[flang][hlfir] Polyhedron/nf 23% performance regression

[vzakhari]

Code with HLFIR lowering runs for 7.4 seconds vs 6 seconds with FIR lowering.

There is some overhead due to extra temporaries at line 261:

256	subroutine NF2DPrecon(x,i1,i2)       ! 2D NF Preconditioning matrix				
257	integer :: i1 , i2				
258	real(dpkind),dimension(i2)::x,t				
259	integer :: i				
260	do i = i1 , i2 , nx				
261	   if ( i>i1 ) x(i:i+nx-1) = x(i:i+nx-1) - au2(i-nx:i-1)*x(i-nx:i-1)
262	   call trisolve(x,i,i+nx-1)				
263	enddo 				
264	do i = i2-2*nx+1 , i1 , -nx				
265	   t(i:i+nx-1) = au2(i:i+nx-1)*x(i+nx:i+2*nx-1)				
266	   call trisolve(t,i,i+nx-1)				
267	   x(i:i+nx-1) = x(i:i+nx-1) - t(i:i+nx-1)				
268	enddo				
269	end subroutine NF2DPrecon            !=========================================				

ArrayValueCopy has special handling for array slices of the form (i:j) and (j+1:k), which allows disambiguating x(i:i+nx-1) with x(i-nx:i-1). We can probably do the same in the optimized bufferization pass or implement something more generic. For example, we can try to use the affine dialect utilities to detect store-load conflicts based on the iteration space constraints derived from the slices configurations and the mapping of the iteration indices to the memory locations (based on the designator indexing).

[llvmbot]

@llvm/issue-subscribers-flang-ir

[vzakhari]

The array sections no-overlap detection merged in #65707

More changes are needed in the alias analysis to disambiguate dummy and host-associated accesses.

[vzakhari]

Fixed by #65919