Register windows

From gem5
Revision as of 21:37, 6 June 2007 by Gblack (talk | contribs)
(diff) ← Older revision | Latest revision (diff) | Newer revision → (diff)
Jump to: navigation, search

Register windows prove to be a bit problematic with the current model of StaticInsts and the read/write register methods. The issue is that based on the register's index, it is possible that it will be one of the registers whose actual index is determined by a combination of its index and the Current Window Pointer (CWP). In SPARC, r0-r7 are non-windowed registers, while r8-r31 are windowed registers.

StaticInst

Register windows provide a form of dynamic renaming that is troublesome with our current StaticInst. Because the CWP's value has nothing to do with the current instruction, it is impossible to statically decode the registers in use based on the machine instruction alone. Thus renaming of registers must happen dynamically, either within the CPU's readReg() or writeReg() function, or in the instruction's execute() function. The latter is difficult because the current readReg() and writeReg() functions refer to a source or destination index, and not the architectural register index.

SimpleCPU

Windowing will be built into the SPARC IntRegFile object, which SimpleCPU will use via CPUExecContext. On writes to the CWP, a callback from the MiscRegFile object to the CPUExecContext (e.g., changeIntRegContext()) will be reflected by the CPUExecContext back down to the IntRegFile so that it can update its internal notion of the CWP.

changeIntRegContext() will be an ISA-independent function. Presumably it will take as an argument a pointer to some ISA-specific object (derived from an ISA-independent base object) that describes the change (sort of like the new Fault objects)... unless anyone has a better idea.

I think a better parameter would be an index which specified what part of the context you were changing and a new value. Otherwise, all of the architectures would have to use a constraining (or trivial) base class for their integer register file context, and keep around the value of the parts they don't want to change. An example would be changing the cwp without changing between sets of global registers.

--Gblack 13:04, 6 April 2006 (EDT)

Detailed CPU

In the detailed CPU case, register window renaming will happen prior to the normal rename. The initial arch reg index will be renamed by adding the appropriate register window shift to them. Then normal renaming can happen given the unique, physical register index. This will allow all current dependence mechanisms to work normally. The first rename must happen at decode/rename. A CWP register at decode or rename will get speculatively updated on SAVE and RESTORE instructions. It will have to be restored upon squashing.

In this case, the MiscRegFile callback changeIntRegContext() will go to the proxy ExecContext, which can forward it up to the CPU object to update its internal CWP. The fact that this update has to be done at decode/rename (before execute) is unresolved... likely we will have to introduce a new virtual StaticInst method to capture decode/rename-time execution effects.

Adding this function raises several questions (that probably ought to have their own wiki page):

  • In the common case where there's nothing to do, is it faster to have a null virtual function or a bool that suppresses calling the function?
  • Are there other behaviors currently hard-coded into the pipeline (or implemented via flags) that could be moved to this function?
  • Should we do the same thing for commit (create a virtual function and move some existing code there)? I think probably yes...