[LLD][COFF] add __{data,bss}_{start,end}__ symbols for Cygwin support (llvm#136180)

jeremyd2019 · GeorgeARM · commit ed0344cfa1d3 · 2025-05-07T16:30:01.000+01:00
Cygwin requires these symbols for its fork emulation to know what data
to copy into the child. GNU ld defines these symbols for MinGW targets
also, so do the same here.

Cygwin also has the `.data_cygwin_nocopy` section, which is merged into
`.data` outside the `__data_start__` to `__data_end__` range. This
excludes it from fork's copying. AFAIK it's only used by the Cygwin DLL
itself (which requires a custom linker script to link, that's not
supported by LLD), but the section is included in GNU ld's default
linker script so handle it here too.

Signed-off-by: Jeremy Drake &lt;github@jdrake.com&gt;
diff --git a/lld/COFF/Driver.cpp b/lld/COFF/Driver.cpp
@@ -2028,6 +2028,7 @@ void LinkerDriver::linkerMain(ArrayRef<const char *> argsArr) {
     parseMerge(".ctors=.rdata");
     parseMerge(".dtors=.rdata");
     parseMerge(".CRT=.rdata");
+    parseMerge(".data_cygwin_nocopy=.data");
   }
 
   // Handle /section
@@ -2484,6 +2485,10 @@ void LinkerDriver::linkerMain(ArrayRef<const char *> argsArr) {
     if (config->mingw) {
       symtab.addAbsolute(symtab.mangle("__CTOR_LIST__"), 0);
       symtab.addAbsolute(symtab.mangle("__DTOR_LIST__"), 0);
+      symtab.addAbsolute("__data_start__", 0);
+      symtab.addAbsolute("__data_end__", 0);
+      symtab.addAbsolute("__bss_start__", 0);
+      symtab.addAbsolute("__bss_end__", 0);
     }
     if (config->debug || config->buildIDHash != BuildIDHash::None)
       if (symtab.findUnderscore("__buildid"))
diff --git a/lld/COFF/Writer.cpp b/lld/COFF/Writer.cpp
@@ -240,6 +240,7 @@ class Writer {
   void createRuntimePseudoRelocs();
   void createECChunks();
   void insertCtorDtorSymbols();
+  void insertBssDataStartEndSymbols();
   void markSymbolsWithRelocations(ObjFile *file, SymbolRVASet &usedSymbols);
   void createGuardCFTables();
   void markSymbolsForRVATable(ObjFile *file,
@@ -315,6 +316,7 @@ class Writer {
 
   OutputSection *textSec;
   OutputSection *hexpthkSec;
+  OutputSection *bssSec;
   OutputSection *rdataSec;
   OutputSection *buildidSec;
   OutputSection *dataSec;
@@ -1077,7 +1079,7 @@ void Writer::createSections() {
   textSec = createSection(".text", code | r | x);
   if (isArm64EC(ctx.config.machine))
     hexpthkSec = createSection(".hexpthk", code | r | x);
-  createSection(".bss", bss | r | w);
+  bssSec = createSection(".bss", bss | r | w);
   rdataSec = createSection(".rdata", data | r);
   buildidSec = createSection(".buildid", data | r);
   dataSec = createSection(".data", data | r | w);
@@ -1260,8 +1262,10 @@ void Writer::createMiscChunks() {
   if (config->autoImport)
     createRuntimePseudoRelocs();
 
-  if (config->mingw)
+  if (config->mingw) {
     insertCtorDtorSymbols();
+    insertBssDataStartEndSymbols();
+  }
 }
 
 // Create .idata section for the DLL-imported symbol table.
@@ -2381,6 +2385,31 @@ void Writer::insertCtorDtorSymbols() {
   }
 }
 
+// MinGW (really, Cygwin) specific.
+// The Cygwin startup code uses __data_start__ __data_end__ __bss_start__
+// and __bss_end__ to know what to copy during fork emulation.
+void Writer::insertBssDataStartEndSymbols() {
+  if (!dataSec->chunks.empty()) {
+    Symbol *dataStartSym = ctx.symtab.find("__data_start__");
+    Symbol *dataEndSym = ctx.symtab.find("__data_end__");
+    Chunk *endChunk = dataSec->chunks.back();
+    replaceSymbol<DefinedSynthetic>(dataStartSym, dataStartSym->getName(),
+                                    dataSec->chunks.front());
+    replaceSymbol<DefinedSynthetic>(dataEndSym, dataEndSym->getName(), endChunk,
+                                    endChunk->getSize());
+  }
+
+  if (!bssSec->chunks.empty()) {
+    Symbol *bssStartSym = ctx.symtab.find("__bss_start__");
+    Symbol *bssEndSym = ctx.symtab.find("__bss_end__");
+    Chunk *endChunk = bssSec->chunks.back();
+    replaceSymbol<DefinedSynthetic>(bssStartSym, bssStartSym->getName(),
+                                    bssSec->chunks.front());
+    replaceSymbol<DefinedSynthetic>(bssEndSym, bssEndSym->getName(), endChunk,
+                                    endChunk->getSize());
+  }
+}
+
 // Handles /section options to allow users to overwrite
 // section attributes.
 void Writer::setSectionPermissions() {
diff --git a/lld/test/COFF/cygwin-symbols.s b/lld/test/COFF/cygwin-symbols.s
@@ -0,0 +1,99 @@
+# REQUIRES: x86
+# RUN: split-file %s %t.dir && cd %t.dir
+
+# RUN: llvm-mc -triple=x86_64-windows-cygnus -filetype=obj -o data-no-bss.obj data-no-bss.s
+# RUN: lld-link -lldmingw -entry:main data-no-bss.obj -out:data-no-bss.exe
+# RUN: llvm-objdump -s data-no-bss.exe | FileCheck --check-prefix=DATANOBSS %s
+
+# RUN: llvm-mc -triple=x86_64-windows-cygnus -filetype=obj -o bss-no-data.obj bss-no-data.s
+# RUN: lld-link -lldmingw -entry:main bss-no-data.obj -out:bss-no-data.exe
+# RUN: llvm-objdump -s bss-no-data.exe | FileCheck --check-prefix=BSSNODATA %s
+
+# RUN: llvm-mc -triple=x86_64-windows-cygnus -filetype=obj -o data-and-bss.obj data-and-bss.s
+# RUN: lld-link -lldmingw -entry:main data-and-bss.obj -out:data-and-bss.exe
+# RUN: llvm-objdump -s data-and-bss.exe | FileCheck --check-prefix=DATAANDBSS %s
+
+#--- data-no-bss.s
+.globl main
+main:
+  nop
+
+.data
+  .quad 1
+  .byte 2
+
+.section .data_cygwin_nocopy, "w"
+  .align 4
+  .quad 3
+  .byte 4
+
+.section .test, "w"
+  .quad __data_start__
+  .quad __data_end__
+  .quad __bss_start__
+  .quad __bss_end__
+
+#--- bss-no-data.s
+.globl main
+main:
+  nop
+
+.bss
+  .zero 8192
+
+.section .test, "w"
+  .quad __data_start__
+  .quad __data_end__
+  .quad __bss_start__
+  .quad __bss_end__
+
+#--- data-and-bss.s
+.globl main
+main:
+  nop
+
+.data
+  .quad 1
+  .byte 2
+
+.section .data_cygwin_nocopy, "w"
+  .align 4
+  .quad 3
+  .byte 4
+
+.bss
+  .zero 8192
+
+.section .test, "w"
+  .quad __data_start__
+  .quad __data_end__
+  .quad __bss_start__
+  .quad __bss_end__
+
+# DATANOBSS:      Contents of section .data:
+# DATANOBSS-NEXT: 140003000 01000000 00000000 02000000 03000000
+# DATANOBSS-NEXT: 140003010 00000000 04
+# __data_start__ pointing at 0x140003000 and
+# __data_end__   pointing at 0x140003009.
+# DATANOBSS-NEXT: Contents of section .test:
+# DATANOBSS-NEXT: 140004000 00300040 01000000 09300040 01000000
+# DATANOBSS-NEXT: 140004010 18300040 01000000 18300040 01000000
+
+# __bss_start__ pointing at 0x140003000 and
+# __bss_end__   pointing at 0x140005000.
+# BSSNODATA-NOT:  Contents of section .data:
+# BSSNODATA:      Contents of section .test:
+# BSSNODATA-NEXT: 140005000 00300040 01000000 00300040 01000000
+# BSSNODATA-NEXT: 140005010 00300040 01000000 00500040 01000000
+
+# DATAANDBSS:      Contents of section .data:
+# DATAANDBSS-NEXT: 140003000 01000000 00000000 02000000 03000000
+# DATAANDBSS-NEXT: 140003010 00000000 04000000 00000000 00000000
+# __data_start__ pointing at 0x140003000 and
+# __data_end__   pointing at 0x140003009.
+# __bss_start__ pointing at 0x140003018 and
+# __bss_end__   pointing at 0x140005018.
+# DATAANDBSS:      1400031f0 00000000 00000000 00000000 00000000
+# DATAANDBSS-NEXT: Contents of section .test:
+# DATAANDBSS-NEXT: 140006000 00300040 01000000 09300040 01000000
+# DATAANDBSS-NEXT: 140006010 18300040 01000000 18500040 01000000
