Re: [PATCH] tcg/tcg: Avoid TS_DEAD for basic block ending

[Richard Henderson]

On 3/20/23 23:44, LIU Zhiwei wrote:
> On 2023/3/21 14:06, Richard Henderson wrote:
> > On 3/20/23 21:53, LIU Zhiwei wrote:
> > > TS_DEAD means we will release the register allocated for this temporary. But
> > > at basic block ending, we can still use the allocted register.
> > >
> > > Signed-off-by: LIU Zhiwei <zhiwei_liu@linux.alibaba.com>
> >
> > Test case?
>
> I have run an Ubuntu image after this patch. It can boot.

That's surprising.  I would have expected an assert with --enable-debug-tcg, but this 
appears to be an oversight in tcg_reg_alloc_bb_end.  We only validate the liveness data 
for TEMP_EBB and TEMP_CONST, but not TEMP_TB or TEMP_GLOBAL.

> But I can't find a direct test case.  Because the IRs supported with flags TCG_OPF_BB_END 
> do not have  input or output parameter, such as the set_label or br.

That's exactly why we want all GLOBAL and TB to be DEAD | MEM, so that they're saved back 
to their home slots and released from their registers.

The register allocator for TCG does not work across extended basic blocks.  Importantly, 
if you have a forward branch like so:


        g1 = func(a)
        brcond ..., L1

        stuff
        g2 = func(b)    
        g1 = g2
        discard g2

L1:

What value should g1->reg have at L1?  The allocator does not do the global control flow 
and allocation required to ensure that g1->reg is the same at the brcond and at the label.

Nominally, I would have expected one value for g1->reg at the branch, a different value 
for g2->reg in the second BB, and for the assignment to steal g2->reg and move it to 
g1->reg (per tcg_reg_alloc_mov of an IS_DEAD_ARG temp).  Which would result in an 
incorrect allocation at L1.

What are you attempting to do?  Is this just guesswork?


r~