Merge .xdata into .rdata by default

[rnk]

Summary

MSVC does this. The user can override it with their own /merge: flag.

[rnk]

Do you think we should merge .idata, .edata, and .didata into .rdata as well? That will be more involved than this.

[smeenai]

What are the advantages to the section merging, apart from matching link's ouput? I personally prefer LLD keeping the separate sections, since it's easier to examine them that way.

[rnk]

In D40197#929200, @smeenai wrote:

What are the advantages to the section merging, apart from matching link's ouput? I personally prefer LLD keeping the separate sections, since it's easier to examine them that way.

Saves alignment padding, mostly. All of .xdata, .edata, .idata, and .didata can be dumped by llvm-readobj by starting from the PE headers, so they aren't that hard to read without section headers. The alignment padding is particularly bad when sections have different protections so we have to pad out to 4096 to get to a page boundary, as in .data vs. .didata, which are typically near each other because we sort output sections alphabetically.

[smeenai]

Ah all right, that makes sense then.

[ruiu]

LGTM

[rnk]

This caused EH-related test failures on 32-bit LLD bots: https://ci.chromium.org/buildbot/chromium.clang/CrWinClangLLD/166

64-bit code was fine. I'm reverting and will investigate later.

lld/COFF/Driver.cpp

void LinkerDriver::link(ArrayRef<const char *> ArgsArr) {
   // Handle /failifmismatch
   for (auto *Arg : Args.filtered(OPT_failifmismatch))
     checkFailIfMismatch(Arg->getValue());
 
   // Handle /merge
   for (auto *Arg : Args.filtered(OPT_merge))
     parseMerge(Arg->getValue());
 
+  // Add default section merging rules after user rules. User rules take
+  // precedence, but we will emit a warning if there is a conflict.
+  parseMerge(".xdata=.rdata");
+
   // Handle /section
   for (auto *Arg : Args.filtered(OPT_section))
     parseSection(Arg->getValue());
 
   // Handle /aligncomm
   for (auto *Arg : Args.filtered(OPT_aligncomm))
     parseAligncomm(Arg->getValue());
 

lld/test/COFF/icf-xdata.s

 # RUN: llvm-mc %s -triple x86_64-windows-msvc -filetype=obj -o %t.obj
+# RUN: lld-link %t.obj -dll -noentry -out:%t.dll -merge:.xdata=.xdata 2>&1 \
+# RUN:     | FileCheck %s --check-prefix=WARN
+# RUN: llvm-readobj -sections %t.dll | FileCheck %s --check-prefix=XDATA
 # RUN: lld-link %t.obj -dll -noentry -out:%t.dll
-# RUN: llvm-readobj -sections %t.dll | FileCheck %s
+# RUN: llvm-readobj -sections %t.dll | FileCheck %s --check-prefix=RDATA
 
 # There shouldn't be much xdata, because all three .pdata entries (12 bytes
 # each) should use the same .xdata unwind info.
-# CHECK:         Name: .pdata
-# CHECK-NEXT:    VirtualSize: 0x24
-# CHECK:         Name: .xdata
-# CHECK-NEXT:    VirtualSize: 0x8
+# XDATA:         Name: .pdata
+# XDATA-NEXT:    VirtualSize: 0x24
+# XDATA:         Name: .xdata
+# XDATA-NEXT:    VirtualSize: 0x8
+#
+# WARN: warning: .xdata=.rdata: already merged into .xdata
+#
+# RDATA:         Name: .pdata
+# RDATA-NEXT:    VirtualSize: 0x24
+# RDATA:         Name: .rdata
+# RDATA-NEXT:    VirtualSize: 0x8
 
   .text
 callee:
   ret
 
   .def   xdata1;
   .scl  2;
   .type 32;