lld: Target-specific section-folding logic for ICF

lld/ELF/Arch/AArch64.cpp

@@ -8,10 +8,13 @@
 
 #include "InputFiles.h"
 #include "OutputSections.h"
+#include "Relocations.h"
 #include "Symbols.h"
 #include "SyntheticSections.h"
 #include "Target.h"
 #include "lld/Common/ErrorHandler.h"
+#include "llvm/ADT/SmallSet.h"
+#include "llvm/ADT/SmallVector.h"
 #include "llvm/BinaryFormat/ELF.h"
 #include "llvm/Support/Endian.h"
 
@@ -83,6 +86,7 @@ class AArch64 : public TargetInfo {
                 uint64_t val) const override;
   RelExpr adjustTlsExpr(RelType type, RelExpr expr) const override;
   void relocateAlloc(InputSectionBase &sec, uint8_t *buf) const override;
+  bool canFoldSection(const SmallVector<Relocation> &relocs) const override;
 
 private:
   void relaxTlsGdToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const;
@@ -975,6 +979,20 @@ void AArch64::relocateAlloc(InputSectionBase &sec, uint8_t *buf) const {
   }
 }
 
+bool AArch64::canFoldSection(const SmallVector<Relocation> &relocs) const {
+  // Section cannot be folded as part of ICF if it contains unpaired relocations
+  // eg: ADR_GOT_PAGE and LD64_GOT_LO12_NC don't point to same symbol.
+  SmallSet<Symbol*, 4> syms;
+  for (const Relocation &reloc : relocs) {
+    if (reloc.type == R_AARCH64_ADR_GOT_PAGE)
+      syms.insert(reloc.sym);
+    else if (reloc.type == R_AARCH64_LD64_GOT_LO12_NC && !syms.contains(reloc.sym))
+      return false;
+  }
+
+  return true;
+}
+
 // AArch64 may use security features in variant PLT sequences. These are:
 // Pointer Authentication (PAC), introduced in armv8.3-a and Branch Target
 // Indicator (BTI) introduced in armv8.5-a. The additional instructions used

lld/ELF/ICF.cpp

@@ -77,9 +77,11 @@
 #include "InputFiles.h"
 #include "LinkerScript.h"
 #include "OutputSections.h"
+#include "Relocations.h"
 #include "SymbolTable.h"
 #include "Symbols.h"
 #include "SyntheticSections.h"
+#include "Target.h"
 #include "llvm/BinaryFormat/ELF.h"
 #include "llvm/Object/ELF.h"
 #include "llvm/Support/Parallel.h"
@@ -239,6 +241,7 @@ template <class ELFT>
 template <class RelTy>
 bool ICF<ELFT>::constantEq(const InputSection *secA, Relocs<RelTy> ra,
                            const InputSection *secB, Relocs<RelTy> rb) {
+  SmallVector<Relocation> relocsA, relocsB;
   if (ra.size() != rb.size())
     return false;
   auto rai = ra.begin(), rae = ra.end(), rbi = rb.begin();
@@ -247,11 +250,21 @@ bool ICF<ELFT>::constantEq(const InputSection *secA, Relocs<RelTy> ra,
         rai->getType(ctx.arg.isMips64EL) != rbi->getType(ctx.arg.isMips64EL))
       return false;
 
-    uint64_t addA = getAddend<ELFT>(*rai);
-    uint64_t addB = getAddend<ELFT>(*rbi);
+    int64_t addA = getAddend<ELFT>(*rai);
+    int64_t addB = getAddend<ELFT>(*rbi);
 
     Symbol &sa = secA->file->getRelocTargetSym(*rai);
     Symbol &sb = secB->file->getRelocTargetSym(*rbi);
+
+    // We only need relocation type and relocation target symbol. Offset is
+    // not used (as of now). So we don't bother populating one.
+    relocsA.push_back(
+        {ctx.target->getRelExpr(rai->getType(ctx.arg.isMips64EL), sa, 0),
+         rai->getType(ctx.arg.isMips64EL), 0, addA, &sa});
+    relocsB.push_back(
+        {ctx.target->getRelExpr(rbi->getType(ctx.arg.isMips64EL), sb, 0),
+         rbi->getType(ctx.arg.isMips64EL), 0, addB, &sb});
+
     if (&sa == &sb) {
       if (addA == addB)
         continue;
@@ -308,6 +321,10 @@ bool ICF<ELFT>::constantEq(const InputSection *secA, Relocs<RelTy> ra,
       return false;
   }
 
+  // Target-specific section-folding logic based on section's relocation list
+  if (!ctx.target->canFoldSection(relocsA) || !ctx.target->canFoldSection(relocsB))
+    return false;
+
   return true;
 }
 

lld/ELF/Target.h

@@ -102,6 +102,9 @@ class TargetInfo {
   virtual void applyJumpInstrMod(uint8_t *loc, JumpModType type,
                                  JumpModType val) const {}
 
+  // Used by ICF to determine if section full of @relocs can be folded safely with another
+  virtual bool canFoldSection(const SmallVector<Relocation> &relocs) const { return true; }
+
   virtual ~TargetInfo();
 
   // This deletes a jump insn at the end of the section if it is a fall thru to

lld/test/ELF/aarch64-icf-unpaired.s

@@ -0,0 +1,48 @@
+// REQUIRES: aarch64
+
+# RUN: llvm-mc -filetype=obj -triple=aarch64 %s -o %t
+# RUN: ld.lld %t -o %t2 --icf=all --print-icf-sections | FileCheck %s
+
+# CHECK-NOT: selected section {{.*}}(.text.f2_0)
+# CHECK-NOT:   removing identical section {{.*}}(.text.f2_1)
+# CHECK-NOT:   removing identical section {{.*}}(.text.f2_2)
+
+.addrsig
+
+callee:
+ret
+
+.macro f, index
+
+.section .text.f1_\index,"ax",@progbits
+f1_\index:
+adrp x0, :got:g\index
+mov x1, #\index
+b f2_\index
+
+.section .text.f2_\index,"ax",@progbits
+f2_\index:
+ldr x0, [x0, :got_lo12:g\index] 
+b callee
+
+.globl g\index
+.section .rodata.g\index,"a",@progbits
+g_\index:
+.long 111
+.long 122
+
+g\index:
+.byte 123
+
+.section .text._start,"ax",@progbits
+bl f1_\index
+
+.endm
+
+.section .text._start,"ax",@progbits
+.globl _start
+_start:
+
+f 0
+f 1
+f 2