DWARF: Compensate for compiler's off-by-one instruction.
The compiler generates stackmaps with PC value which is *after*
the instruction rather PC of the instruction itself. This causes
trouble when generating native line number mapping since the
branch instruction itself will not be mapped, and consequently
we get incorrect line number in backtraces. Add code to compensate
for this. See the in-code comment for more details.
Change-Id: I72c992e6d08a767f314290d562421b251ae60732
diff --git a/compiler/debug/elf_debug_line_writer.h b/compiler/debug/elf_debug_line_writer.h
index ed26d96..0d3673d 100644
--- a/compiler/debug/elf_debug_line_writer.h
+++ b/compiler/debug/elf_debug_line_writer.h
@@ -114,6 +114,40 @@
continue;
}
+ // Compensate for compiler's off-by-one-instruction error.
+ //
+ // The compiler generates stackmap with PC *after* the branch instruction
+ // (because this is the PC which is easier to obtain when unwinding).
+ //
+ // However, the debugger is more clever and it will ask us for line-number
+ // mapping at the location of the branch instruction (since the following
+ // instruction could belong to other line, this is the correct thing to do).
+ //
+ // So we really want to just decrement the PC by one instruction so that the
+ // branch instruction is covered as well. However, we do not know the size
+ // of the previous instruction, and we can not subtract just a fixed amount
+ // (the debugger would trust us that the PC is valid; it might try to set
+ // breakpoint there at some point, and setting breakpoint in mid-instruction
+ // would make the process crash in spectacular way).
+ //
+ // Therefore, we say that the PC which the compiler gave us for the stackmap
+ // is the end of its associated address range, and we use the PC from the
+ // previous stack map as the start of the range. This ensures that the PC is
+ // valid and that the branch instruction is covered.
+ //
+ // This ensures we have correct line number mapping at call sites (which is
+ // important for backtraces), but there is nothing we can do for non-call
+ // sites (so stepping through optimized code in debugger is not possible).
+ //
+ // We do not adjust the stackmaps if the code was compiled as debuggable.
+ // In that case, the stackmaps should accurately cover all instructions.
+ if (!mi->is_native_debuggable) {
+ for (size_t i = pc2dex_map.size() - 1; i > 0; --i) {
+ pc2dex_map[i].from_ = pc2dex_map[i - 1].from_;
+ }
+ pc2dex_map[0].from_ = 0;
+ }
+
Elf_Addr method_address = base_address + mi->code_address;
PositionInfos dex2line_map;