New option back-references are local, beefup manual.

* doc/grep.texi : Document the new options and the new behaviour
back-references are local.  Use excerpt from Karl Berry regex
texinfo.
* bootstrap/Makefile.try : Added xstrtoumax.o xstrtoul.o hard-local.o
From Guglielmo 'bond' Bondioni :
The bug was that using a multi line file that contained REs (one per
line), backreferences in the REs were considered global (to the file)
and not local (to the line).
That is, \1 in line n refers to the first \(.\) in the whole file,
rather than in the line itself.
From Tapani Tarvainen :
# Re: grep -e '\(a\)\1' -e '\(b\)\1'
That's not the way it should work: multiple -e arguments
should be treated as independent patterns and back references
should not refer to previous ones.
From Paul Eggert :
GNU grep currently does not issue
diagnostics for the following two cases, both of which are erroneous:
grep -e '[' -e ']'
grep '[
]'
POSIX requires a diagnostic in both cases because '[' is not a valid
regular expression.
To overcome those problems, grep no longer pass the concatenate
patterns to GNU regex but rather compile each patterns separately
and keep the result in an array.
* src/search.c (patterns) : New global variable; a structure array
holding the compiled patterns.
Declare function prototypes to minimize error.
(dfa, kswset, regexbuf, regs) : Removed, no longer static globals, but
rather fields in patterns[] structure per motif.
(Fcompile) : Alloc an entry in patterns[] to hold the regex.
(Ecompile) : Alloc an entry per motif in the patterns[] array.
(Gcompile) : Likewise.
(EGexecute) : Loop through of array of patterns[] for a match.
From Bernd Strieder :
# tail -f logfile | grep important | do_something_urgent
# tail -f logfile | grep important | do_something_taking_very_long
If grep does full buffering in these cases then the urgent operation
does not happen as it should in the first case, and in the second case
time is lost due to waiting for the buffer to be filled.
This is clearly spoken not grep's fault in the first place, but libc's.
There is a heuristic in libc that make a stream line-buffered only if a
terminal is on the other end. This doesn't take care of the cases where
this connection is somehow indirect.
* src/grep.c (line_buffered) : new option variable.
(prline) : if line_buffered is set fflush() is call.
(usage) : line_buffered new option.
Input from Paul Eggert, doing setvbuf() may not be portable
and breaks grep -z.
This patch prevent kwset_matcher from following problems.
For example, in SJIS encoding, one character has the codepoint 0x895c.
So the second byte of the character can match with '\' incorrectly.
And in eucJP encoding, there are the characters whose codepoints are
0xa5b9, 0xa5c8.  On the other hand, there is one character whose
codepoint is 0xb9a5.  So 0xb9a5 can match with 2nd byte of 0xa5b9
and 1st byte of 0xa5c8.
(EGexecute) : call check_multibyte_string when kwset is set.
(Fexecute) : call to check_multibyte_string.
(MB_CUR_MAX) : new macro.

2 files changed, 6 insertions, 0 deletions

diff --git a/bootstrap/.cvsignore b/bootstrap/.cvsignore
new file mode 100644
index 00000000..282522db
--- /dev/null
+++ b/bootstrap/.cvsignore
@@ -0,0 +1,2 @@
+Makefile
+Makefile.in
diff --git a/bootstrap/Makefile.try b/bootstrap/Makefile.try
index bd9aef16..b34f0b2d 100644
--- a/bootstrap/Makefile.try
+++ b/bootstrap/Makefile.try
@@ -13,11 +13,15 @@ OBJS = alloca.$(OBJEXT) \
       getopt1.$(OBJEXT) \
       grep.$(OBJEXT) \
       kwset.$(OBJEXT) \
+      hard-local.$(OBJEXT) \
+      isdir.$(OBJEXT) \
       memchr.$(OBJEXT) \
       obstack.$(OBJEXT) \
       regex.$(OBJEXT) \
       savedir.$(OBJEXT) \
       search.$(OBJEXT) \
+      xstrtol.$(OBJEXT) \
+      xstrtoumax.$(OBJEXT) \
       stpcpy.$(OBJEXT)
 
 # Where is DIR and opendir/readdir defined.