Hello,
I am seeking your opinion on what is the best way to solve a problem that I
observed.
Compiler assumes that formal parameters based on different symbols in Fortran
programs are not aliased. This leads to a runtime error for CPU2006 416.gamess
when "-Ofast -LNO:fusion=2" is used.
In the following code snippet, formal parameter array "D" and "DO" in
subroutine “SYMTRD” have the same storage as their corresponding actual
parameter "X(I40)". Therefore the read of "D" in the 1st loop must alias with
the write of "DO" in the 2nd loop. When the two outermost loops are fused
together, we got incorrect results.
In summary, the Fortran parameter no-alias rule makes loop transformations
illegal.
For legality concern, we could rely on interprocedual analysis to track whether
formal parameters have the same base. But compiler analysis tends to be
conservative and therefore could hurt performance. It might not always be
feasible to implement "copy-in copy-out" which may hurt performance as well.
Will that be reasonable to treat this as an user error?
Code snippet:
callsite:
CALL SYMTRD(X(I30),X(I40),X(I10),X(I40),X(I60),IA,L1,L1)
callee:
SUBROUTINE SYMTRD(SH,D,DT,DO,DOD,IA,N,NDIM)
DIMENSION SH(*),D(*),DT(NDIM,*),DO(*),DOD(*),IA(*)
......
DO 140 I = 1,N
DO 120 J = 1,N
DTEMP = ZERO
DO 100 K = 1,N
IK = IA(I)+K
IF (I .LT. K) IK = IA(K)+I
JK = IA(J)+K
IF (J .LT. K) JK = IA(K)+J
DTEMP = DTEMP+SH(IK)*D(JK) // read of D
100 CONTINUE
DT(I,J) = DTEMP
120 CONTINUE
140 CONTINUE
IJ = 0
DO 200 I = 1,N
DO 180 J = 1,I
IJ = IJ+1
DTEMP = ZERO
DO 160 K = 1,N
JK = IA(J)+K
IF (J .LT. K) JK = IA(K)+J
DTEMP = DTEMP+DT(I,K)*SH(JK)
160 CONTINUE
DO(IJ) = DTEMP // write of DO
180 CONTINUE
200 CONTINUE
-Mei Ye
------------------------------------------------------------------------------
Special Offer-- Download ArcSight Logger for FREE (a $49 USD value)!
Finally, a world-class log management solution at an even better price-free!
Download using promo code Free_Logger_4_Dev2Dev. Offer expires
February 28th, so secure your free ArcSight Logger TODAY!
http://p.sf.net/sfu/arcsight-sfd2d
_______________________________________________
Open64-devel mailing list
Open64-devel@lists.sourceforge.net
https://lists.sourceforge.net/lists/listinfo/open64-devel
