On Wed, 3 Oct 2012, Jakub Jelinek wrote:
> Hi!
>
> With the PR54519 patch I've just posted, I've noticed, I've noticed on the
> same testcase from yesterday's IRC:
> static inline void foo (int x, int y) { asm volatile ("nop"); }
> static inline void bar (int z) { foo (z, 0); foo (z, 1); }
> int main ()
> {
> bar (0);
> bar (1);
> return 0;
> }
> that while I can print x and y just fine, if I do bt, x, y and z printed
> in the backtrace are all optimized out.
> The problem is that first tree versioning for foo.isra.* or bar.isra.*
> deposits the optimized away parameters as VAR_DECLs into the DECL_INITIAL
> block (which is fine), but then during inlining they end up in the remapped
> block of DECL_INITIAL, not the new block added above it into which inliner
> puts parameters. So in the debug info we have
> DW_TAG_inlined_subroutine
> DW_TAG_formal_parameter for non-optimized away parameters
> DW_TAG_lexical_block
> DW_TAG_formal_parameter for optimized away parameters
> and the debugger (expectably) looks only at DW_TAG_inlined_subroutine
> DIE's immediate children for the formal parameters to print during
> backtrace.
> Fixed up by moving the VAR_DECLs for parameters optimized away by versioning
> to BLOCK_SUPERCONTEXT during inlining, at that point we know both of the
> blocks have the same scope, and if the original DECL_INITIAL doesn't contain
> any other vars, we can actually end up with shorter/more correct debug info
> as well as memory savings due to being able to GC the remapped DECL_INITIAL
> block.
>
> Bootstrapped/regtested on x86_64-linux and i686-linux, ok for trunk?
This looks like the wrong place to fix things to me ... either we can
fix this at the point we create the VAR_DECLs for the optimized away
PARM_DECLs (or we should delay that until here?) or we fix it up
in dwarf2out.c (how does this fix interact with stabs and the other
debuginfo formats? mentioning DWARF in tree-inline looks odd,
unless we get rid of the other formats - something I'd of course
welcome ;))
Thanks,
Richard.
> 2012-10-03 Jakub Jelinek <ja...@redhat.com>
>
> * tree-inline.c (expand_call_inline): Move VAR_DECLs with
> PARM_DECL origins from remapped DECL_INITIAL's BLOCK_VARS
> into id->block's BLOCK_VARS.
>
> --- gcc/tree-inline.c.jj 2012-10-02 17:43:13.000000000 +0200
> +++ gcc/tree-inline.c 2012-10-02 19:43:52.576382413 +0200
> @@ -3946,7 +3946,29 @@ expand_call_inline (basic_block bb, gimp
> initialize_inlined_parameters (id, stmt, fn, bb);
>
> if (DECL_INITIAL (fn))
> - prepend_lexical_block (id->block, remap_blocks (DECL_INITIAL (fn), id));
> + {
> + tree *var;
> +
> + prepend_lexical_block (id->block, remap_blocks (DECL_INITIAL (fn),
> id));
> + gcc_checking_assert (BLOCK_SUBBLOCKS (id->block)
> + && (BLOCK_CHAIN (BLOCK_SUBBLOCKS (id->block))
> + == NULL_TREE));
> + /* Move vars for PARM_DECLs from DECL_INITIAL block to id->block,
> + otherwise DW_TAG_formal_parameter will not be children of
> + DW_TAG_inlined_subroutine, but of a DW_TAG_lexical_block
> + under it. The parameters can be then evaluated in the debugger,
> + but don't show in backtraces. */
> + for (var = &BLOCK_VARS (BLOCK_SUBBLOCKS (id->block)); *var; )
> + if (TREE_CODE (DECL_ORIGIN (*var)) == PARM_DECL)
> + {
> + tree v = *var;
> + *var = TREE_CHAIN (v);
> + TREE_CHAIN (v) = BLOCK_VARS (id->block);
> + BLOCK_VARS (id->block) = v;
> + }
> + else
> + var = &TREE_CHAIN (*var);
> + }
>
> /* Return statements in the function body will be replaced by jumps
> to the RET_LABEL. */
>
>
> Jakub
>
>
--
Richard Biener <rguent...@suse.de>
SUSE / SUSE Labs
SUSE LINUX Products GmbH - Nuernberg - AG Nuernberg - HRB 16746
GF: Jeff Hawn, Jennifer Guild, Felix Imend
