Well, what I mean is that I dont see how in C the conditional is only evaluated once. I dont know enough about how C implements a case statement to say for sure But i dont see how a case statement could be implemented in such a way without adding a lot of code to the statement, which would potentially be far more expensive than converting it to a series of conditionals. For instance something like this

switch (val) {
  case 1 : handle_case1; break;
  case 2 : handle_case2; break;
  case 5 : handle_case5;
  case 10: handle_case10;
  default: handle_default; 
}
[download]

I would guess would get converted to something like

if (val == 1) {
  handle_case1;
} elsif (val == 2) {
  handle_case2;
} else {
  if (val == 5) {
    goto CASE5;
  } elsif (val == 10) {
    goto CASE10;
  } else {
    goto DEFAULT;
  }
CASE5:   handle_case5;
CASE10:  handle_case10;
DEFAULT: handle_default;
}
[download]

Or something like it. I just dont see how it could be handled otherwise. Perhaps the fact that it only operates on ints means that there is a neater optimisation. But if you extend switch to handle strings as it does in pascal then I think the situation becomes even more difficult.

Thanks to the CB for clarifying that switch only handles ints. That issue may make my ruminations on this invalid. I dont know. Id be interested to hear from someone that does. As I said, this issue confuses me. :-)

No. On every implementation I've seen, switch/case is implemented almost exactly like perl's goto-EXPR, except that the target label is actually a machine instruction address known at compile time.

This simple C program

#include <stdio.h>

int main() {

    switch ( 4 ) {
        case 1: printf( "case1\n" ); break;
        case 2: printf( "case2\n" ); break;
        case 3: printf( "case3\n" ); break;
        case 4: printf( "case4\n" ); break;
        case 5: printf( "case5\n" ); break;
        case 6: printf( "case6\n" ); break;
    }
}
[download]

Is translated into this assembler

Read more... (3 kB)

The salient part of which is

   ;        switch ( 4 ) {
   ;    
@1:
    mov       eax,4 // Load the selector expression (4)
    cmp       eax,6 // Test if its outside the range of options (6)
    ja        short @2 // If it is, jump to the default
/*
    Otherwise, add the value * the size of the lookup table  
    entries (4bytes) and add it to the base address of the
    dispatch table, then jump to the address help at that
    location in the table.
*/
    jmp       dword ptr [@10+4*eax]
@10:
    dd        @2
    dd        @9
    dd        @8
    dd        @7
    dd        @6
    dd        @5
    dd        @4
   ;    
   ;            case 1: printf( "case1\n" ); break;
   ;    
@9:
[download]

Originally, the offsets were byte offsets and the table size had a maximum of 255. These days, the table entries are absolute 32-bit addresses and the table size can theoretically be 2 GB in size. In all cases, it is very fast.

The interesting part is coding the switch expression so that it converts your range of possible matches to an integer. String lookups can be index using str(i)(n)cmp() quite easily, but I've had to code some more esoteric versions in the past.

---

Examine what is said, not who speaks.
"Efficiency is intelligent laziness." -David Dunham
"Think for yourself!" - Abigail
Hooray!

So would I be correct in understanding that

switch (ulong_val) {
  case 0       : handle_case0; break;
  case ULONGMAX: handle_caseMAX; break;
  default      : handle_default; break;
}
[download]

Generates a _huge_ dispatch table? Can that be correct? I found your dissambly to be very interesting, but I personaly would have like to have seen what happens when the input is discontiguous and operated on a variable and not a constant. Any chance of an update BrowserUk?



---
demerphq



First they ignore you, then they laugh at you, then they fight you, then you win.
-- Gandhi