I don't know if this is what you mean now I've seen your discussion in comments, but it certainly works for the cases in the original question.
This code does not rely on any package or code explicitly designated as experimental by the L3 developers.
Note, however, that I have no idea what I am doing.
Caveat emptor ....
Counts are included to show that the grouping within the token lists is preserved e.g. that a{bcde}f is counted as 3 tokens and not 6 or 8 when the token list is reassembled. During processing, the string is obviously counted as having more tokens since this is necessary to search and replace within the groups.
The result of the replacement operation is stored in a globally set variable \g_henri_mod_tl.
\documentclass{article}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{expl3}
\begin{document}
\ExplSyntaxOn
\str_new:N \l_henri_mod_str
\int_new:N \l_henri_tmpa_int
\int_new:N \l_henri_tmpb_int
\int_new:N \l_henri_tmpc_int
\tl_new:N \g_henri_mod_tl
\cs_new_protected:Npn \henri_replace_all:nnn #1 #2 #3
{
\group_begin:
\str_clear:N \l_henri_mod_str
\int_zero:N \l_henri_tmpa_int
\str_set:Nn \l_tmpa_str { #1 }
\str_set:Nn \l_tmpb_str { #2 }
\int_set:Nn \l_henri_tmpb_int { \str_count:N \l_tmpa_str }
\int_set:Nn \l_henri_tmpc_int { \str_count:N \l_tmpb_str }
\int_compare:nTF { \l_henri_tmpc_int = 1 }
{
\int_do_until:nn { \l_henri_tmpb_int = \l_henri_tmpa_int }
{
\str_if_eq_x:nnTF { #2 } { \str_head:N \l_tmpa_str }
{
\str_put_right:Nx \l_henri_mod_str { #3 }
}
{
\str_put_right:Nx \l_henri_mod_str { \str_head:N \l_tmpa_str }
}
\str_set:Nx \l_tmpa_str { \str_tail:N \l_tmpa_str }
\int_incr:N \l_henri_tmpa_int
}
}
{
\int_do_until:nn { \l_henri_tmpb_int = \l_henri_tmpa_int }
{
\str_if_eq_x:nnTF { #2 } { \str_range:Nnn \l_tmpa_str { 1 } { \l_henri_tmpc_int } }
{
\str_put_right:Nx \l_henri_mod_str { #3 }
\int_set:Nn \l_tmpa_int { \str_count:N \l_tmpa_str }
\str_set:Nx \l_tmpa_str { \str_range:Nnn \l_tmpa_str { 1 + \l_henri_tmpc_int } { \l_tmpa_int } }
\int_add:Nn \l_henri_tmpa_int { \l_henri_tmpc_int }
}
{
\str_put_right:Nx \l_henri_mod_str { \str_head:N \l_tmpa_str }
\str_set:Nx \l_tmpa_str { \str_tail:N \l_tmpa_str }
\int_incr:N \l_henri_tmpa_int
}
}
}
\tl_gset_rescan:Nno \g_henri_mod_tl {} { \l_henri_mod_str }
\group_end:
}
\cs_generate_variant:Nn \henri_replace_all:nnn { Vnn }
\henri_replace_all:nnn { abc{ab{abc}c} } { b } { d }
\g_henri_mod_tl {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\tl_set:Nn \l_tmpa_tl { abc{ab{abc}c} }
\henri_replace_all:Vnn \l_tmpa_tl { b } { d }
\g_henri_mod_tl {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\henri_replace_all:nnn { {a=b}\,{[]} } { [ } { \sqsubseteq }
$\g_henri_mod_tl$ {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\henri_replace_all:nnn { gydihŵs } { y } { w }
\g_henri_mod_tl {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\henri_replace_all:nnn { abc{ab{abc}c} } { bc } { doodle }
\g_henri_mod_tl {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\henri_replace_all:nnn { {a=[b}\,{[]} } { =[ } { \sqsubseteq }
$\g_henri_mod_tl$ {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\ExplSyntaxOff
\end{document}
EDITED to deal with searches for strings of more than one character. This can correctly substitute =[ with \sqsubseteq as mentioned in a comment.
EDIT
It is possible to define a further command sequence which obeys the target syntax. However, it should be noticed that this will not work in all cases. In particular, it fails to work correctly with gwdihŵs.
The idea is just to do the replacement and then spit out the global variable. I am not sure that it is correct to call the macro \tl_replace_allrecursive:nnn as this lacks any appropriate prefix, but if the macro is for purely personal use and you're not worried about future breakage, that's up to you. Personally, I'd call it something like \henri_replace_allrecursive:nnn and be safe since I don't see anything to be gained from violating the naming rules.
\cs_new_protected:Npn \tl_replace_allrecursive:nnn #1 #2 #3
{
\henri_replace_all:nnn { #1 } { #2 } { #3 }
\g_henri_mod_tl
}
Then we can say
\tl_set:Nx \l_tmpa_tl { \tl_replace_allrecursive:nnn { abc{ab{abc}c} } { b } { d } }
\l_tmpa_tl \par
\tl_set:Nx \l_tmpa_tl { \tl_replace_allrecursive:nnn { {a=b}\,{[]} } { [ } { \sqsubseteq } }
$\l_tmpa_tl$ \par
\tl_set:Nx \l_tmpa_tl { \tl_replace_allrecursive:nnn { abc{ab{abc}c} } { bc } { doodle } }
\l_tmpa_tl \par
%
\tl_set:Nx \l_tmpa_tl { \tl_replace_allrecursive:nnn { {a=[b}\,{[]} } { =[ } { \sqsubseteq } }
$\l_tmpa_tl$ \par
and, comparing with the original results, we can see that the replacements are as expected (less gwdihŵs, of course).
I take it the count of tokens here is irrelevant since everything is being expanded.
Complete code:
\documentclass{article}
\usepackage[utf8]{inputenc}
\usepackage[T1]{fontenc}
\usepackage{expl3}
\begin{document}
\ExplSyntaxOn
\str_new:N \l_henri_mod_str
\int_new:N \l_henri_tmpa_int
\int_new:N \l_henri_tmpb_int
\int_new:N \l_henri_tmpc_int
\tl_new:N \g_henri_mod_tl
\cs_new_protected:Npn \henri_replace_all:nnn #1 #2 #3
{
\group_begin:
\str_clear:N \l_henri_mod_str
\int_zero:N \l_henri_tmpa_int
\str_set:Nn \l_tmpa_str { #1 }
\str_set:Nn \l_tmpb_str { #2 }
\int_set:Nn \l_henri_tmpb_int { \str_count:N \l_tmpa_str }
\int_set:Nn \l_henri_tmpc_int { \str_count:N \l_tmpb_str }
\int_compare:nTF { \l_henri_tmpc_int = 1 }
{
\int_do_until:nn { \l_henri_tmpb_int = \l_henri_tmpa_int }
{
\str_if_eq_x:nnTF { #2 } { \str_head:N \l_tmpa_str }
{
\str_put_right:Nx \l_henri_mod_str { #3 }
}
{
\str_put_right:Nx \l_henri_mod_str { \str_head:N \l_tmpa_str }
}
\str_set:Nx \l_tmpa_str { \str_tail:N \l_tmpa_str }
\int_incr:N \l_henri_tmpa_int
}
}
{
\int_do_until:nn { \l_henri_tmpb_int = \l_henri_tmpa_int }
{
\str_if_eq_x:nnTF { #2 } { \str_range:Nnn \l_tmpa_str { 1 } { \l_henri_tmpc_int } }
{
\str_put_right:Nx \l_henri_mod_str { #3 }
\int_set:Nn \l_tmpa_int { \str_count:N \l_tmpa_str }
\str_set:Nx \l_tmpa_str { \str_range:Nnn \l_tmpa_str { 1 + \l_henri_tmpc_int } { \l_tmpa_int } }
\int_add:Nn \l_henri_tmpa_int { \l_henri_tmpc_int }
}
{
\str_put_right:Nx \l_henri_mod_str { \str_head:N \l_tmpa_str }
\str_set:Nx \l_tmpa_str { \str_tail:N \l_tmpa_str }
\int_incr:N \l_henri_tmpa_int
}
}
}
\tl_gset_rescan:Nno \g_henri_mod_tl {} { \l_henri_mod_str }
\group_end:
}
\cs_generate_variant:Nn \henri_replace_all:nnn { Vnn }
\cs_new_protected:Npn \tl_replace_allrecursive:nnn #1 #2 #3
{
\henri_replace_all:nnn { #1 } { #2 } { #3 }
\g_henri_mod_tl
}
\verb|\henri_replace_all:nnn { } { } { } \g_henri_mod_tl|
\smallskip\par
\henri_replace_all:nnn { abc{ab{abc}c} } { b } { d }
\g_henri_mod_tl {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\tl_set:Nn \l_tmpa_tl { abc{ab{abc}c} }
\henri_replace_all:Vnn \l_tmpa_tl { b } { d }
\g_henri_mod_tl {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\henri_replace_all:nnn { {a=b}\,{[]} } { [ } { \sqsubseteq }
$\g_henri_mod_tl$ {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\henri_replace_all:nnn { gydihŵs } { y } { w }
\g_henri_mod_tl {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\henri_replace_all:nnn { abc{ab{abc}c} } { bc } { doodle }
\g_henri_mod_tl {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\henri_replace_all:nnn { {a=[b}\,{[]} } { =[ } { \sqsubseteq }
$\g_henri_mod_tl$ {} ~ has ~ \tl_count:N \g_henri_mod_tl {} ~ tokens.\par
\bigskip\par
\verb|\tl_set:Nx \l_tmpa_tl { \tl_replace_allrecursive:nnn { ... } { ... } { ... } }|
\smallskip\par
\tl_set:Nx \l_tmpa_tl { \tl_replace_allrecursive:nnn { abc{ab{abc}c} } { b } { d } }
\l_tmpa_tl \par
\tl_set:Nx \l_tmpa_tl { \tl_replace_allrecursive:nnn { {a=b}\,{[]} } { [ } { \sqsubseteq } }
$\l_tmpa_tl$ \par
\tl_set:Nx \l_tmpa_tl { \tl_replace_allrecursive:nnn { abc{ab{abc}c} } { bc } { doodle } }
\l_tmpa_tl \par
%
\tl_set:Nx \l_tmpa_tl { \tl_replace_allrecursive:nnn { {a=[b}\,{[]} } { =[ } { \sqsubseteq } }
$\l_tmpa_tl$ \par
\ExplSyntaxOff
\end{document}
l3regex? – May 26 '16 at 21:35l3regexbefore. I am happy to read your answer using it. – Henri Menke May 26 '16 at 21:37\regex_replace_all:nnNis here for this purpose. In any case, I think they actually do something similar for\tl_(lower|upper)_case:nnso you might want to look at the implementation until an answer arrives. – Manuel May 26 '16 at 21:37\regex_replace_all:nnNdoes something similar internally what Henri tried to achieve with the\tl_map_inline– May 26 '16 at 21:45l3regexso I thought he was trying to solve it without that, hence my proposal of looking into\tl_.._case:nn. I thinkl3regexdoes internally by\detokenizeing the token list, then doing replacements. – Manuel May 26 '16 at 21:48\tl_..._case:n, because these are fully expandable. – Henri Menke May 26 '16 at 21:54\tl_set:Nx \l_tmpa_tl { \tl_upper_case:n { abc } }– Henri Menke May 27 '16 at 06:56