-ftest-coverage and .got sections on alpha

-ftest-coverage and .got sections on alpha

[Brad Lucier]

I have some large program files that compile fine on alphaev6 unless
I use -ftest-coverage and -fprofile-arcs to prepare them for gcov
analysis.  When I do that I get the infamous .got section over 64K
message from the linux loader.  This implies to me that gcc is emitting
many symbols that eventually get into the .o file just for test coverage.

This general problem is dicsussed regularly on the RH alpha-linux mailing
list and the general advice is to place many small objects into an
array, etc., so that they don't require entries in the .got table.
Would it be possible, or a good idea, to do something like this
for the extra information that -ftest-coverage and -fprofile-arcs places
into the .o file?

Brad Lucier

Re: -ftest-coverage and .got sections on alpha

[Martin v. Loewis]

> I have some large program files that compile fine on alphaev6 unless
> I use -ftest-coverage and -fprofile-arcs to prepare them for gcov
> analysis.  When I do that I get the infamous .got section over 64K
> message from the linux loader. 

Not being very familiar with that architecture: Where does this limit
come from, and how could it be removed? I find it very strange that
a 64 bit target suffers from a 16 bit limit...

Regards,
Martin

Re: -ftest-coverage and .got sections on alpha

[Richard Henderson]

On Mon, May 29, 2000 at 12:07:08AM +0200, Martin v. Loewis wrote:
> Not being very familiar with that architecture: Where does this limit
> come from, and how could it be removed?

From the 16-bit memory offset in the instruction.  The .got limit
is per-object file, but apparently Brad's files are large enough 
that he's hitting it anyway.


r~

Re: -ftest-coverage and .got sections on alpha

[Richard Henderson]

On Sun, May 28, 2000 at 02:36:53PM -0500, Brad Lucier wrote:
> When I do that I get the infamous .got section over 64K message from
> the linux loader.  This implies to me that gcc is emitting many symbols
> that eventually get into the .o file just for test coverage.

I just looked again.  It isn't.  It's referencing exactly 2,
no matter the size of the input.


r~

Re: -ftest-coverage and .got sections on alpha

[Brad Lucier]

> 
> On Sun, May 28, 2000 at 02:36:53PM -0500, Brad Lucier wrote:
> > When I do that I get the infamous .got section over 64K message from
> > the linux loader.  This implies to me that gcc is emitting many symbols
> > that eventually get into the .o file just for test coverage.
> 
> I just looked again.  It isn't.  It's referencing exactly 2,
> no matter the size of the input.

Sorry, I was confused, I had this difficulty (and still have this difficulty)
ith 2.95.1, I cannot reproduce it with the development compiler.

Brad

With the 2.95.1 compiler:

popov-201% gcc -O0 -ftest-coverage -fprofile-arcs -mieee -D___SINGLE_HOST -c _meroon.c all.c test.c test_.c
popov-202% gcc -O0 -ftest-coverage -fprofile-arcs -mieee -D___SINGLE_HOST -o test  _meroon.o all.o test.o test_.o -lgambc -lm -ldl
/export/u10/binutils-2.9.5.0.16/bin/ld: _meroon.o: .got subsegment exceeds 64K (size 91000)
/export/u10/binutils-2.9.5.0.16/bin/ld: failed to set dynamic section sizes: No such file or directory
collect2: ld returned 1 exit status

Re: -ftest-coverage and .got sections on alpha

[Richard Henderson]

On Sun, May 28, 2000 at 08:25:12PM -0500, Brad Lucier wrote:
> Sorry, I was confused, I had this difficulty (and still have this difficulty)
> ith 2.95.1, I cannot reproduce it with the development compiler.

Ah, yes.  Before a recent rewrite of that code we did reference
lots of little 8-byte pieces.


r~