Remove __llvm_mipmap section in llvm-strip

[ellis]

Summary

The __llvm_mipmap section should be considered a debug section because it provides no use at runtime. The llvm-strip tool should remove that section.

[alexander-shaposhnikov]

I'm wondering if you can use llvm-objcopy --remove-section for these purposes instead?

[ellis]

In D104164#2815489, @alexshap wrote:

I'm wondering if you can use llvm-objcopy --remove-section for these purposes instead?

Yes, we can use llvm-objcopy --remove-section=__llvm_mipmap a.out. I like this patch because it allows you to use llvm-strip easily, but it technically isn't required.

[alexander-shaposhnikov]

my 0.02$ - (the diff lacks tests, isDebugSection has a more specific meaning in this context - but these are minor concerns), the main concern - I'm not sure that putting this logic into llvm-objcopy is a good idea, there are many tools which create extra sections and making llvm-objcopy aware of them doesn't feel right. Btw - llvm-strip has --remove-section as well, so e.g. you can simply add it to the same invocation of the tool, no need to run it as a separate build step. But I'd like to hear what others think, cc @jhenderson, @MaskRay

[MaskRay]

__llvm_mipmap is not a debug section, so not suitable in isDebugSection.

--strip-debug/--strip-nonalloc/--strip-all-gnu cannot remove SHF_ALLOC sections. The proposed __llvm_mipmap has the SHF_ALLOC flag so not suitable.

Every SHF_ALLOC section is part of the program image. A tool like llvm-strip needs to be conservative and assumes removal of every such section can affect runtime behaviors and probably correctness.
(A self-introspection program can dump its own non-SHF_ALLOC section; this sentence does not discuss such programs.)
Sometimes there are indeed special SHF_ALLOC sections which are strippable, users can remove them with --remove-section. It is not strip's business to understand every such section.

[ellis]

In D104164#2815798, @MaskRay wrote:

__llvm_mipmap is not a debug section, so not suitable in isDebugSection.

--strip-debug/--strip-nonalloc/--strip-all-gnu cannot remove SHF_ALLOC sections. The proposed __llvm_mipmap has the SHF_ALLOC flag so not suitable.

Every SHF_ALLOC section is part of the program image. A tool like llvm-strip needs to be conservative and assumes removal of every such section can affect runtime behaviors and probably correctness.
(A self-introspection program can dump its own non-SHF_ALLOC section; this sentence does not discuss such programs.)
Sometimes there are indeed special SHF_ALLOC sections which are strippable, users can remove them with --remove-section. It is not strip's business to understand every such section.

I have no problem using --remove-section so I'll update this stack to use that. Thanks for the feedback!

llvm/tools/llvm-objcopy/ELF/ELFObjcopy.cpp

 namespace elf {
 
 using namespace object;
 using namespace ELF;
 using SectionPred = std::function<bool(const SectionBase &Sec)>;
 
 static bool isDebugSection(const SectionBase &Sec) {
   return StringRef(Sec.Name).startswith(".debug") ||
-         StringRef(Sec.Name).startswith(".zdebug") || Sec.Name == ".gdb_index";
+         StringRef(Sec.Name).startswith(".zdebug") ||
+         Sec.Name == ".gdb_index" || Sec.Name == "__llvm_mipmap";
 }
 
 static bool isDWOSection(const SectionBase &Sec) {
   return StringRef(Sec.Name).endswith(".dwo");
 }
 
 static bool onlyKeepDWOPred(const Object &Obj, const SectionBase &Sec) {
   // We can't remove the section header string table.