[Bug target/68837] PowerPC switch statement performance

[Bug target/68837] PowerPC switch statement performance

[guihaoc at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68837

HaoChen Gui <guihaoc at gcc dot gnu.org> changed:

           What    |Removed                     |Added
----------------------------------------------------------------------------
                 CC|                            |guihaoc at gcc dot gnu.org

--- Comment #2 from HaoChen Gui <guihaoc at gcc dot gnu.org> ---
(In reply to David Edelsohn from comment #0)
> Improve performance of switch statements:
> 
> 1) Heuristics for decision tree vs tablejump
> 
> 2) Avoid sign extended lwa for offset

I think #1 was already implemented in current GC. Jump tables coexists with
conditional jumps. jump-table-max-growth-ratio-for-speed/size decide how large
a jump table could be and case-values-threshold defines if a jump table is
beneficial. 

An example of GIMPLE switch
;; GIMPLE switch case clusters: 35 37 JT(values:5 comparisons:5 range:8
density: 62.50%):65-72 JT(values:6 comparisons:6 range:12 density:
50.00%):111-122

For #2, the offset could be negative in a multiple jump table case. Right now
it uses lwax and there is no overhead for sing extend, I think. Please correct
me if I am wrong.

[Bug target/68837] PowerPC switch statement performance

[dje at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68837

--- Comment #3 from David Edelsohn <dje at gcc dot gnu.org> ---
The PR was from 2015, before Martin's improvements.

Also, sign-extend load instructions were less efficient at the time.  We need
to re-examine the sequence on more recent microarchitectures.

[Bug target/68837] PowerPC switch statement performance

[segher at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68837

--- Comment #4 from Segher Boessenkool <segher at gcc dot gnu.org> ---
On Power9 the lwa insn is cracked into an lwz and an extsw, just like
on older CPUs.  Cracked instructions have fewer constraints on p9 than
they did on most older CPUs though (it doesn't have to be first in a
fetch group any more).

It still saves one insn to execute if lwz would be enough.  Another
option may be to just do an ld even: this saves the number of insns
needed, but it takes more space for the table (so it won't be cached /
prefetched as well), and absolute addresses require relocations.

[Bug target/68837] PowerPC switch statement performance

[guihaoc at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=68837

--- Comment #5 from HaoChen Gui <guihaoc at gcc dot gnu.org> ---
I think there are two ways avoiding sign extension for offset loading.

a. Make sure all offsets be positive.  There exists backward jumps as well as
STC will reorder the basic block. So the offset might be negative. We can check
the offsets after bb-reorder pass and remove sign extension if all offsets are
positive. It may need a new pass.

b. Use absolute addresses instead of offsets. It's controlled by macro
CASE_VECTOR_PC_RELATIVE and flag_pic. x86 uses absolute address(labels) in
their jump table, though the size of jump is bigger. Power haven't supported
absolute address yet.