Re: Assigning multiple contiguous registers to a virtual register

shrey <shreyas76@gmail.com>
Tue, 16 Mar 2010 09:49:47 -0700 (PDT)

          From comp.compilers

Related articles
Assigning multiple contiguous registers to a virtual register shreyas76@gmail.com (shrey) (2010-03-13)
Re: Assigning multiple contiguous registers to a virtual register pronesto@gmail.com (Fernando) (2010-03-15)
Re: Assigning multiple contiguous registers to a virtual register shreyas76@gmail.com (shrey) (2010-03-16)
| List of all articles for this month |

From: shrey <shreyas76@gmail.com>
Newsgroups: comp.compilers
Date: Tue, 16 Mar 2010 09:49:47 -0700 (PDT)
Organization: Compilers Central
References: 10-03-025 10-03-037
Keywords: registers
Posted-Date: 16 Mar 2010 23:36:12 EDT

Thats some good data except I am not sure how general it is esp
considering new languages.


On a different note, if its an architecture without alignment
restrictions, then would it be a possibility to present to the
register allocator a homogeneous virtual register world ? Something
like if x86 could allow AL, BH to be allocated to a 64 bit instead of
(AH, Al)/ AX.


Of course,. register allocator has to be taught how to still consider
contiguous registers.
Is there any existing work along those lines ?




thanks
shrey






On Mar 15, 6:45 pm, Fernando <prone...@gmail.com> wrote:
> Usually we do not see many virtuals with different sizes. I have some
> numbers in here:
>
> http://www2.dcc.ufmg.br/laboratorios/llp/wiki/doku.php?id=benchmarkpz...
>
> This page above talks about register allocation by puzzle solving,
> i.e, we use this abstraction called puzzles:
>
> http://compilers.cs.ucla.edu/fernando/projects/puzzles/tutorial/
>
> If a puzzle has pieces of different sizes, then it means that you are
> using virtuals of different sizes, i.e, 32 and 64-bit floating point
> values, for instance. But most of the floating point benchmarks that I
> know are too homogeneous: either all the pieces are 32-bits, or 64-
> bits.


Post a followup to this message

Return to the comp.compilers page.
Search the comp.compilers archives again.