Re: Ada bootstrap failure due to loop optimization and builtin_stack_alloc

[Richard Henderson <rth at redhat dot com>]

On Tue, Aug 03, 2004 at 03:16:25PM -0400, Richard Kenner wrote:
>     Yes, absolutely we can.  Patch in testing.
> 
> Can you let me know how?  Ada has some types that this should also be
> set on (unrelated to this issue).

I've created a DECL_VALUE_EXPR that may be set for a decl.  After
this has been set, the original decl should never be referenced,
and the value of the decl can be computed from the given expression.

In the case of variable sized data, we get

	char xyzzy[i];
	bar(xyzzy);
	return xyzzy[1];
=>
        ...
	T.9 = __builtin_alloca (T.7);
	xyzzy.8 = (char[T.1] *)T.9;
	bar (xyzzy.8);
	T.11 = (*xyzzy.8)[1];
	T.10 = (int)T.11;
	return T.10;

So we've replaced xyzzy (the VLA) with xyzzy.8 (a pointer to the array).
We'll have set DECL_VALUE_EXPR(xyzzy) = <INDIRECT_REF xyzzy.8>.

With a little bit of rearrangement in dwarf2out.c, we can render this as

        .ascii "xyzzy\0"        # DW_AT_name
	...
        .byte   0x2     # DW_AT_location
        .byte   0x73    # DW_OP_breg3
        .sleb128 0

The breg3 is because xyzzy.8 got allocated to reg3, %ebx.  If I add
an asm with some clobbers, to force the register allocator to spill
xyzzy.8 to memory, I get

        .byte   0x3     # DW_AT_location
        .byte   0x91    # DW_OP_fbreg
        .sleb128 -16
        .byte   0x6     # DW_OP_deref

Here we find the address of xyzzy.8 on the stack, and dereference.
All as we expected.

The full generality of DECL_VALUE_EXPR should be able to handle
debug info for SRA.  Of course, IIRC gdb doesn't actually handle
DW_OP_piece at the moment, so it's largely irrelevant for now.

I'm in the process of running the gdb testsuite to make sure I didn't
somehow bust things up for some corner case.  Perhaps I'll have it
all done tomorrow.


r~