[LoongArch] Allow delayed decision for ADD/SUB relocations (llvm#72960)

Refer to RISCV [1], LoongArch also need delayed decision for ADD/SUB
relocations. In handleAddSubRelocations, just return directly if SecA !=
SecB, handleFixup usually will finish the the rest of creating PCRel
relocations works. Otherwise we emit relocs depends on whether
relaxation is enabled. If not, we return true and avoid record ADD/SUB
relocations.
Now the two symbols separated by alignment directive will return without
folding symbol offset in AttemptToFoldSymbolOffsetDifference, which has
the same effect when relaxation is enabled.

[1] https://reviews.llvm.org/D155357

Author: MQ-mengqing

llvm/lib/MC/MCExpr.cpp

@@ -632,7 +632,8 @@ static void AttemptToFoldSymbolOffsetDifference(
   // instructions and InSet is false (not expressions in directive like
   // .size/.fill), disable the fast path.
   if (Layout && (InSet || !SecA.hasInstructions() ||
-                 !Asm->getContext().getTargetTriple().isRISCV())) {
+                 !(Asm->getContext().getTargetTriple().isRISCV() ||
+                   Asm->getContext().getTargetTriple().isLoongArch()))) {
     // If both symbols are in the same fragment, return the difference of their
     // offsets. canGetFragmentOffset(FA) may be false.
     if (FA == FB && !SA.isVariable() && !SB.isVariable()) {

llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.cpp

@@ -177,6 +177,34 @@ bool LoongArchAsmBackend::shouldForceRelocation(const MCAssembler &Asm,
   }
 }
 
+static inline std::pair<MCFixupKind, MCFixupKind>
+getRelocPairForSize(unsigned Size) {
+  switch (Size) {
+  default:
+    llvm_unreachable("unsupported fixup size");
+  case 6:
+    return std::make_pair(
+        MCFixupKind(FirstLiteralRelocationKind + ELF::R_LARCH_ADD6),
+        MCFixupKind(FirstLiteralRelocationKind + ELF::R_LARCH_SUB6));
+  case 8:
+    return std::make_pair(
+        MCFixupKind(FirstLiteralRelocationKind + ELF::R_LARCH_ADD8),
+        MCFixupKind(FirstLiteralRelocationKind + ELF::R_LARCH_SUB8));
+  case 16:
+    return std::make_pair(
+        MCFixupKind(FirstLiteralRelocationKind + ELF::R_LARCH_ADD16),
+        MCFixupKind(FirstLiteralRelocationKind + ELF::R_LARCH_SUB16));
+  case 32:
+    return std::make_pair(
+        MCFixupKind(FirstLiteralRelocationKind + ELF::R_LARCH_ADD32),
+        MCFixupKind(FirstLiteralRelocationKind + ELF::R_LARCH_SUB32));
+  case 64:
+    return std::make_pair(
+        MCFixupKind(FirstLiteralRelocationKind + ELF::R_LARCH_ADD64),
+        MCFixupKind(FirstLiteralRelocationKind + ELF::R_LARCH_SUB64));
+  }
+}
+
 bool LoongArchAsmBackend::writeNopData(raw_ostream &OS, uint64_t Count,
                                        const MCSubtargetInfo *STI) const {
   // We mostly follow binutils' convention here: align to 4-byte boundary with a
@@ -191,6 +219,56 @@ bool LoongArchAsmBackend::writeNopData(raw_ostream &OS, uint64_t Count,
   return true;
 }
 
+bool LoongArchAsmBackend::handleAddSubRelocations(const MCAsmLayout &Layout,
+                                                  const MCFragment &F,
+                                                  const MCFixup &Fixup,
+                                                  const MCValue &Target,
+                                                  uint64_t &FixedValue) const {
+  std::pair<MCFixupKind, MCFixupKind> FK;
+  uint64_t FixedValueA, FixedValueB;
+  const MCSection &SecA = Target.getSymA()->getSymbol().getSection();
+  const MCSection &SecB = Target.getSymB()->getSymbol().getSection();
+
+  // We need record relocation if SecA != SecB. Usually SecB is same as the
+  // section of Fixup, which will be record the relocation as PCRel. If SecB
+  // is not same as the section of Fixup, it will report error. Just return
+  // false and then this work can be finished by handleFixup.
+  if (&SecA != &SecB)
+    return false;
+
+  // In SecA == SecB case. If the linker relaxation is enabled, we need record
+  // the ADD, SUB relocations. Otherwise the FixedValue has already been
+  // calculated out in evaluateFixup, return true and avoid record relocations.
+  if (!STI.hasFeature(LoongArch::FeatureRelax))
+    return true;
+
+  switch (Fixup.getKind()) {
+  case llvm::FK_Data_1:
+    FK = getRelocPairForSize(8);
+    break;
+  case llvm::FK_Data_2:
+    FK = getRelocPairForSize(16);
+    break;
+  case llvm::FK_Data_4:
+    FK = getRelocPairForSize(32);
+    break;
+  case llvm::FK_Data_8:
+    FK = getRelocPairForSize(64);
+    break;
+  default:
+    llvm_unreachable("unsupported fixup size");
+  }
+  MCValue A = MCValue::get(Target.getSymA(), nullptr, Target.getConstant());
+  MCValue B = MCValue::get(Target.getSymB());
+  auto FA = MCFixup::create(Fixup.getOffset(), nullptr, std::get<0>(FK));
+  auto FB = MCFixup::create(Fixup.getOffset(), nullptr, std::get<1>(FK));
+  auto &Asm = Layout.getAssembler();
+  Asm.getWriter().recordRelocation(Asm, Layout, &F, FA, A, FixedValueA);
+  Asm.getWriter().recordRelocation(Asm, Layout, &F, FB, B, FixedValueB);
+  FixedValue = FixedValueA - FixedValueB;
+  return true;
+}
+
 std::unique_ptr<MCObjectTargetWriter>
 LoongArchAsmBackend::createObjectTargetWriter() const {
   return createLoongArchELFObjectWriter(

llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchAsmBackend.h

@@ -31,10 +31,15 @@ class LoongArchAsmBackend : public MCAsmBackend {
 public:
   LoongArchAsmBackend(const MCSubtargetInfo &STI, uint8_t OSABI, bool Is64Bit,
                       const MCTargetOptions &Options)
-      : MCAsmBackend(llvm::endianness::little), STI(STI), OSABI(OSABI),
-        Is64Bit(Is64Bit), TargetOptions(Options) {}
+      : MCAsmBackend(llvm::endianness::little,
+                     LoongArch::fixup_loongarch_relax),
+        STI(STI), OSABI(OSABI), Is64Bit(Is64Bit), TargetOptions(Options) {}
   ~LoongArchAsmBackend() override {}
 
+  bool handleAddSubRelocations(const MCAsmLayout &Layout, const MCFragment &F,
+                               const MCFixup &Fixup, const MCValue &Target,
+                               uint64_t &FixedValue) const override;
+
   void applyFixup(const MCAssembler &Asm, const MCFixup &Fixup,
                   const MCValue &Target, MutableArrayRef<char> Data,
                   uint64_t Value, bool IsResolved,

llvm/lib/Target/LoongArch/MCTargetDesc/LoongArchFixupKinds.h

@@ -106,7 +106,9 @@ enum Fixups {
   // 20-bit fixup corresponding to %gd_pc_hi20(foo) for instruction pcalau12i.
   fixup_loongarch_tls_gd_pc_hi20,
   // 20-bit fixup corresponding to %gd_hi20(foo) for instruction lu12i.w.
-  fixup_loongarch_tls_gd_hi20
+  fixup_loongarch_tls_gd_hi20,
+  // Generate an R_LARCH_RELAX which indicates the linker may relax here.
+  fixup_loongarch_relax = FirstLiteralRelocationKind + ELF::R_LARCH_RELAX
 };
 } // end namespace LoongArch
 } // end namespace llvm

llvm/test/MC/LoongArch/Misc/subsection.s

@@ -0,0 +1,38 @@
+# RUN: not llvm-mc --filetype=obj --triple=loongarch64 --mattr=-relax %s -o /dev/null 2>&1 | FileCheck %s --check-prefixes=ERR,NORELAX --implicit-check-not=error:
+## TODO: not llvm-mc --filetype=obj --triple=loongarch64 --mattr=+relax %s -o /dev/null 2>&1 | FileCheck %s --check-prefixes=ERR,RELAX --implicit-check-not=error:
+
+a:
+  nop
+b:
+  la.pcrel $t0, a
+c:
+  nop
+d:
+
+.data
+## Positive subsection numbers
+## With relaxation, report an error as c-b is not an assemble-time constant.
+# RELAX: :[[#@LINE+1]]:14: error: cannot evaluate subsection number
+.subsection c-b
+# RELAX: :[[#@LINE+1]]:14: error: cannot evaluate subsection number
+.subsection d-b
+# RELAX: :[[#@LINE+1]]:14: error: cannot evaluate subsection number
+.subsection c-a
+
+.subsection b-a
+.subsection d-c
+
+## Negative subsection numbers
+# NORELAX: :[[#@LINE+2]]:14: error: subsection number -8 is not within [0,2147483647]
+# RELAX:   :[[#@LINE+1]]:14: error: cannot evaluate subsection number
+.subsection b-c
+# NORELAX: :[[#@LINE+2]]:14: error: subsection number -12 is not within [0,2147483647]
+# RELAX:   :[[#@LINE+1]]:14: error: cannot evaluate subsection number
+.subsection b-d
+# NORELAX: :[[#@LINE+2]]:14: error: subsection number -12 is not within [0,2147483647]
+# RELAX:   :[[#@LINE+1]]:14: error: cannot evaluate subsection number
+.subsection a-c
+# ERR:     :[[#@LINE+1]]:14: error: subsection number -4 is not within [0,2147483647]
+.subsection a-b
+# ERR:     :[[#@LINE+1]]:14: error: subsection number -4 is not within [0,2147483647]
+.subsection c-d

llvm/test/MC/LoongArch/Relocations/relax-addsub.s

@@ -0,0 +1,68 @@
+# RUN: llvm-mc --filetype=obj --triple=loongarch64 --mattr=-relax %s \
+# RUN:     | llvm-readobj -r -x .data - | FileCheck %s --check-prefix=NORELAX
+# RUN: llvm-mc --filetype=obj --triple=loongarch64 --mattr=+relax %s \
+# RUN:     | llvm-readobj -r -x .data - | FileCheck %s --check-prefix=RELAX
+
+# NORELAX:       Relocations [
+# NORELAX-NEXT:    Section ({{.*}}) .rela.text {
+# NORELAX-NEXT:      0x10 R_LARCH_PCALA_HI20 .text 0x0
+# NORELAX-NEXT:      0x14 R_LARCH_PCALA_LO12 .text 0x0
+# NORELAX-NEXT:    }
+# NORELAX-NEXT:  ]
+
+# NORELAX:      Hex dump of section '.data':
+# NORELAX-NEXT: 0x00000000 04040004 00000004 00000000 0000000c
+# NORELAX-NEXT: 0x00000010 0c000c00 00000c00 00000000 00000808
+# NORELAX-NEXT: 0x00000020 00080000 00080000 00000000 00
+
+# RELAX:       Relocations [
+# RELAX-NEXT:    Section ({{.*}}) .rela.text {
+# RELAX-NEXT:      0x10 R_LARCH_PCALA_HI20 .L1 0x0
+# RELAX-NEXT:      0x14 R_LARCH_PCALA_LO12 .L1 0x0
+# RELAX-NEXT:    }
+# RELAX-NEXT:    Section ({{.*}}) .rela.data {
+# RELAX-NEXT:      0xF R_LARCH_ADD8 .L3 0x0
+# RELAX-NEXT:      0xF R_LARCH_SUB8 .L2 0x0
+# RELAX-NEXT:      0x10 R_LARCH_ADD16 .L3 0x0
+# RELAX-NEXT:      0x10 R_LARCH_SUB16 .L2 0x0
+# RELAX-NEXT:      0x12 R_LARCH_ADD32 .L3 0x0
+# RELAX-NEXT:      0x12 R_LARCH_SUB32 .L2 0x0
+# RELAX-NEXT:      0x16 R_LARCH_ADD64 .L3 0x0
+# RELAX-NEXT:      0x16 R_LARCH_SUB64 .L2 0x0
+# RELAX-NEXT:    }
+# RELAX-NEXT:  ]
+
+# RELAX:      Hex dump of section '.data':
+# RELAX-NEXT: 0x00000000 04040004 00000004 00000000 00000000
+# RELAX-NEXT: 0x00000010 00000000 00000000 00000000 00000808
+# RELAX-NEXT: 0x00000020 00080000 00080000 00000000 00
+
+.text
+.L1:
+  nop
+.L2:
+  .align 4
+.L3:
+  la.pcrel $t0, .L1
+.L4:
+  ret
+
+.data
+## Not emit relocs
+.byte  .L2 - .L1
+.short .L2 - .L1
+.word  .L2 - .L1
+.dword .L2 - .L1
+## With relaxation, emit relocs because of the .align making the diff variable.
+## TODO Handle alignment directive. Why they emit relocs now? They returns
+## without folding symbols offset in AttemptToFoldSymbolOffsetDifference().
+.byte  .L3 - .L2
+.short .L3 - .L2
+.word  .L3 - .L2
+.dword .L3 - .L2
+## TODO
+## With relaxation, emit relocs because la.pcrel is a linker-relaxable inst.
+.byte  .L4 - .L3
+.short .L4 - .L3
+.word  .L4 - .L3
+.dword .L4 - .L3