Hi all,
I noticed during current builgs, on Ubuntu 16.10 (Yakkety) and now also on Debian 9 (Stretch) was changed the default compiler to GCC6. This brings the number of new FTBFS. That's why I started to concentrate on these FTBFS. Note - for this reason the packages for Yakkety in Preliminary Stable Builds are now inconsistent.
As a first, of course, I started with tdelibs. Most were simple conflict signed / unsigned char or int. However, one serious problem is here.
In the 'kjs' and 'tdehtml/ecma' is used HashEntry structure that contains the item 'value' of type short int - see 'kjs/lookup.h'. Into this item was stored values DOM::NodeFilter::ShowCode::SHOW_ALL = 0xFFFFFFFF (see tdehtml/dom/dom2_traversal.h) and kjs event CHANGE with value 32768 (see tdehtml/ecma/kjs_events.cpp). Both are therefore outside the range of short int.
Changing the type of 'value' in HashEntry from short to long would apparently cause a change in ABI - kjs/lookup.h is part of the public includes. That's why I made the change values 0xFFFFFFFF and 32768 to -1. I think that GCC compilers version less than 6 will store these values in the same way. Therefore, I believe that the proposed patch would not cause problems == can be pushed. What is your opinion?
Cheers
On Wed, Aug 17, 2016 at 03:00 (+0200), Slávek Banko wrote:
Hi all,
I noticed during current builgs, on Ubuntu 16.10 (Yakkety) and now also on Debian 9 (Stretch) was changed the default compiler to GCC6. This brings the number of new FTBFS. That's why I started to concentrate on these FTBFS. Note - for this reason the packages for Yakkety in Preliminary Stable Builds are now inconsistent.
As a first, of course, I started with tdelibs. Most were simple conflict signed / unsigned char or int. However, one serious problem is here.
In the 'kjs' and 'tdehtml/ecma' is used HashEntry structure that contains the item 'value' of type short int - see 'kjs/lookup.h'. Into this item was stored values DOM::NodeFilter::ShowCode::SHOW_ALL = 0xFFFFFFFF (see tdehtml/dom/dom2_traversal.h) and kjs event CHANGE with value 32768 (see tdehtml/ecma/kjs_events.cpp). Both are therefore outside the range of short int.
Changing the type of 'value' in HashEntry from short to long would apparently cause a change in ABI - kjs/lookup.h is part of the public includes. That's why I made the change values 0xFFFFFFFF and 32768 to -1. I think that GCC compilers version less than 6 will store these values in the same way. Therefore, I believe that the proposed patch would not cause problems == can be pushed. What is your opinion?
Is anyone compiling this for a 1's complement hardware platform? If so, that could be an issue (now or in the future).
I don't have any gcc 6.0 machine to play with, so here are some alternative thoughts, for what they are worth.
Would (short)0xFFFFFFFF make it happy? Or how about ~0 (or ~(short)0) to get all 1 bits?
Cheers. Jim
On Wednesday 17 of August 2016 16:26:26 Jim wrote:
On Wed, Aug 17, 2016 at 03:00 (+0200), Slávek Banko wrote:
Hi all,
I noticed during current builgs, on Ubuntu 16.10 (Yakkety) and now also on Debian 9 (Stretch) was changed the default compiler to GCC6. This brings the number of new FTBFS. That's why I started to concentrate on these FTBFS. Note - for this reason the packages for Yakkety in Preliminary Stable Builds are now inconsistent.
As a first, of course, I started with tdelibs. Most were simple conflict signed / unsigned char or int. However, one serious problem is here.
In the 'kjs' and 'tdehtml/ecma' is used HashEntry structure that contains the item 'value' of type short int - see 'kjs/lookup.h'. Into this item was stored values DOM::NodeFilter::ShowCode::SHOW_ALL = 0xFFFFFFFF (see tdehtml/dom/dom2_traversal.h) and kjs event CHANGE with value 32768 (see tdehtml/ecma/kjs_events.cpp). Both are therefore outside the range of short int.
Changing the type of 'value' in HashEntry from short to long would apparently cause a change in ABI - kjs/lookup.h is part of the public includes. That's why I made the change values 0xFFFFFFFF and 32768 to -1. I think that GCC compilers version less than 6 will store these values in the same way. Therefore, I believe that the proposed patch would not cause problems == can be pushed. What is your opinion?
Is anyone compiling this for a 1's complement hardware platform? If so, that could be an issue (now or in the future).
I don't have any gcc 6.0 machine to play with, so here are some alternative thoughts, for what they are worth.
Would (short)0xFFFFFFFF make it happy? Or how about ~0 (or ~(short)0) to get all 1 bits?
Cheers. Jim
It does not make sense to have in the code (short)0xFFFFFFFF. Combination 'short' together with the value for which is insufficient signed long, with the idea 'compiler will do something with this' has no meaningful reason.
Format ~0 looks good, but I'm afraid of possible complications:
1) hash tables are generated by kjs/create_hash_table (perl script) - it would have to verify what would be the behavior with expressions instead of simple values.
2) enum values can be used in the code for the bindings to other languages - such as python. Use expressions instead of simple values could cause problems.
Therefore, I believe that the use of the simple value -1 makes more sense.
On Thu, Aug 18, 2016 at 20:55 (+0200), Slávek Banko wrote:
On Wednesday 17 of August 2016 16:26:26 Jim wrote:
On Wed, Aug 17, 2016 at 03:00 (+0200), Slávek Banko wrote:
Hi all,
Various irrelevancies <snip>ed...
However, one serious problem is here.
In the 'kjs' and 'tdehtml/ecma' is used HashEntry structure that contains the item 'value' of type short int - see 'kjs/lookup.h'. Into this item was stored values DOM::NodeFilter::ShowCode::SHOW_ALL = 0xFFFFFFFF (see tdehtml/dom/dom2_traversal.h) and kjs event CHANGE with value 32768 (see tdehtml/ecma/kjs_events.cpp). Both are therefore outside the range of short int.
Changing the type of 'value' in HashEntry from short to long would apparently cause a change in ABI - kjs/lookup.h is part of the public includes. That's why I made the change values 0xFFFFFFFF and 32768 to -1. I think that GCC compilers version less than 6 will store these values in the same way. Therefore, I believe that the proposed patch would not cause problems == can be pushed. What is your opinion?
Is anyone compiling this for a 1's complement hardware platform? If so, that could be an issue (now or in the future).
I don't have any gcc 6.0 machine to play with, so here are some alternative thoughts, for what they are worth.
how about ~0 (or ~(short)0) to get all 1 bits?
Format ~0 looks good, but I'm afraid of possible complications:
- hash tables are generated by kjs/create_hash_table (perl script) - it
would have to verify what would be the behavior with expressions instead of simple values.
Is that verification difficult?
- enum values can be used in the code for the bindings to other
languages - such as python. Use expressions instead of simple values could cause problems.
Is it difficult to check that?
Therefore, I believe that the use of the simple value -1 makes more sense.
(FWIW) Wikipedia says The representation used in most current computing devices is two's complement, although the Unisys ClearPath Dorado series mainframes use ones' complement. Putting a -1 in there will not give you 0xFFFF on such machines. Maybe no one cares...
Jim
On 2016/08/17 09:00 AM, Slávek Banko wrote:
Hi all,
I noticed during current builgs, on Ubuntu 16.10 (Yakkety) and now also on Debian 9 (Stretch) was changed the default compiler to GCC6. This brings the number of new FTBFS. That's why I started to concentrate on these FTBFS. Note - for this reason the packages for Yakkety in Preliminary Stable Builds are now inconsistent.
As a first, of course, I started with tdelibs. Most were simple conflict signed / unsigned char or int. However, one serious problem is here.
In the 'kjs' and 'tdehtml/ecma' is used HashEntry structure that contains the item 'value' of type short int - see 'kjs/lookup.h'. Into this item was stored values DOM::NodeFilter::ShowCode::SHOW_ALL = 0xFFFFFFFF (see tdehtml/dom/dom2_traversal.h) and kjs event CHANGE with value 32768 (see tdehtml/ecma/kjs_events.cpp). Both are therefore outside the range of short int.
Changing the type of 'value' in HashEntry from short to long would apparently cause a change in ABI - kjs/lookup.h is part of the public includes. That's why I made the change values 0xFFFFFFFF and 32768 to -1. I think that GCC compilers version less than 6 will store these values in the same way. Therefore, I believe that the proposed patch would not cause problems == can be pushed. What is your opinion?
Cheers
I think the int/long values would simply be truncated when stored into a short int, giving unpredictable values. Does -1 have any special meaning when used in "value" field? If not, it should be safe to use it.
Cheers Michele
On Wednesday 17 of August 2016 18:18:23 Michele Calgaro wrote:
On 2016/08/17 09:00 AM, Slávek Banko wrote:
Hi all,
I noticed during current builgs, on Ubuntu 16.10 (Yakkety) and now also on Debian 9 (Stretch) was changed the default compiler to GCC6. This brings the number of new FTBFS. That's why I started to concentrate on these FTBFS. Note - for this reason the packages for Yakkety in Preliminary Stable Builds are now inconsistent.
As a first, of course, I started with tdelibs. Most were simple conflict signed / unsigned char or int. However, one serious problem is here.
In the 'kjs' and 'tdehtml/ecma' is used HashEntry structure that contains the item 'value' of type short int - see 'kjs/lookup.h'. Into this item was stored values DOM::NodeFilter::ShowCode::SHOW_ALL = 0xFFFFFFFF (see tdehtml/dom/dom2_traversal.h) and kjs event CHANGE with value 32768 (see tdehtml/ecma/kjs_events.cpp). Both are therefore outside the range of short int.
Changing the type of 'value' in HashEntry from short to long would apparently cause a change in ABI - kjs/lookup.h is part of the public includes. That's why I made the change values 0xFFFFFFFF and 32768 to -1. I think that GCC compilers version less than 6 will store these values in the same way. Therefore, I believe that the proposed patch would not cause problems == can be pushed. What is your opinion?
Cheers
I think the int/long values would simply be truncated when stored into a short int, giving unpredictable values. Does -1 have any special meaning when used in "value" field? If not, it should be safe to use it.
Cheers Michele
There is generated more hash tables where the 'value' has the meaning specific for each table. The tables, for which two mentioned changes concern, has no other item with the value -1. I dare to say that values even today could not be stored as anything other than -1.
Any other ideas, suggestions, objections?
Cheers
On 2016/08/19 03:05 AM, Slávek Banko wrote:
There is generated more hash tables where the 'value' has the meaning specific for each table. The tables, for which two mentioned changes concern, has no other item with the value -1. I dare to say that values even today could not be stored as anything other than -1.
Any other ideas, suggestions, objections?
Cheers
I will try to take a look in more detail over the weekend. I will come back on this. Cheers Michele
-
Jim -
Michele Calgaro -
Slávek Banko