Re: [PATCH] wip: change ISSPACE to use ascii whitespace check.

Alejandro Colomar via Mutt-dev <[email protected]> Thu, 19 Feb 2026 14:32:45 +0100
Newsgroups	gmane.mail.mutt.devel
Message-ID	<aZcQBd0v7RaKXv3w@devuan>
On 2026-02-19T12:52:00+0100, Alejandro Colomar wrote:
> Hi Ian,
> 
> On 2026-02-19T10:30:09+0000, Ian Collier via Mutt-dev wrote:
> > On Thu, Feb 19, 2026 at 01:46:34AM +0100, Alejandro Colomar via Mutt-dev wrote:
> > > On 2026-02-18T23:42:28+0000, Ian Collier via Mutt-dev wrote:
> > > > I do not understand why strchrnul and strcmp are even being invoked here.
> > 
> > > Because they are necessary.  (Well, strcmp(3) is not, as you could do
> > > manual byte operations, but it's simple, more readable, and optimized
> > > out.)  strchrnul(3) is necessary.
> > 
> > > 	strchr(s, '\0') != NULL
> > 
> > > would evaluate to true, which is a misbehavior: it would treat '\0' as
> > > white space, while it is not. 
> > 
> > The nul character is a bit of an edge case when considering strxxx functions,
> > of course.  But that's easily fixed:
> > 
> >  c && strchr(s, c) != NULL
> > 
> > No doubt you'll now be telling me that you don't want to evaluate c twice
> > in the context of a macro.  Well then, use an inline function. :-)
> 
> An inline function with that would be an alternative.  A macro is more
> compact, though, and can be implemented as a one-liner.  When
> implementing several of these, it can help readability of the entire set
> of APIs if they fit in a compact space and they are very similar:
> 
> 	// isascii_c - is [:ascii:] C-locale
> 	#define isascii_c(c)   (!!strchr(CTYPE_ASCII_C, c))
> 	#define iscntrl_c(c)   (!!strchr(CTYPE_CNTRL_C, c))
> 	#define islower_c(c)   (!streq(strchrnul(CTYPE_LOWER_C, c), ""))
> 	#define isupper_c(c)   (!streq(strchrnul(CTYPE_UPPER_C, c), ""))
> 	#define isdigit_c(c)   (!streq(strchrnul(CTYPE_DIGIT_C, c), ""))
> 	#define ispunct_c(c)   (!streq(strchrnul(CTYPE_PUNCT_C, c), ""))
> 	#define isspace_c(c)   (!streq(strchrnul(CTYPE_SPACE_C, c), ""))
> 	#define isalpha_c(c)   (!streq(strchrnul(CTYPE_ALPHA_C, c), ""))
> 	#define isalnum_c(c)   (!streq(strchrnul(CTYPE_ALNUM_C, c), ""))
> 	#define isgraph_c(c)   (!streq(strchrnul(CTYPE_GRAPH_C, c), ""))
> 	#define isprint_c(c)   (!streq(strchrnul(CTYPE_PRINT_C, c), ""))
> 	#define isxdigit_c(c)  (!streq(strchrnul(CTYPE_XDIGIT_C, c), ""))
> 
> Here's a comparison of the assembly they produce:
> 
> 	alx@devuan:~/tmp$ diff -u isspace?.c
> 	--- isspace1.c	2026-02-19 12:43:01.016632233 +0100
> 	+++ isspace2.c	2026-02-19 12:39:40.987744755 +0100
> 	@@ -1,8 +1,7 @@
> 	 #include <string.h>
> 	 #define CTYPE_SPACE_C   " \t\n\v\f\r"
> 	-#define streq(a,b)  (strcmp(a,b) == 0)
> 	 bool
> 	 isspace_c(int c)
> 	 {
> 	-	return !streq(strchrnul(CTYPE_SPACE_C, c), "");
> 	+	return c && strchr(CTYPE_SPACE_C, c) != NULL;
> 	 }
> 	alx@devuan:~/tmp$ gcc -S -O2 isspace?.c
> 	alx@devuan:~/tmp$ diff -u isspace?.s
> 	--- isspace1.s	2026-02-19 12:44:24.264776257 +0100
> 	+++ isspace2.s	2026-02-19 12:44:24.276776354 +0100
> 	@@ -1,4 +1,4 @@
> 	-	.file	"isspace1.c"
> 	+	.file	"isspace2.c"
> 		.text
> 		.section	.rodata.str1.1,"aMS",@progbits,1
> 	 .LC0:
> 	@@ -10,16 +10,23 @@
> 	 isspace_c:
> 	 .LFB0:
> 		.cfi_startproc
> 	+	testl	%edi, %edi
> 	+	je	.L5
> 		subq	$8, %rsp
> 		.cfi_def_cfa_offset 16
> 		movl	%edi, %esi
> 		leaq	.LC0(%rip), %rdi
> 	-	call	strchrnul@PLT
> 	-	cmpb	$0, (%rax)
> 	+	call	strchr@PLT
> 	+	testq	%rax, %rax
> 		setne	%al
> 		addq	$8, %rsp
> 		.cfi_def_cfa_offset 8
> 		ret
> 	+	.p2align 4,,10
> 	+	.p2align 3
> 	+.L5:
> 	+	xorl	%eax, %eax
> 	+	ret
> 		.cfi_endproc
> 	 .LFE0:
> 		.size	isspace_c, .-isspace_c
> 
> The strchrnul(3) version seems to have significantly less instructions.
> 
> And strchrnul(3) is internally simpler than strchr(3).  In musl, for
> example, strchrnul(3) is used to implement strchr(3).

Here's the comparison between isspace_c() using non-system strchrnul()
vs the brancy implementation with strchr(3):

	alx@devuan:~/tmp$ diff -u isspace[23].c
	--- isspace2.c	2026-02-19 12:39:40.987744755 +0100
	+++ isspace3.c	2026-02-19 14:26:39.575137468 +0100
	@@ -1,7 +1,14 @@
	 #include <string.h>
	 #define CTYPE_SPACE_C   " \t\n\v\f\r"
	+#define streq(a,b)  (strcmp(a,b) == 0)
	+#define strnul(s)  strchr(s, '\0')
	+static inline char *
	+my_strchrnul(const char *s, int c)
	+{
	+	return strchr(s, c) ?: strnul(s);
	+}
	 bool
	 isspace_c(int c)
	 {
	-	return c && strchr(CTYPE_SPACE_C, c) != NULL;
	+	return !streq(my_strchrnul(CTYPE_SPACE_C, c), "");
	 }
	alx@devuan:~/tmp$ gcc -S -O2 isspace[23].c
	alx@devuan:~/tmp$ diff -u isspace[23].s
	--- isspace2.s	2026-02-19 14:27:20.193097036 +0100
	+++ isspace3.s	2026-02-19 14:27:20.205097129 +0100
	@@ -1,4 +1,4 @@
	-	.file	"isspace2.c"
	+	.file	"isspace3.c"
		.text
		.section	.rodata.str1.1,"aMS",@progbits,1
	 .LC0:
	@@ -8,27 +8,25 @@
		.globl	isspace_c
		.type	isspace_c, @function
	 isspace_c:
	-.LFB0:
	+.LFB1:
		.cfi_startproc
	-	testl	%edi, %edi
	-	je	.L5
	+	movl	%edi, %esi
		subq	$8, %rsp
		.cfi_def_cfa_offset 16
	-	movl	%edi, %esi
		leaq	.LC0(%rip), %rdi
		call	strchr@PLT
	-	testq	%rax, %rax
	+	movq	%rax, %rdx
	+	xorl	%eax, %eax
	+	testq	%rdx, %rdx
	+	je	.L1
	+	cmpb	$0, (%rdx)
		setne	%al
	+.L1:
		addq	$8, %rsp
		.cfi_def_cfa_offset 8
		ret
	-	.p2align 4,,10
	-	.p2align 3
	-.L5:
	-	xorl	%eax, %eax
	-	ret
		.cfi_endproc
	-.LFE0:
	+.LFE1:
		.size	isspace_c, .-isspace_c
		.ident	"GCC: (Debian 15.2.0-13) 15.2.0"
		.section	.note.GNU-stack,"",@progbits

I expect the strchr(3) version would be faster, since the branch is
rarely used, and the main body is smaller.  However, I wouldn't worry
about performance in systems that don't implement strchrnul(3); those
are rare, and the difference would be small-ish.  Let the systems
implement strchrnul(3) if they care about performance.


Cheers,
Alex

-- 
<https://www.alejandro-colomar.es>
signature.asc (application/pgp-signature, 833 B)