On 3/4/20 6:33 PM, Jakub Jelinek wrote:
On Wed, Mar 04, 2020 at 06:23:10PM +0100, Martin Liška wrote:
On 3/4/20 4:25 PM, Egeyar Bagcioglu wrote:
Thanks Richard.
I do not have write-access to the GCC repo. I'd be glad if someone commits it
for me.
Can we please wait? I'm really convinced we do not want one another very similar
functionality. I would definitely recommend to change the semantics
of -frecord-gcc-switches to what the patchset does.
That's why I added Nick as he's the author of the original implementation.
Reasoning is provided in my previous email.
Also, what is the reason for storing the option in a file, rather than say
putting them into an environment variable from which the backend can pick it
up?
I needed to pass that information from one process to the other and I
picked a way to do so. Moreover, passing files is what gcc does;
therefore, it didn't seem unnatural to do so. Furthermore, gcc specs
already has its way to create temporary file names so I didn't have to
worry about what level of thread safety is acceptable. If you tell me
why an environment variable is better, we can discuss and I might
implement that too.
I must say I don't really see advantages of this over
-grecord-gcc-switches, recording all options looks very bloaty and will
include mostly stuff you don't really care about (such as, e.g. the -I
options without knowing what was the current directory when the source file
has been compiled), on the other side will not record interesting options
that -grecord-gcc-switches records (say, if some code is compiled with
-march=native, this new option will record that, rather than what it really
is),
1) -grecord-gcc-switches is quite similar to -frecord-gcc-switches and
my first and main arguement about the latter is valid for the former:
You cannot easily construct the right command line to invoke the same
compilation from the switches. Switches are useful to check that the
compiler is doing the expected. The command line is useful when you want
the compiler to do the same thing again.
2) The output is exactly what gcc takes as command line to make that
compilation. It is bloaty, surely, when so is what's provided to gcc to
make it compile as desired. On the other hand, -grecord-gcc-switches is
just much bloatier since it does not work alone. It's embedded in dwarf.
3) Many systems automatically strip dwarf, together with this
information. I do not think there is an easy way to separate this
information and keep it.
Since this comes up often, I'd like to emphasize that the internal
switches of gcc are different than the user-given options of gcc. The
purpose of this new option is really to catch the user given options of
gcc and only them, so much so that I mark it SWITCH_IGNORE within the
specs to avoid it propagated to any child processes of gcc. This option
is simply to retrieve how the call was made to GCC. Currently, there are
no other options giving us this information.
but I won't stand in a way unless such an option would be on by
default.
I totally agree that this should not be on by default.
Best regards
Egeyar
