Repost: FastFPE results for Linux/ARM

Started by Marko Kreenover 21 years ago3 messagespatches
Jump to latest
#1Marko Kreen
markokr@gmail.com

On Linux/ARM PostgreSQL may encounter 2 types of floating point emulation:
1) standard precision: Linux/NWFPE, GCC soft-float
2) reduced precision: Linux/FastFPE

Currently pg_regress allows only unconditional alternatives - thus
different results are accepted on all platforms. This should be
avoided as acceptable behaviour on one platform may be bug on another.

Following patch:

* Lets pg_regress.sh accept multiple values from resultmap
* Adds FastFPE results for 'point' test

--
marko

Attachments:

arm-fastfpe.difftext/plain; charset=us-asciiDownload+253-32
#2Tom Lane
tgl@sss.pgh.pa.us
In reply to: Marko Kreen (#1)
Re: Repost: FastFPE results for Linux/ARM

Marko Kreen <marko@l-t.ee> writes:

On Linux/ARM PostgreSQL may encounter 2 types of floating point emulation:
1) standard precision: Linux/NWFPE, GCC soft-float
2) reduced precision: Linux/FastFPE

I don't really see why we should consider the latter as a "pass".
The regression tests exist in part to inform you when you are using
a substandard platform. Lately it seems that people have adopted the
goal that all the tests should "pass" no matter what. I fundamentally
disagree with that.

regards, tom lane

#3Marko Kreen
markokr@gmail.com
In reply to: Tom Lane (#2)
Re: Repost: FastFPE results for Linux/ARM

On Fri, Feb 11, 2005 at 12:08:22PM -0500, Tom Lane wrote:

Marko Kreen <marko@l-t.ee> writes:

On Linux/ARM PostgreSQL may encounter 2 types of floating point emulation:
1) standard precision: Linux/NWFPE, GCC soft-float
2) reduced precision: Linux/FastFPE

I don't really see why we should consider the latter as a "pass".
The regression tests exist in part to inform you when you are using
a substandard platform. Lately it seems that people have adopted the
goal that all the tests should "pass" no matter what. I fundamentally
disagree with that.

Well, the main problem is the confusion a failure causes. How
should a user decide whether he can run the database safely or
not, after a failed regression test? Or even what the reason
for failure was?

FastFPE on ARM is expected and seems to belong to "dont worry"
category. And I dont like saying: "Couple of regression failures
are normal".

Although I can understand that 'reduced precision' may sound
worrysome to lot of people. Maybe there should be new class
of failures in pg_regress: "expected failures" - which still
count as failures but have little note explaining the problem
so user can decide about the severity.

--
marko