[lld] [lld][ELF] Support relax R_LARCH_ALIGN (PR #78692)

[Jinyang He via llvm-commits]

================
@@ -659,6 +663,156 @@ void LoongArch::relocate(uint8_t *loc, const Relocation &rel,
   }
 }
 
+static bool relax(InputSection &sec) {
+  const uint64_t secAddr = sec.getVA();
+  const MutableArrayRef<Relocation> relocs = sec.relocs();
+  auto &aux = *sec.relaxAux;
+  bool changed = false;
+  ArrayRef<SymbolAnchor> sa = ArrayRef(aux.anchors);
+  uint64_t delta = 0;
+
+  std::fill_n(aux.relocTypes.get(), relocs.size(), R_LARCH_NONE);
+  aux.writes.clear();
+  for (auto [i, r] : llvm::enumerate(relocs)) {
+    const uint64_t loc = secAddr + r.offset - delta;
+    uint32_t &cur = aux.relocDeltas[i], remove = 0;
+    switch (r.type) {
+    case R_LARCH_ALIGN: {
+      const uint64_t addend =
+          r.sym->isUndefined() ? Log2_64(r.addend) + 1 : r.addend;
+      const uint64_t allBytes = (1 << (addend & 0xff)) - 4;
+      const uint64_t align = 1 << (addend & 0xff);
+      const uint64_t maxBytes = addend >> 8;
+      const uint64_t off = loc & (align - 1);
+      const uint64_t curBytes = off == 0 ? 0 : align - off;
+      // All bytes beyond the alignment boundary should be removed.
+      // If emit bytes more than max bytes to emit, remove all.
+      if (maxBytes != 0 && curBytes > maxBytes)
+        remove = allBytes;
+      else
+        remove = allBytes - curBytes;
+      assert(static_cast<int32_t>(remove) >= 0 &&
+             "R_LARCH_ALIGN needs expanding the content");
+      break;
+    }
+    }
+
+    // For all anchors whose offsets are <= r.offset, they are preceded by
+    // the previous relocation whose `relocDeltas` value equals `delta`.
+    // Decrease their st_value and update their st_size.
+    for (; sa.size() && sa[0].offset <= r.offset; sa = sa.slice(1)) {
+      if (sa[0].end)
+        sa[0].d->size = sa[0].offset - delta - sa[0].d->value;
+      else
+        sa[0].d->value = sa[0].offset - delta;
+    }
+    delta += remove;
+    if (delta != cur) {
+      cur = delta;
+      changed = true;
+    }
+  }
+
+  for (const SymbolAnchor &a : sa) {
+    if (a.end)
+      a.d->size = a.offset - delta - a.d->value;
+    else
+      a.d->value = a.offset - delta;
+  }
+  // Inform assignAddresses that the size has changed.
+  if (!isUInt<32>(delta))
+    fatal("section size decrease is too large: " + Twine(delta));
+  sec.bytesDropped = delta;
+  return changed;
+}
+
+// When relaxing just R_LARCH_ALIGN, relocDeltas is usually changed only once in
+// the absence of a linker script. For call and load/store R_LARCH_RELAX, code
+// shrinkage may reduce displacement and make more relocations eligible for
+// relaxation. Code shrinkage may increase displacement to a call/load/store
+// target at a higher fixed address, invalidating an earlier relaxation. Any
+// change in section sizes can have cascading effect and require another
+// relaxation pass.
+bool LoongArch::relaxOnce(int pass) const {
+  if (config->relocatable)
+    return false;
+
+  if (pass == 0)
+    initSymbolAnchors();
+
+  SmallVector<InputSection *, 0> storage;
+  bool changed = false;
+  for (OutputSection *osec : outputSections) {
+    if (!(osec->flags & SHF_EXECINSTR))
+      continue;
+    for (InputSection *sec : getInputSections(*osec, storage))
+      changed |= relax(*sec);
+  }
+  return changed;
+}
+
+void LoongArch::finalizeRelax(int passes) const {
+  SmallVector<InputSection *, 0> storage;
+  for (OutputSection *osec : outputSections) {
+    if (!(osec->flags & SHF_EXECINSTR))
+      continue;
+    for (InputSection *sec : getInputSections(*osec, storage)) {
+      RelaxAux &aux = *sec->relaxAux;
+      if (!aux.relocDeltas)
+        continue;
+
+      MutableArrayRef<Relocation> rels = sec->relocs();
+      ArrayRef<uint8_t> old = sec->content();
+      size_t newSize = old.size() - aux.relocDeltas[rels.size() - 1];
+      uint8_t *p = context().bAlloc.Allocate<uint8_t>(newSize);
+      uint64_t offset = 0;
+      int64_t delta = 0;
+      sec->content_ = p;
+      sec->size = newSize;
+      sec->bytesDropped = 0;
+
+      // Update section content: remove NOPs for R_LARCH_ALIGN and rewrite
+      // instructions for relaxed relocations.
+      for (size_t i = 0, e = rels.size(); i != e; ++i) {
+        uint32_t remove = aux.relocDeltas[i] - delta;
+        delta = aux.relocDeltas[i];
+        if (remove == 0 && aux.relocTypes[i] == R_LARCH_NONE)
+          continue;
+
+        // Copy from last location to the current relocated location.
+        const Relocation &r = rels[i];
+        uint64_t size = r.offset - offset;
+
+        // The rel.type may be fixed to R_LARCH_RELAX and the instruction
+        // will be deleted, then cause size<0
+        if (remove == 0 && (int64_t)size < 0) {
+          assert(aux.relocTypes[i] == R_LARCH_RELAX && "Unexpected size");
+          continue;
+        }
+
+        memcpy(p, old.data() + offset, size);
----------------
MQ-mengqing wrote:

The first memcpy is copy from the last offset to current reloc offset. The second memcpy is copy from final offset the to the section end.


https://github.com/llvm/llvm-project/pull/78692