| Re: BLOCK WAS VECTORIZED. |
|
 |
|
 |
|
 |
|
 |
Group: comp.lang.fortran · Group Profile
Author: glen herrmannsfeldtglen herrmannsfeldt Date: Sep 20, 2008 09:28
Steve Lionel wrote:
(snip)
>> main-goodformat2.f90(147): (col. 7) remark: BLOCK WAS VECTORIZED.
>> Can these lines cause error in the result?
> No - they are optimization information messages from the compiler
> letting you know that it figured out it could use vector instructions to
> perform as many as four operations in one instruction. That they appear
> by default is somewhat historical.
Since some optimizations can generate different results,
it does seem worth asking about.
(snip)
> 4. All x64 processors and all Intel-based Mac processors support SSE2,
> for IA-32 the compiler has a switch to restrict it to generic "Pentium"
> instructions (-mia32 on Linux, /arch:ia32 on Windows.)
SSE2 instead of x87, or something else?
If you compare to x87 code with intermediate results in
the 80 bit registers (64 bit significand) then the results
might be different. There is a discussion on another list on
the advantages of extra precision in floating point products.
It would be nice to have a high-level language allowing for
extra precision in multiply/add, such as generated by the x87.
-- glen
| 
