The java stuff generates sections with names like .rodata.jutf8.10 and
.rodata.jutf.14 that have the entry size (sh_entsize) set to an
integerer that isn't a power of two. When merging these sections into
the final output, lld adjusts the alignment to be at least the entry
size. However, the alignment of a section should always be a power of
2, and this is checked later on. At that point lld throws us a fatal
error.
The diff below fixes this issue by sticking to the original alignment
if the entry size isn't a power of two. Presumably lld is trying to
optimize things here by making sure that entries are word or
cache-line aligned when possible. But such an optimization doesn't
make sense if the entry size isn't a power of 2
This makes the ports gcc build on amd64 systems with lld as the default linker.
ok?
Index: gnu/llvm/tools/lld/ELF/SyntheticSections.cpp
===================================================================
RCS file: /cvs/src/gnu/llvm/tools/lld/ELF/SyntheticSections.cpp,v
retrieving revision 1.8
diff -u -p -r1.8 SyntheticSections.cpp
--- gnu/llvm/tools/lld/ELF/SyntheticSections.cpp 3 Jun 2018 12:34:03
-0000 1.8
+++ gnu/llvm/tools/lld/ELF/SyntheticSections.cpp 20 Oct 2018 13:48:46
-0000
@@ -2512,7 +2512,9 @@ void elf::mergeSections() {
continue;
StringRef OutsecName = getOutputSectionName(MS);
- uint32_t Alignment = std::max<uint32_t>(MS->Alignment, MS->Entsize);
+ uint32_t Alignment = MS->Alignment;
+ if (isPowerOf2_32(MS->Entsize))
+ Alignment = std::max<uint32_t>(Alignment, MS->Entsize);
auto I = llvm::find_if(MergeSections, [=](MergeSyntheticSection *Sec) {
// While we could create a single synthetic section for two different
