VBCC - What's going on here? - Page 4

phx · 26 July 2018, 12:45

There are certainly bugs in several mathieee#?.library versions, but I don't want to exclude a problem with vbcc or vasm at this point. There may be strange side-effects from both sides.

When compiling the examples for Kickstart 1.3 (using the provided config file or the default kick13 one) I can reproduce the issue on a Kickstart 1.3 Amiga (mathieeedoubbas.library 34.46). I was unable to reproduce it on other configurations, as I have mostly V40 setups.

While trying to isolate the problem it disappeared, after I recompiled vfprintf.c with my current development versions of vbccm68k and vasm. There were only minor changes in the m68k backend during the last months. In my list for the next release I found this:

Code:

- m68k: use asr-sequence for signed div with power-of-two
- m68k: fixed peephole bug
- m68k: New attribute __fp0ret, to force float-return in fp0, disregarding a -no-fp-return option.
- m68k: Fixed bug with unsigned short array indexes > 32767.

The biggest change concerns vasm's peephole optimizer. We finally started to eliminate the last bits of PhxAss-specific directives from the source and completely switched to vasm-native. Which means the output uses the vasm-native OPT directive now, to enable optimizations.

Anyway, this strange issue certainly needs further analyzation, but the recompiled update of m13.lib works for me. Please try that.

Probably I also have to recompile mieee.lib for 2.0+, but it is difficult to verify, as I cannot reproduce any bug there at the moment.

deimos · 26 July 2018, 13:08

I've copied those files into my lib directory and rebuilt my code...

It works!

Thanks!

I should be able to verify the 2.0 libs if you'd like.

Quote:

Originally Posted by phx

There are certainly bugs in several mathieee#?.library versions, but I don't want to exclude a problem with vbcc or vasm at this point. There may be strange side-effects from both sides.

When compiling the examples for Kickstart 1.3 (using the provided config file or the default kick13 one) I can reproduce the issue on a Kickstart 1.3 Amiga (mathieeedoubbas.library 34.46). I was unable to reproduce it on other configurations, as I have mostly V40 setups.

While trying to isolate the problem it disappeared, after I recompiled vfprintf.c with my current development versions of vbccm68k and vasm. There were only minor changes in the m68k backend during the last months. In my list for the next release I found this:

Code:

- m68k: use asr-sequence for signed div with power-of-two
- m68k: fixed peephole bug
- m68k: New attribute __fp0ret, to force float-return in fp0, disregarding a -no-fp-return option.
- m68k: Fixed bug with unsigned short array indexes > 32767.

The biggest change concerns vasm's peephole optimizer. We finally started to eliminate the last bits of PhxAss-specific directives from the source and completely switched to vasm-native. Which means the output uses the vasm-native OPT directive now, to enable optimizations.

Anyway, this strange issue certainly needs further analyzation, but the recompiled update of m13.lib works for me. Please try that.

Probably I also have to recompile mieee.lib for 2.0+, but it is difficult to verify, as I cannot reproduce any bug there at the moment.

Photon · 26 July 2018, 14:31

Quote:

Originally Posted by deimos

Printing enough decimal places also makes the problem go away.

Code:

printf("%0.13f\n%0.13f\n%0.13f\n%0.13f\n%0.13f\n\n", neg_one, zero, one, two, three); // good
printf("%0.12f\n%0.12f\n%0.12f\n%0.12f\n%0.12f\n\n", neg_one, zero, one, two, three); // bad

Aren't the 2.0 libs verified by this? I.e. result OK but destroyed later.

deimos · 26 July 2018, 15:35

No, it doesn't say anything about the 2.0 libs. This, and the other recent tests point to the problem being in printf, and it seems that phx has shown that it's not a bug in printf, but some optimisations gone wrong when building printf. A new version of m13.lib has fixed the problem, and new versions of the 2.0+ equivalents will probably fix the problem there too.

Quote:

Originally Posted by Photon

Aren't the 2.0 libs verified by this? I.e. result OK but destroyed later.

phx · 26 July 2018, 15:53

Quote:

Originally Posted by deimos

It works!

Great! Thanks for the confirmation.

Quote:

I should be able to verify the 2.0 libs if you'd like.

Ok. I will rebuild the m68k-amigaos target libs as soon as possible. Probably I will also consider an official patch release.

In the meantime, in case you have a lot of time, you might want to test with the mieee.lib from older vbcc releases. I wonder if this is a problem which appeared recently, with the last release.
http://server.owl.de/~frank/vbcc/201...8k-amigaos.lha
http://server.owl.de/~frank/vbcc/201...8k-amigaos.lha
http://server.owl.de/~frank/vbcc/201...8k-amigaos.lha
(Just take the mieee.lib from the target archives.)

deimos · 26 July 2018, 17:04

Well... of the three links you give, only the 2011-08-05 is free of the problem.

Rather than copy files over my installation I edited my vc.vfg to point to each version. Since the result is surprising I'm going to now go and copy the maths libs over instead, just in case.

Quote:

Originally Posted by phx

Great! Thanks for the confirmation.

Ok. I will rebuild the m68k-amigaos target libs as soon as possible. Probably I will also consider an official patch release.

In the meantime, in case you have a lot of time, you might want to test with the mieee.lib from older vbcc releases. I wonder if this is a problem which appeared recently, with the last release.
http://server.owl.de/~frank/vbcc/201...8k-amigaos.lha
http://server.owl.de/~frank/vbcc/201...8k-amigaos.lha
http://server.owl.de/~frank/vbcc/201...8k-amigaos.lha
(Just take the mieee.lib from the target archives.)

deimos · 26 July 2018, 17:13

Yup, 2011-08-05 is good, 2014-12-30 and 2016-03-23 are bad.

Sorry.

Quote:

Originally Posted by phx

Great! Thanks for the confirmation.

Ok. I will rebuild the m68k-amigaos target libs as soon as possible. Probably I will also consider an official patch release.

In the meantime, in case you have a lot of time, you might want to test with the mieee.lib from older vbcc releases. I wonder if this is a problem which appeared recently, with the last release.
http://server.owl.de/~frank/vbcc/201...8k-amigaos.lha
http://server.owl.de/~frank/vbcc/201...8k-amigaos.lha
http://server.owl.de/~frank/vbcc/201...8k-amigaos.lha
(Just take the mieee.lib from the target archives.)

Quote:

Originally Posted by deimos

Well... of the three links you give, only the 2011-08-05 is free of the problem.

Rather than copy files over my installation I edited my vc.vfg to point to each version. Since the result is surprising I'm going to now go and copy the maths libs over instead, just in case.

phx · 26 July 2018, 21:52

Quote:

Originally Posted by deimos

Yup, 2011-08-05 is good, 2014-12-30 and 2016-03-23 are bad.

Wow... the problem is older than I thought. But as it only became visible with old mathieeedoubbas.libraries it was not noticed until now. Thanks for your tests!

Attached you find a recompilation of all m68k-amigaos and m68k-kick13 math linker libraries.

deimos · 27 July 2018, 11:35

Thanks. I can confirm that both the 1.3 and 2.0 libs in that archive work for me.

phx · 28 July 2018, 16:14

Great! Thanks for testing.

26 July 2018, 12:45	#61
phx Natteravn Join Date: Nov 2009 Location: Herford / Germany Posts: 2,496	There are certainly bugs in several mathieee#?.library versions, but I don't want to exclude a problem with vbcc or vasm at this point. There may be strange side-effects from both sides. When compiling the examples for Kickstart 1.3 (using the provided config file or the default kick13 one) I can reproduce the issue on a Kickstart 1.3 Amiga (mathieeedoubbas.library 34.46). I was unable to reproduce it on other configurations, as I have mostly V40 setups. While trying to isolate the problem it disappeared, after I recompiled vfprintf.c with my current development versions of vbccm68k and vasm. There were only minor changes in the m68k backend during the last months. In my list for the next release I found this: Code: - m68k: use asr-sequence for signed div with power-of-two - m68k: fixed peephole bug - m68k: New attribute __fp0ret, to force float-return in fp0, disregarding a -no-fp-return option. - m68k: Fixed bug with unsigned short array indexes > 32767. The biggest change concerns vasm's peephole optimizer. We finally started to eliminate the last bits of PhxAss-specific directives from the source and completely switched to vasm-native. Which means the output uses the vasm-native OPT directive now, to enable optimizations. Anyway, this strange issue certainly needs further analyzation, but the recompiled update of m13.lib works for me. Please try that. Probably I also have to recompile mieee.lib for 2.0+, but it is difficult to verify, as I cannot reproduce any bug there at the moment. Last edited by phx; 07 December 2020 at 11:50.

Similar Threads
Thread	Thread Starter	Forum	Replies	Last Post
vbcc 0.9f released	phx	News	49	01 April 2018 11:33
vbcc V0.9e released	phx	News	17	31 October 2016 21:18
VBCC and #include	majikeyric	Coders. C/C++	3	03 March 2016 15:07
vbcc 0.9d	phx	News	43	13 July 2015 19:41
VBCC 0.8j for Windows	hitchhikr	Coders. General	11	09 October 2008 00:58

27 July 2018, 11:35	#69
deimos It's coming back! Join Date: Jul 2018 Location: comp.sys.amiga Posts: 762	Thanks. I can confirm that both the 1.3 and 2.0 libs in that archive work for me.

28 July 2018, 16:14	#70
phx Natteravn Join Date: Nov 2009 Location: Herford / Germany Posts: 2,496	Great! Thanks for testing.

Currently Active Users Viewing This Thread: 1 (0 members and 1 guests)