On 09/23/2015 07:55 AM, Kevin O'Connor wrote: > On Tue, Sep 22, 2015 at 01:24:47PM -0700, Richard Henderson wrote: >> With an eye toward having this data replace the gen_opc_* arrays >> that each target collects in order to enable restore_state_from_tb. > > Hi Richard, > > Instead of having each architecture front-end determine the constants > to be restored on an exception, have you considered having the tcg > liveness pass automatically detect them? > > What I was thinking was if: > - each front-end stored each constant on every instruction using > regular "movi" ops > - tcg_liveness_analysis() tracked which global memory "sync" writes > are purely due to an op that can raise an exception > - then tcg_liveness_analysis() could remove "movi" instructions with > outputs that are needed only during an exception and place the > constant directly in the compressed table itself. > > I'm curious if this was considered and if there is a reason it > wouldn't work well.
We certainly don't have enough information to infer something like that. The moment we reach a helper that isn't marked as TCG_CALL_NO_WG, all that inference has to go out the window ans we have to assume that the "movi op" is both necessary and overwritten. The knowledge of which helpers modify a field such as cc_op is baked into the translators at a different level. r~
