OMPOverwrite¶

The iterations of this region assign every element of an array before they read it, when a condition holds.

Declaration¶

Syntax¶

OMPOverwrite(expr array, expr condition)

Arguments¶

Argument name	Argument type	Argument description
`array`	`expr`	The array, a variable or a component of one.
`condition`	`expr`	A logical host expression, evaluated on entry to the region.

Return values¶

None.

Description¶

When condition holds on entry to a parallel region assigned to the device, its iterations together assign every element of array before any of them reads one, so what the host holds in array need not be copied to the device. The offload hands this to the launch as the overwritten and overwritten_if of a GpuKernelLaunch.

The offload shows this itself for a loop whose assignment has the loop variables as subscripts. A loop transformation that rewrites such a loop into a form the offload cannot read it from, but that runs the same iterations and assignments, records it from the loop as it was: index-set splitting does, since the frame it adds runs the original body on indices it computes. A pass that changes which iterations the region runs drops the clause, and so does the offload when it runs the region on one thread inside device code, where there is no launch.

The source never writes it.

Examples¶

(OMPOverwrite
    (Var 2 d)
    (LogicalBinOp
        (IntegerCompare (Cast (IntegerConstant 1 (Integer 4) Decimal) IntegerToInteger (Integer 8) () ()) Eq (ArrayBound (Var 2 d) (IntegerConstant 1 (Integer 4) Decimal) (Integer 8) LBound ()) (Logical 4) ())
        And
        (IntegerCompare (Cast (Var 2 n) IntegerToInteger (Integer 8) () ()) Eq (ArrayBound (Var 2 d) (IntegerConstant 1 (Integer 4) Decimal) (Integer 8) UBound ()) (Logical 4) ())
        (Logical 4)
        ()
    )
)