Register Spilling

Register Spilling

[Umar Janjua]

During register allocation , when certain virtual registers are decided to be 
spilled to memory, is there a specific scheme/strategy on the basis of which 
the spilled values are laid out on memory.

regards,

Re: Register Spilling

[Ian Lance Taylor]

Umar Janjua <Umar.Janjua@cl.cam.ac.uk> writes:

> During register allocation , when certain virtual registers are decided to be 
> spilled to memory, is there a specific scheme/strategy on the basis of which 
> the spilled values are laid out on memory.

Well, yes and no.  The spilled values are put into the stack frame.
The space is allocated by calls to assign_stack_local().  But there is
no predictable ordering of the spilled pseudo-registers, if that is
what you mean.  The registers are spilled as needed as the compiler
works through the instructions.

Ian

Re: Register Spilling

[Umar Janjua]

Well, if that is the case, then , how advantageous would be considering cache 
locality for such register spilled values, so that instructions that require 
spilled values do not cause cache miss.
Does the famous register allocation by graph coloring algorithm takes care of 
cache locality for spilled values.

regards



Umar Janjua <Umar.Janjua@cl.cam.ac.uk> writes:

> During register allocation , when certain virtual registers are decided to be 
> spilled to memory, is there a specific scheme/strategy on the basis of which 
> the spilled values are laid out on memory.

Well, yes and no.  The spilled values are put into the stack frame.
The space is allocated by calls to assign_stack_local().  But there is
no predictable ordering of the spilled pseudo-registers, if that is
what you mean.  The registers are spilled as needed as the compiler
works through the instructions.

Re: Register Spilling

[Falk Hueffner]

Umar Janjua <Umar.Janjua@cl.cam.ac.uk> writes:

> Well, if that is the case, then , how advantageous would be
> considering cache locality for such register spilled values, so that
> instructions that require spilled values do not cause cache miss.

Most stack frames easily fit into L1, so probably not a lot. There has
been some work on this motivated by other things, though; see "Optimal
Stack Slot Assignment in GCC" by Naveen Sharma and Sanjiv Kumar Gupta
in the GCC summit proceedings
(http://www.linux.org.uk/~ajh/gcc/gccsummit-2003-proceedings.pdf)

-- 
	Falk