Hi Wilco,

On 12/23/22 19:35, Wilco Dijkstra wrote:
> Hi Alex,

(a)

> 
>>        if (dst == end)
>>            return end;
>>        if (stp_unlikely(dst == NULL))  // Allow chaining with stpeprintf().
>>            return NULL;
> 
>> Oh, and the two branches above can be optimized into a branch that returns dst.
> 
> How? There will be 2 branches since you're doing 2 checks here...

That is equivalent to:

(b)

         if (dst == end)
             return dst;
         if (stp_unlikely(dst == NULL))  // Allow chaining with stpeprintf().
             return dst;

which itself is equivalent to:

(c)

         if ((dst == end) || stp_unlikely(dst == NULL))
             return dst;

which still has a branch in ||, due to the shortcut of the boolean operator.
However, the compiler is allowed to transform it into bitwise, since there are 
no side effects, no UB, and the result would be the same:

(d)

         if ((dst == end) | stp_unlikely(dst == NULL))
             return dst;

Which doesn't have a hidden branch.

I tried GCC, and (b) and (c) produce the same assembly code, but slightly 
different than (a) (but almost identical, nothing significant).  (d) produces 
considerably different assembly.  Tried under -O3 -march=native.  I don't know 
enough assembly to judge which is better; I'll copy the results here for the 
curious.  I guess the compiler seems to prefer an extra branch here over 
unconditionally doing bitwise operations; it very likely knows more than I do.

Cheers,

Alex



alx@asus5775:~/src/alx/libstp$ diff -u a.s b.s
--- a.s	2022-12-23 23:27:57.788103834 +0100
+++ b.s	2022-12-23 23:28:13.463919271 +0100
@@ -59,9 +59,9 @@
  	.cfi_offset 3, -16
  	movq	%rsi, %rbx
  	cmpq	%rsi, %rdi
-	je	.L9
-	testq	%rdi, %rdi
  	je	.L12
+	testq	%rdi, %rdi
+	je	.L13
  	movq	%rbx, %rcx
  	movq	%rdx, %rsi
  	xorl	%edx, %edx
@@ -79,7 +79,7 @@
  .L11:
  	.cfi_restore_state
  	movb	$0, -1(%rbx)
-.L9:
+.L12:
  	movq	%rbx, %rax
  	popq	%rbx
  	.cfi_remember_state
@@ -87,7 +87,7 @@
  	ret
  	.p2align 4,,10
  	.p2align 3
-.L12:
+.L13:
  	.cfi_restore_state
  	xorl	%eax, %eax
  	popq	%rbx
alx@asus5775:~/src/alx/libstp$ diff -u b.s c.s
alx@asus5775:~/src/alx/libstp$ diff -u c.s d.s
--- c.s	2022-12-23 23:29:07.315367548 +0100
+++ d.s	2022-12-23 23:28:58.007455133 +0100
@@ -11,16 +11,16 @@
  	.cfi_offset 3, -16
  	movq	%rsi, %rbx
  	cmpq	%rsi, %rdi
-	je	.L2
+	je	.L4
  	testq	%rdi, %rdi
-	je	.L5
+	je	.L4
  	movq	%rbx, %rcx
  	movq	%rdx, %rsi
  	xorl	%edx, %edx
  	subq	%rdi, %rcx
  	call	memccpy@PLT
  	testq	%rax, %rax
-	je	.L4
+	je	.L3
  	decq	%rax
  	popq	%rbx
  	.cfi_remember_state
@@ -28,20 +28,19 @@
  	ret
  	.p2align 4,,10
  	.p2align 3
-.L4:
+.L3:
  	.cfi_restore_state
-	movb	$0, -1(%rbx)
-.L2:
  	movq	%rbx, %rax
+	movb	$0, -1(%rbx)
  	popq	%rbx
  	.cfi_remember_state
  	.cfi_def_cfa_offset 8
  	ret
  	.p2align 4,,10
  	.p2align 3
-.L5:
+.L4:
  	.cfi_restore_state
-	xorl	%eax, %eax
+	movq	%rdi, %rax
  	popq	%rbx
  	.cfi_def_cfa_offset 8
  	ret
@@ -59,16 +58,16 @@
  	.cfi_offset 3, -16
  	movq	%rsi, %rbx
  	cmpq	%rsi, %rdi
-	je	.L12
+	je	.L10
  	testq	%rdi, %rdi
-	je	.L13
+	je	.L10
  	movq	%rbx, %rcx
  	movq	%rdx, %rsi
  	xorl	%edx, %edx
  	subq	%rdi, %rcx
  	call	memccpy@PLT
  	testq	%rax, %rax
-	je	.L11
+	je	.L9
  	decq	%rax
  	popq	%rbx
  	.cfi_remember_state
@@ -76,20 +75,19 @@
  	ret
  	.p2align 4,,10
  	.p2align 3
-.L11:
+.L9:
  	.cfi_restore_state
-	movb	$0, -1(%rbx)
-.L12:
  	movq	%rbx, %rax
+	movb	$0, -1(%rbx)
  	popq	%rbx
  	.cfi_remember_state
  	.cfi_def_cfa_offset 8
  	ret
  	.p2align 4,,10
  	.p2align 3
-.L13:
+.L10:
  	.cfi_restore_state
-	xorl	%eax, %eax
+	movq	%rdi, %rax
  	popq	%rbx
  	.cfi_def_cfa_offset 8
  	ret


> 
> Cheers,
> Wilco

-- 
<http://www.alejandro-colomar.es/>