[lld] [LLD][COFF] Add basic ARM64X dynamic relocations support (PR #118035)

[Martin Storsjö via llvm-commits]

================
@@ -1147,4 +1148,85 @@ uint32_t ImportThunkChunkARM64EC::extendRanges() {
   return sizeof(arm64Thunk) - sizeof(uint32_t);
 }
 
+size_t Arm64XDynamicRelocEntry::getSize() const {
+  switch (type) {
+  case IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE:
+    return sizeof(uint16_t) + size; // A header and a payload.
+  case IMAGE_DVRT_ARM64X_FIXUP_TYPE_DELTA:
+  case IMAGE_DVRT_ARM64X_FIXUP_TYPE_ZEROFILL:
+    llvm_unreachable("unsupported type");
+  }
+}
+
+void Arm64XDynamicRelocEntry::writeTo(uint8_t *buf) const {
+  auto out = reinterpret_cast<ulittle16_t *>(buf);
+  *out = (offset & 0xfff) | (type << 12);
+
+  switch (type) {
+  case IMAGE_DVRT_ARM64X_FIXUP_TYPE_VALUE:
+    *out |= ((bit_width(size) - 1) << 14); // Encode the size.
+    switch (size) {
+    case 2:
+      out[1] = value;
+      break;
+    case 4:
+      *reinterpret_cast<ulittle32_t *>(out + 1) = value;
+      break;
+    case 8:
+      *reinterpret_cast<ulittle64_t *>(out + 1) = value;
+      break;
+    default:
+      llvm_unreachable("invalid size");
+    }
+    break;
+  case IMAGE_DVRT_ARM64X_FIXUP_TYPE_DELTA:
+  case IMAGE_DVRT_ARM64X_FIXUP_TYPE_ZEROFILL:
+    llvm_unreachable("unsupported type");
+  }
+}
+
+void DynamicRelocsChunk::finalize() {
+  llvm::stable_sort(arm64xRelocs, [=](const Arm64XDynamicRelocEntry &a,
+                                      const Arm64XDynamicRelocEntry &b) {
+    return a.offset < b.offset;
----------------
mstorsjo wrote:

This `finalize()` method gets called multiple times, if we need to add range extension thunks - while I don't think anything at least at this point within the function really needs to be recalculated after the layout gets redone. But perhaps some future dynamic relocations may need to be re-sorted? (Technically, if we're inserting thunks and have relocations based on symbols, the relative ordering of relocs probably still is the same, but we may need to recalculate offsets. At that point, it's probably fine to keep this in the part of the code that gets rerun, for clarity.)

https://github.com/llvm/llvm-project/pull/118035