https://gcc.gnu.org/bugzilla/show_bug.cgi?id=85961
Bug ID: 85961 Summary: scratch register rsi used after function call Product: gcc Version: 7.3.0 Status: UNCONFIRMED Severity: normal Priority: P3 Component: c++ Assignee: unassigned at gcc dot gnu.org Reporter: bki at hacon dot de Target Milestone: --- Created attachment 44203 --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=44203&action=edit Minimal cpp-File producing the bug With g++ 7.0.0 and g++ 7.3.0 and compiling the attached file as detailed in the comments, produces an object file containing the following instruction sequence: 76: 48 89 c6 mov %rax,%rsi 79: e8 82 ff ff ff callq 0 <_ZNKSt7__cxx1112basic_stringIcSt11char_traitsIcESaIcEE6lengthEv.isra.0> 7e: 48 8d 14 06 lea (%rsi,%rax,1),%rdx According to the System V AMD64 ABI (https://software.intel.com/sites/default/files/article/402129/mpx-linux64-abi.pdf), %rsi is a scratch register not preserved across function calls. I therefore think that this instruction sequence is wrong. Why is this a problem? The function which is called does not clobber %rsi itself. But compiling with '-pg -mfentry' will instrument that function call and allow a user specified function to clobber %rsi. A call to __fentry__ should not clobber %rsi, since %rsi is used for parameter passing and must be unchanged when __fentry__ returns. But by temporarily replacing the return address __fentry__ can instrument the function exit. There, clobbering %rsi (and all other scratch registers) should be acceptable.