www.digitalmars.com         C & C++   DMDScript  

digitalmars.D - array traversal

reply arun <1986.arun gmail.com> writes:
i want to know during array traversal ,whether pointers do a better job than
indices by considering the code snippet

int [10]array;
foreach( value;array) {
   //something
}

i heard that implementations decides which one to use.i want to know how those
two traversal differs? 
Mar 20 2007
next sibling parent Sean Kelly <sean f4.ca> writes:
arun wrote:
 i want to know during array traversal ,whether pointers do a
 better job than indices
It depends on the platform and on the compiler. Nowadays though, there generally isn't enough of a performance difference to matter either way. Just write the code that looks the cleanest. Sean
Mar 20 2007
prev sibling parent reply Tyler Knott <tywebmail mailcity.com> writes:
arun wrote:
 i want to know during array traversal ,whether pointers do a better job than
indices by considering the code snippet
 
 int [10]array;
 foreach( value;array) {
    //something
 }
 
 i heard that implementations decides which one to use.i want to know how those
two traversal differs? 
 
 
At least on x86/x64 they shouldn't differ. They should both compile down to the pseudo-code x86 instruction mov value,[array.ptr+sizeof(int)*index] which is read as "move the value at the address (array.ptr+size(int)*index) into value" using flat pointer arithmetic (i.e. not C-style; use 1-byte increments always, even for types where sizeof(type) > 1 byte). The above examples uses a bit of pseudo-code. The real version uses a few more instructions to make sure that index is within array's bounds (in debug mode) and temporarily store array.ptr and index in a CPU registers.
Mar 20 2007
parent reply BCS <ao pathlink.com> writes:
Reply to Tyler,

 arun wrote:
 
 i want to know during array traversal ,whether pointers do a better
 job than indices by considering the code snippet
 
 int [10]array;
 foreach( value;array) {
 //something
 }
 i heard that implementations decides which one to use.i want to know
 how those two traversal differs?
 
At least on x86/x64 they shouldn't differ. They should both compile down to the pseudo-code x86 instruction mov value,[array.ptr+sizeof(int)*index] which is read as "move the value at the address (array.ptr+size(int)*index) into value" using flat pointer arithmetic (i.e. not C-style; use 1-byte increments always, even for types where sizeof(type) > 1 byte). The above examples uses a bit of pseudo-code. The real version uses a few more instructions to make sure that index is within array's bounds (in debug mode) and temporarily store array.ptr and index in a CPU registers.
I think that options he is taking about are these for(T* ptr = &start; ptr !is &stop; ptr++) { T value = *ptr } vs. for(int i = 0; i< length; i++) { T value = ptr[i]; } only the second ever uses a multiplication
Mar 20 2007
parent reply "Stewart Gordon" <smjg_1998 yahoo.com> writes:
"BCS" <ao pathlink.com> wrote in message 
news:ce0a3343870c8c939107e7e1a1e news.digitalmars.com...
<snip>
 I think that options he is taking about are these
 for(T* ptr = &start; ptr !is &stop; ptr++)
 {
  T value = *ptr
 }

 vs.

 for(int i = 0; i< length; i++)
 {
   T value = ptr[i];
 }

 only the second ever uses a multiplication
And even then not necessarily - some compilers may optimise one form to the other. Stewart.
Mar 28 2007
parent reply Dan <murpsoft hotmail.com> writes:
Stewart Gordon Wrote:

 "BCS" <ao pathlink.com> wrote in message 
 news:ce0a3343870c8c939107e7e1a1e news.digitalmars.com...
 <snip>
 I think that options he is taking about are these
 for(T* ptr = &start; ptr !is &stop; ptr++)
 {
  T value = *ptr
 }

 vs.

 for(int i = 0; i< length; i++)
 {
   T value = ptr[i];
 }

 only the second ever uses a multiplication
And even then not necessarily - some compilers may optimise one form to the other. Stewart.
Hmm... let me think of how that looks in ASM. The first one is wrong, because you're only ptr++, it should be ptr += OFFSET. Apart from that, you're right the first code looks mildly better in ASM. Looping through by adding to the pointer instead of adding to the array index and performing a 'lea'. The good news is that D *should be* optimizing this out, and it ultimately only costs a half cycle which may either align or throw off the u/v pipes - something more important than 1/2 a cycle. I wonder if Walter has it aligning the pipes as appropriate?
Mar 28 2007
next sibling parent reply "Stewart Gordon" <smjg_1998 yahoo.com> writes:
"Dan" <murpsoft hotmail.com> wrote in message 
news:euecda$16oq$1 digitalmars.com...
 Stewart Gordon Wrote:

 "BCS" <ao pathlink.com> wrote in message
 news:ce0a3343870c8c939107e7e1a1e news.digitalmars.com...
 <snip>
 I think that options he is taking about are these
 for(T* ptr = &start; ptr !is &stop; ptr++)
 {
  T value = *ptr
 }
<snip>
 Hmm...  let me think of how that looks in ASM.  The first one
 is wrong, because you're only ptr++, it should be ptr +=
 OFFSET.  Apart from that, you're right the first code looks
 mildly better in ASM.
No it isn't. The ++ operator (and similarly --, +, -, +=, -=) applied to a pointer works in the units of the size of the data type it points to.
 Looping through by adding to the pointer instead of adding to
 the array index and performing a 'lea'.

 The good news is that D *should be* optimizing this out, and
 it ultimately only costs a half cycle which may either align
 or throw off the u/v pipes - something more important than 1/2
 a cycle.
What on earth is a u/v pipe? Stewart.
Mar 28 2007
parent Ary Manzana <ary esperanto.org.ar> writes:
Stewart Gordon escribió:
 "Dan" <murpsoft hotmail.com> wrote in message 
 news:euecda$16oq$1 digitalmars.com...
 Stewart Gordon Wrote:

 "BCS" <ao pathlink.com> wrote in message
 news:ce0a3343870c8c939107e7e1a1e news.digitalmars.com...
 <snip>
 I think that options he is taking about are these
 for(T* ptr = &start; ptr !is &stop; ptr++)
 {
  T value = *ptr
 }
<snip>
 Hmm...  let me think of how that looks in ASM.  The first one
 is wrong, because you're only ptr++, it should be ptr +=
 OFFSET.  Apart from that, you're right the first code looks
 mildly better in ASM.
No it isn't. The ++ operator (and similarly --, +, -, +=, -=) applied to a pointer works in the units of the size of the data type it points to.
 Looping through by adding to the pointer instead of adding to
 the array index and performing a 'lea'.

 The good news is that D *should be* optimizing this out, and
 it ultimately only costs a half cycle which may either align
 or throw off the u/v pipes - something more important than 1/2
 a cycle.
What on earth is a u/v pipe? Stewart.
http://en.wikipedia.org/wiki/Pentium (read: Major changes from the 486)
Mar 28 2007
prev sibling parent Don Clugston <dac nospam.com.au> writes:
Dan wrote:
 Stewart Gordon Wrote:
 
 "BCS" <ao pathlink.com> wrote in message 
 news:ce0a3343870c8c939107e7e1a1e news.digitalmars.com...
 <snip>
 I think that options he is taking about are these
 for(T* ptr = &start; ptr !is &stop; ptr++)
 {
  T value = *ptr
 }

 vs.

 for(int i = 0; i< length; i++)
 {
   T value = ptr[i];
 }

 only the second ever uses a multiplication
And even then not necessarily - some compilers may optimise one form to the other. Stewart.
Hmm... let me think of how that looks in ASM. The first one is wrong, because you're only ptr++, it should be ptr += OFFSET. Apart from that, you're right the first code looks mildly better in ASM. Looping through by adding to the pointer instead of adding to the array index and performing a 'lea'.
But if T.sizeof == 2, 4, or 8, the multiplication can be done in hardware for free. EG if i is stored in esi, mov eax, [ptr + esi*8]; I'd be surprised if there are any modern compilers that actually perform a multiply.
 
 The good news is that D *should be* optimizing this out, and it ultimately
only costs a half cycle which may either align or throw off the u/v pipes -
something more important than 1/2 a cycle.
 
 I wonder if Walter has it aligning the pipes as appropriate?
BTW, the u-v pipe thing is not relevant for recent Pentiums any more. Core2 CPUs are more likely to be limited by the decoding stage than by the execution units.
Mar 30 2007