Re: [PATCH 0/6] Use correct function address in dwarf assembler

[Doug Evans <dje at google dot com>]

Yao Qi writes:
 > When we use dwarf assembler, it is common to generate DW_AT_low_pc and
 > DW_AT_high_pc attributes like this:
 > 
 >  	    subprogram {
 > 		{name main}
 > 		{low_pc main addr}
 > 		{high_pc main+0x100 addr}
 >             }
 > 
 > however, main is not resolved to function main's address on some
 > targets such as arm thumb mode and powerpc64.  One approach is to
 > get function address by labels, like this:
 > 
 >   asm ("func_start: .globl func_start");
 >   static void func (void) {}
 >   asm ("func_end: .globl func_end");
 > 
 > however, some compiler, such as clang, can't guarantee the order of
 > these labels and function, so it isn't portable.
 > 
 > In this patch series, we propose a new approach 
 > 
 >  1. get function start and end address correctly in a portable way,
 >  2. don't affect too much on existing test cases.
 > 
 > In patch 2/6, a new proc function_range is added.  In this proc, the
 > source file is compiled with debug info.
 > 
 >   int main (void)
 >   {
 >     asm ("main_label: .globl main_label");
 >     return 0;
 >   }
 > 
 > we can get the offset of main_label in function main, and then compute
 > the start address of main via main_label - main_label_offset.  This is
 > portable.
 > 
 > In order to avoid much changes to existing test cases, we have to use
 > function_range inside dwarf assembler.  Then, I invent some attribute
 > macros which look like DW attributes, but can be expanded to one or
 > more standard attributes.  The TAG_subprogram above will be simplified
 > with macro attribute used, it becomes:
 > 
 >     subprogram {
 > 	{MACRO_AT_func { func ${srcdir}/${subdir}/${srcfile} }}
 >     }
 > 
 > With this macro attribute, attributes name, low_pc and high_pc are
 > generated and set correctly.  See more details in patch 2/6.
 > 
 > Patch 4/6 is to use dwarf assembler to generate .S file, which is
 > identical to .S current one we are using.  Patch 3 and 5 is to use
 > macro attributes to get function start and end address correctly.
 > Patch 6 is remove labels and switch to use macro attribute to
 > get function start address to make clang happy.
 > 
 > I test the whole series by running tests under gdb.dwarf2/ on
 > powerpc-linux (both 32-bit and 64-bit), arm-none-eabi (arm mode
 > and thumb mode) and x86_64-linux.
 > 
 > Comments are very welcome!

Hi.
In general I like this a lot.
Just a few outstanding questions mentioned in replies to
the relevant patches.