GCC "switch" - When does GCC create a jumptable ?

[Frank B]

Hi,

I have a big switch () {..}:

unsigned i;
....
switch(i) {
case 0: ...break;
case 1: ...break;
...
case 63: ...break;
}

64 cases, no default, no holes. I thought, GCC creates a jumptable, but it does not.
Which criteria must be met ?
Is there a way to force it ?

I'm using -O3 with LTO.

 

[tni]

It usually does. Maybe your code has a fair amount of overlap and GCC merges those parts. You can try using O1 / O2 / Os - that may change code generation.

You can create a jump table manually with computed gotos. It's a GCC extension, not standard C or C++. It's mostly supported by Clang as well.

 

[Frank B]

Yes, there is some overlapping.
The "&&" is new to me - Thank you
The other way would have been to create (nearly) 64 functions and an array with functionpointers.

 

[MichaelMeissner]

There are various knobs that the backend can control to influence how switch tables or if-statements are generated. I don't know what decisions the arm maintainers have made for this. On the PowerPC, we tend to prefer more if/then/elses than a table jump for small numbers of case entries, due to the way the PowerPC handles indirect jumps.

There is a parameter that you can control via command line that biases the decision, and an option to enable/disable jump tables if the table is dense enough:

• --param case-values-threshold=<n>: The smallest number of different values for which it is best to use a jump-table instead of a tree of conditional branches, if 0, use the default for the machine.
• -fjump-table: Use jump tables for sufficiently large switch statements.
• -fno-jump-table: Do not use jump tables for sufficiently large switch statements.