On 06/20/2017 11:32 AM, Jakub Jelinek wrote:
> On Tue, Jun 20, 2017 at 11:23:36AM +0200, Martin Liška wrote:
>>> Then something needs to be done for debugging too.  If it is without VTA,
>>> then probably just having DECL_VALUE_EXPR is good enough, otherwise
>>> (VTA) you probably don't want that (or can reset it at that point), but
>>> instead emit after the initialization stmt a debug stmt that the variable
>>> value now lives in a different var.  Though ideally we want the debugger
>>> to be able to also change the value of the var, that might be harder.
>>> With DECL_VALUE_EXPR on the other side the debug info will be incorrect in
>>> the prologue until it is assigned to the slot.
>>
>> Here I'm not sure about how to distinguish whether to build or not to build
>> the debug statement. According to flag_var_tracking?
> 
> More like if (target_for_debug_bind (arg))
> And if that is false, just make sure DECL_VALUE_EXPR is set to var.
> 
>> You mean something like:
>> g = gimple_build_debug_bind (arg, var, g);
>> ?
> 
> Well, there is no stmt, so the last argument would be just NULL.
> 
>>> I don't understand the distinction.  If you turn the original parm
>>> for complex/vector DECL_GIMPLE_REG_P, you should need the exact same code
>>> (but I think it would be better to use the default SSA_NAME of the PARM_DECL
>>> if it is a gimple reg type, rather than use the PARM_DECL itself
>>> and wait for update_ssa).
>>
>> Yes, the test-case /gcc/testsuite/g++.dg/asan/function-argument-3.C fails for me
>> as one needs to have a temporary SSA name, otherwise:
>>
>> /home/marxin/Programming/gcc/gcc/testsuite/g++.dg/asan/function-argument-3.C:13:1: error: invalid rhs for gimple memory store
>>  foo (v4si arg)
>>  ^~~
>> arg
>>
>> arg
>>
>> # .MEM_4 = VDEF <.MEM_1(D)>
>> arg = arg;
>> during GIMPLE pass: sanopt
>>
>> If I see correctly the function in my test-case does not have default def SSA name for the parameter.
>> Thus I guess I need to create a SSA name?
> 
> I'd expect if you have DECL_GIMPLE_REG_P set on the PARM_DECL and
> use the default def, you shouldn't run into this.
> 
> 	Jakub
> 

Good I fixed that in v2, that passes regression tests.
Ale objections should be resolved in the version.

Ready for trunk?
Martin