+    //   ldr x3, [x2, #off]
+    Ldr literalLdr = ldr2;
+    literalLdr.offset = rel1->referentVA - rel2->rel.offset - isec->getVA();
+    if (isLiteralLdrEligible(literalLdr)) {
+      writeNop(buf + hint.offset0);
+      writeLiteralLdr(buf + hint.offset0 + hint.delta[0], literalLdr);
+    }
+  }
+}
+
 void ARM64::applyOptimizationHints(uint8_t *buf, const ConcatInputSection *isec,
                                    ArrayRef<uint64_t> relocTargets) const {
   assert(isec);
   assert(relocTargets.size() == isec->relocs.size());
 
   // Note: Some of these optimizations might not be valid when shared regions
   // are in use. Will need to revisit this if splitSegInfo is added.
 
   OptimizationHintContext ctx1(buf, isec, relocTargets);
   for (const OptimizationHint &hint : isec->optimizationHints) {
     switch (hint.type) {
     case LOH_ARM64_ADRP_ADRP:
       // This is done in another pass because the other optimization hints
       // might cause its targets to be turned into NOPs.
       break;
     case LOH_ARM64_ADRP_LDR:
       ctx1.applyAdrpLdr(hint);
       break;
     case LOH_ARM64_ADRP_ADD_LDR:
+      // TODO: Implement this
+      break;
     case LOH_ARM64_ADRP_LDR_GOT_LDR:
+      ctx1.applyAdrpLdrGotLdr(hint);
+      break;
     case LOH_ARM64_ADRP_ADD_STR:
     case LOH_ARM64_ADRP_LDR_GOT_STR:
       // TODO: Implement these
       break;
     case LOH_ARM64_ADRP_ADD:
       ctx1.applyAdrpAdd(hint);
       break;
     case LOH_ARM64_ADRP_LDR_GOT:

lld/test/MachO/loh-adrp-ldr-got-ldr.s

+# REQUIRES: aarch64
+
+# RUN: rm -rf %t; split-file %s %t
+# RUN: llvm-mc -filetype=obj -triple=arm64-apple-darwin %t/lib.s -o %t/lib.o
+# RUN: %lld -arch arm64 -dylib -o %t/lib.dylib %t/lib.o
+# RUN: llvm-mc -filetype=obj -triple=arm64-apple-darwin %t/external.s -o %t/near-got.o
+# RUN: llvm-mc -filetype=obj -triple=arm64-apple-darwin %t/external.s -defsym=PADDING=1 -o %t/far-got.o
+# RUN: llvm-mc -filetype=obj -triple=arm64-apple-darwin %t/local.s -o %t/local.o
+# RUN: %lld -arch arm64 %t/near-got.o %t/lib.dylib -o %t/NearGot
+# RUN: %lld -arch arm64 %t/far-got.o %t/lib.dylib -o %t/FarGot
+# RUN: %lld -arch arm64 %t/local.o -o %t/Local
+# RUN: llvm-objdump -d --macho %t/NearGot | FileCheck %s -check-prefix=NEAR-GOT
+# RUN: llvm-objdump -d --macho %t/FarGot | FileCheck %s -check-prefix=FAR-GOT
+# RUN: llvm-objdump -d --macho %t/Local | FileCheck %s -check-prefix=LOCAL
+
+#--- external.s
+.text
+.align 2
+.globl _main
+_main:
+
+## Basic test
+L1: adrp x0, _external@GOTPAGE
+L2: ldr  x1, [x0, _external@GOTPAGEOFF]
+L3: ldr  x2, [x1]
+# NEAR-GOT-LABEL: _main:
+# NEAR-GOT-NEXT: nop
+# NEAR-GOT-NEXT: ldr x1, #{{.*}} ; literal pool symbol address: _external
+# NEAR-GOT-NEXT: ldr x2, [x1]
+# FAR-GOT-LABEL: _main:
+# FAR-GOT-NEXT:  adrp x0
+# FAR-GOT-NEXT:  ldr x1
+# FAR-GOT-NEXT:  ldr x2, [x1]
+
+## The second load has an offset
+L4: adrp x0, _external@GOTPAGE
+L5: ldr  x1, [x0, _external@GOTPAGEOFF]
+L6: ldr  q2, [x1, #16]
+# NEAR-GOT-NEXT: nop
+# NEAR-GOT-NEXT: ldr x1, #{{.*}} ; literal pool symbol address: _external
+# NEAR-GOT-NEXT: ldr q2, [x1, #16]
+# FAR-GOT-NEXT:  adrp x0
+# FAR-GOT-NEXT:  ldr x1
+# FAR-GOT-NEXT:  ldr q2, [x1, #16]
+
+### Tests for invalid inputs
+.ifndef PADDING
+## Registers don't match
+L7: adrp x0, _external@GOTPAGE
+L8: ldr  x1, [x1, _external@GOTPAGEOFF]
+L9: ldr  x2, [x1]
+# NEAR-GOT-NEXT: adrp x0
+# NEAR-GOT-NEXT: ldr x1
+# NEAR-GOT-NEXT: ldr x2, [x1]
+
+## Registers don't match
+L10: adrp x0, _external@GOTPAGE
+L11: ldr  x1, [x0, _external@GOTPAGEOFF]
+L12: ldr  x2, [x0]
+# NEAR-GOT-NEXT: adrp x0
+# NEAR-GOT-NEXT: ldr x1
+# NEAR-GOT-NEXT: ldr x2, [x0]
+
+## Not an LDR (immediate)
+L13: adrp x0, _external@GOTPAGE
+L14: ldr  x1, 0
+L15: ldr  x2, [x1]
+# NEAR-GOT-NEXT: adrp x0
+# NEAR-GOT-NEXT: ldr x1
+# NEAR-GOT-NEXT: ldr x2, [x1]
+
+.loh AdrpLdrGotLdr L7, L8, L9
+.loh AdrpLdrGotLdr L10, L11, L12
+.loh AdrpLdrGotLdr L13, L14, L15
+.endif
+
+.loh AdrpLdrGotLdr L1, L2, L3
+.loh AdrpLdrGotLdr L4, L5, L6
+
+.ifdef PADDING
+.space 1048576
+.endif
+.data
+
+
+#--- lib.s
+.data
+.align 4
+.globl _external
+_external:
+    .zero 32
+
+#--- local.s
+.text
+.align 2
+.globl _main
+_main:
+
+### Transformation to a literal LDR
+## Basic case
+L1: adrp x0, _close@GOTPAGE
+L2: ldr  x1, [x0, _close@GOTPAGEOFF]
+L3: ldr  x2, [x1]
+# LOCAL-LABEL: _main:
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldr x2
+
+## Load with offset
+L4: adrp x0, _close@GOTPAGE
+L5: ldr  x1, [x0, _close@GOTPAGEOFF]
+L6: ldr  x2, [x1, #8]
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldr x2
+
+## 32 bit load
+L7: adrp x0, _close@GOTPAGE
+L8: ldr  x1, [x0, _close@GOTPAGEOFF]
+L9: ldr  w1, [x1]
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldr w1, _close
+
+## Floating point
+L10: adrp x0, _close@GOTPAGE
+L11: ldr  x1, [x0, _close@GOTPAGEOFF]
+L12: ldr  s1, [x1]
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldr s1, _close
+
+L13: adrp x0, _close@GOTPAGE
+L14: ldr  x1, [x0, _close@GOTPAGEOFF]
+L15: ldr  d1, [x1, #8]
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldr d1, _close8
+
+L16: adrp x0, _close@GOTPAGE
+L17: ldr  x1, [x0, _close@GOTPAGEOFF]
+L18: ldr  q0, [x1]
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldr q0, _close
+
+
+### Transformation to ADR+LDR
+## 1 byte floating point load
+L19: adrp x0, _close@GOTPAGE
+L20: ldr  x1, [x0, _close@GOTPAGEOFF]
+L21: ldr  b2, [x1]
+# LOCAL-NEXT: adr x1
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldr b2, [x1]
+
+## 1 byte GPR load, zero extend
+L22: adrp x0, _close@GOTPAGE
+L23: ldr  x1, [x0, _close@GOTPAGEOFF]
+L24: ldrb w2, [x1]
+# LOCAL-NEXT: adr x1
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldrb w2, [x1]
+
+## 1 byte GPR load, sign extend
+L25: adrp  x0, _close@GOTPAGE
+L26: ldr   x1, [x0, _close@GOTPAGEOFF]
+L27: ldrsb x2, [x1]
+# LOCAL-NEXT: adr x1
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldrsb x2, [x1]
+
+## Unaligned
+L28: adrp x0, _unaligned@GOTPAGE
+L29: ldr  x1, [x0, _unaligned@GOTPAGEOFF]
+L30: ldr  x2, [x1]
+# LOCAL-NEXT: adr x1
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldr x2, [x1]
+
+
+### Transformation to ADRP + imediate LDR
+## Basic test: target is far
+L31: adrp x0, _far@GOTPAGE
+L32: ldr  x1, [x0, _far@GOTPAGEOFF]
+L33: ldr  x2, [x1]
+# LOCAL-NEXT: adrp x0
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldr x2
+
+## With offset
+L34: adrp x0, _far@GOTPAGE
+L35: ldr  x1, [x0, _far@GOTPAGEOFF]
+L36: ldr  x2, [x1, #8]
+# LOCAL-NEXT: adrp x0
+# LOCAL-NEXT: nop
+# LOCAL-NEXT: ldr x2
+
+### No changes other than GOT relaxation
+## Far and unaligned
+L37: adrp x0, _far_unaligned@GOTPAGE
+L38: ldr  x1, [x0, _far_unaligned@GOTPAGEOFF]
+L39: ldr  x2, [x1]
+# LOCAL-NEXT: adrp x0
+# LOCAL-NEXT: add x1, x0
+# LOCAL-NEXT: ldr x2, [x1]
+
+## Far with large offset (_far_offset@GOTPAGEOFF + #255 > 4095)
+L40: adrp x0, _far_offset@GOTPAGE
+L41: ldr  x1, [x0, _far_offset@GOTPAGEOFF]
+L42: ldrb w2, [x1, #255]
+# LOCAL-NEXT: adrp x0
+# LOCAL-NEXT: add x1, x0
+# LOCAL-NEXT: ldrb w2, [x1, #255]
+
+### Tests for invalid inputs, only GOT relaxation should happen
+## Registers don't match
+L43: adrp x0, _far@GOTPAGE
+L44: ldr  x1, [x0, _far@GOTPAGEOFF]
+L45: ldr  x2, [x2]
+# LOCAL-NEXT: adrp x0
+# LOCAL-NEXT: add x1, x0
+# LOCAL-NEXT: ldr x2, [x2]
+
+.data
+.align 4
+    .quad 0
+_close:
+    .quad 0
+_close8:
+    .quad 0
+    .byte 0
+_unaligned:
+    .quad 0
+
+.space 1048576
+.align 12
+    .quad 0
+_far:
+     .quad 0
+    .byte 0
+_far_unaligned:
+    .quad 0
+.space 4000
+_far_offset:
+    .byte 0
+
+
+.loh AdrpLdrGotLdr L1, L2, L3
+.loh AdrpLdrGotLdr L4, L5, L6
+.loh AdrpLdrGotLdr L7, L8, L9
+.loh AdrpLdrGotLdr L10, L11, L12
+.loh AdrpLdrGotLdr L13, L14, L15
+.loh AdrpLdrGotLdr L16, L17, L18
+.loh AdrpLdrGotLdr L19, L20, L21
+.loh AdrpLdrGotLdr L22, L23, L24
+.loh AdrpLdrGotLdr L25, L26, L27
+.loh AdrpLdrGotLdr L28, L29, L30
+.loh AdrpLdrGotLdr L31, L32, L33
+.loh AdrpLdrGotLdr L34, L35, L36
+.loh AdrpLdrGotLdr L37, L38, L39
+.loh AdrpLdrGotLdr L40, L41, L42
+.loh AdrpLdrGotLdr L43, L44, L45