hoist boundary check code to the out of loop

[imasahiro]

Current implementation of sprintf(..., "%x",...) always execute boundary check.
(e.g. https://github.com/h2o/qrintf/blob/master/share/qrintf/qrintf.h#L1321)

This pull request hoists this check to the out of loop to simplify main code of hexadecimal to string conversion.

[kazuho]

Thank you for the PR.

I agree that boundary check should not be performed for every byte, I am basically fine to merge the changes. Before that, would you mind making following changes so that the diff would be minimal?

• there is no need to introduce *p, IMO using ctx.str[ctx.off++] should be fine in terms of performance
  • the reasons why I believe so is:
  • both ctx.str and ctx.off would be stored on register since they are always passed-by-value
  • the expression can be compiled into movb %r1, (%r2,%r3) or something alike
  • ctx needs to be returned, and using the value to store both intermediate and final results will save the number of registers used
• do not introduce wlen as well - if possible, please update len for the purpose in order to save the registers

It would be great if you could make such changes so that the diff of gets limited to the topic of the branch (i.e. only change the _qrintf_chk_?x functions).

I would appreciate it if you could either open a separate PR or a separate commit within this PR to change other things in case you are convinced that my understanding described in the above listing is wrong.

[imasahiro]

@kazuho
Thanks for the feedback. I updated the branch with the suggested changes:

• remove *p and use ctx.str[off++] in _qrintf_chk_?x, use ctx.str[ctx.off++] in _qrintf_nck_?x.
• remove wlen

[kazuho]

Thanks for the changes. I think we are almost complete. Please consider the nit-picking comments on the commit.

[imasahiro]

Thank you for your comment.

But we cannot modify the code as above because it breaks a compatibility with snprintf.
Acording to man page of printf(http://linux.die.net/man/3/sprintf), when the
buffer does not have enough space to write, the return value of the snprintf
is the number of characters which would have been written to the final string
if enough space has been available.

Return Value
...
If the output was truncated due to this limit then the return value
is the number of characters (excluding the terminating null byte)
which would have been written to the final string if enough space
had been available.

[imasahiro]

I tried to reconsider about above comments and I updated the branch.

• Remove local variable off
• Introduce new local variable rest
• Both qrintf_chk?x and qrintf_nck?x now use ctx.str[ctx.off++]

[kazuho]

Hmm. Let me see. Thought that it was possible to not add any local vars...

[kazuho]

Sorry for the fuss. The proposed code seems optimal. Merged to master with little tweaks.

[imasahiro]

@kazuho
Thank you very much for your review and final merge.