AOCC Flang ICEs on computed goto in geom_hnd.F

[jeffhammond]

This is not a bug in NWChem, but I don't know where else to put it right now. As soon as I figure out whether it is an upstream Flang problem or an AOCC-specific problem, I'll close and redirect to the appropriate place.

A standalone reproducer is below. The problematic line of code is IF(N2) 280,270,120.

flang -m64 -ffast-math  -mcmodel=medium -mcmodel=medium -fno-backslash -Wuninitialized  -fopenmp -fdefault-integer-8 -finline-functions -fno-openmp -g -c geom_hnd_goto.F 
F90-F-0000-Internal compiler error. addili: illegal IL_TYPE(opc)       0  (geom_hnd_goto.F: 21)

flang -v
AOCC.LLVM.2.1.0.B1030.2019_11_12 clang version 9.0.0 (CLANG: Build#1030) (based on LLVM AOCC.LLVM.2.1.0.B1030.2019_11_12)
Target: x86_64-unknown-linux-gnu
Thread model: posix
InstalledDir: /nfs/pdx/home/jrhammon/ROME/aocc-compiler-2.1.0/bin
Found candidate GCC installation: /usr/lib/gcc/x86_64-redhat-linux/4.8.2
Found candidate GCC installation: /usr/lib/gcc/x86_64-redhat-linux/4.8.5
Selected GCC installation: /usr/lib/gcc/x86_64-redhat-linux/4.8.5
Candidate multilib: .;@m64
Candidate multilib: 32;@m32
Selected multilib: .;@m64

      SUBROUTINE HND_GIVDIA(A,VEC,EIG,IA,N,NDIM,
     1           w,gamma,beta,betasq,p,q,iposv,ivpos,iord)
      implicit none
      integer n,ndim,n1,n2,nr,ik
      integer IA(1)
      integer IPOSV(*),IVPOS(*),IORD(*)
      double precision A(1),VEC(NDIM,1),EIG(1)
      double precision W(*),GAMMA(*),BETA(*),BETASQ(*)
      double precision P(*),Q(*)
      double precision zero,pt5,one,two,rhosq
      double precision b,s,sum,sgn,temp,d,sqrts,g,a2,cosa,COSAP,DIF
      double precision DIA,QJ,PPBR,PP,PPBS,SINA2,SINA,R1,R2,R12,SHIFT
      double precision U,WTAW,WJ
      integer i1,i,il,ij,ii,j,itemp,jj,k,lv,l,m,nrr,np,npas,nv,nt
      DATA ZERO,PT5,ONE,TWO /0.0D+00,0.5D+00,1.0D+00,2.0D+00/
      DATA RHOSQ /1.0D-22/
      IF(N.EQ.0) GO TO 560
      N1=N-1
      N2=N-2
      GAMMA(1)=A(1)
      IF(N2) 280,270,120
  120 DO 260 NR=1,N2
      IK=IA(NR+1)+NR
      B=A(IK)
      S=ZERO
      DO 130 I=NR,N2
      IJ=IA(I+2)+NR
  130 S=S+A(IJ)**2
      A(IK)=ZERO
      IF(S.LE.ZERO) GO TO 250
      S=S+B*B
      SGN=+ONE
      IF(B.GE.ZERO) GO TO 160
      SGN=-ONE
  160 SQRTS= SQRT(S)
      D=SGN/(SQRTS+SQRTS)
      TEMP= SQRT(PT5+B*D)
      W(NR)=TEMP
      A(IK)=TEMP
      D=D/TEMP
      B=-SGN*SQRTS
      DO 170 I=NR,N2
      IJ=IA(I+2)+NR
      TEMP=D*A(IJ)
      W(I+1)=TEMP
  170 A(IJ)=TEMP
      WTAW=ZERO
      DO 220 I=NR,N1
      SUM=ZERO
      II=IA(I+1)
      DO 180 J=NR,I
      IJ=II+J+1
  180 SUM=SUM+A(IJ)*W(J)
      I1=I+1
      IF(N1.LT.I1) GO TO 210
      DO 200 J=I1,N1
      IJ=IA(J+1)+I+1
  200 SUM=SUM+A(IJ)*W(J)
  210 P(I)=SUM
  220 WTAW=WTAW+SUM*W(I)
      DO 230 I=NR,N1
  230 Q(I)=P(I)-WTAW*W(I)
      DO 240 J=NR,N1
      QJ=Q(J)
      WJ=W(J)
      JJ=J+1
      DO 240 I=J,N1
      IJ=IA(I+1)+JJ
  240 A(IJ)=A(IJ)-TWO*(W(I)*QJ+WJ*Q(I))
  250 BETA(NR)=B
      BETASQ(NR)=B*B
      IL=IK+1
  260 GAMMA(NR+1)=A(IL)
  270 IJ=IA(N)+N-1
      B=A(IJ)
      BETA(N-1)=B
      BETASQ(N-1)=B*B
      IJ=IJ+1
      GAMMA(N)=A(IJ)
  280 BETASQ(N)=ZERO
      DO 300 I=1,N
      DO 299 J=1,N
  299 VEC(I,J)=ZERO
  300 VEC(I,I)=ONE
      M=N
      SUM=ZERO
      NPAS=1
      GO TO 400
  310 SUM=SUM+SHIFT
      COSA=ONE
      G=GAMMA(1)-SHIFT
      PP=G
      PPBS=PP*PP+BETASQ(1)
      PPBR= SQRT(PPBS)
      DO 370 J=1,M
      COSAP=COSA
      IF(PPBS.NE.ZERO) GO TO 320
      SINA=ZERO
      SINA2=ZERO
      COSA=ONE
      GO TO 350
  320 SINA=BETA(J)/PPBR
      SINA2=BETASQ(J)/PPBS
      COSA=PP/PPBR
      NT=J+NPAS
      IF(NT.LT.N) GO TO 330
      NT=N
  330 CONTINUE
      DO 340 I=1,NT
      TEMP=COSA*VEC(I,J)+SINA*VEC(I,J+1)
      VEC(I,J+1)=-SINA*VEC(I,J)+COSA*VEC(I,J+1)
  340 VEC(I,J)=TEMP
  350 DIA=GAMMA(J+1)-SHIFT
      U=SINA2*(G+DIA)
      GAMMA(J)=G+U
      G=DIA-U
      PP=DIA*COSA-SINA*COSAP*BETA(J)
      IF(J.NE.M) GO TO 360
      BETA(J)=SINA*PP
      BETASQ(J)=SINA2*PP*PP
      GO TO 380
  360 PPBS=PP*PP+BETASQ(J+1)
      PPBR= SQRT(PPBS)
      BETA(J)=SINA*PPBR
  370 BETASQ(J)=SINA2*PPBS
  380 GAMMA(M+1)=G
      NPAS=NPAS+1
      IF(BETASQ(M).GT.RHOSQ) GO TO 410
  390 EIG(M+1)=GAMMA(M+1)+SUM
  400 BETA(M)=ZERO
      BETASQ(M)=ZERO
      M=M-1
      IF(M.EQ.0) GO TO 430
      IF(BETASQ(M).LE.RHOSQ) GO TO 390
  410 A2=GAMMA(M+1)
      R2=PT5*A2
      R1=PT5*GAMMA(M)
      R12=R1+R2
      DIF=R1-R2
      TEMP= SQRT(DIF*DIF+BETASQ(M))
      R1=R12+TEMP
      R2=R12-TEMP
      DIF=  ABS(A2-R1)-  ABS(A2-R2)
      IF(DIF.LT.ZERO) GO TO 420
      SHIFT=R2
      GO TO 310
  420 SHIFT=R1
      GO TO 310
  430 EIG(1)=GAMMA(1)+SUM
      DO 440 J=1,N
      IPOSV(J)=J
      IVPOS(J)=J
  440  IORD(J)=J
      M=N
      GO TO 470
  450 DO 460 J=1,M
      IF(EIG(J).LE.EIG(J+1)) GO TO 460
      TEMP=EIG(J)
      EIG(J)=EIG(J+1)
      EIG(J+1)=TEMP
      ITEMP=IORD(J)
      IORD(J)=IORD(J+1)
      IORD(J+1)=ITEMP
  460 CONTINUE
  470 M=M-1
      IF(M.NE.0) GO TO 450
      IF(N1.EQ.0) GO TO 500
      DO 490 L=1,N1
      NV=IORD(L)
      NP=IPOSV(NV)
      IF(NP.EQ.L) GO TO 490
      LV=IVPOS(L)
      IVPOS(NP)=LV
      IPOSV(LV)=NP
      DO 480 I=1,N
      TEMP=VEC(I,L)
      VEC(I,L)=VEC(I,NP)
  480 VEC(I,NP)=TEMP
  490 CONTINUE
  500 DO 550 NRR=1,N
      K=N1
  510 K=K-1
      IF(K.LE.0) GO TO 550
      SUM=ZERO
      DO 520 I=K,N1
      IJ=IA(I+1)+K
  520 SUM=SUM+VEC(I+1,NRR)*A(IJ)
      SUM=SUM+SUM
      DO 530 I=K,N1
      IJ=IA(I+1)+K
  530 VEC(I+1,NRR)=VEC(I+1,NRR)-SUM*A(IJ)
      GO TO 510
  550 CONTINUE
  560 CONTINUE
      RETURN
      END

[jeffhammond]

@dmcdougall Do you know where I should send this bug report? I don't have a recent upstream Flang installed on any of my machines, so I would prefer to file a bug against AOCC Flang and let them figure it out.

[edoapra]

@jeffhammond This is a know bug in AOCC 2.1
flang-compiler/flang#844

I have already introduced work-arounds in master and 7.0.0
cfcd024
aca6e0f

[dmcdougall]

@edoapra Thanks for digging that ticket up.

@jeffhammond It looks like the issue is already known. Let me know if I can do anything to help.

[jeffhammond]

Damon: NWChem is the only code I care about that needs 64-bit INTEGER types so if Edo has a fix for this, then there is no urgency to update AOCC. Thanks.

[dmcdougall]

Understood.

[edoapra]

@jeffhammond I am inclined to close this issue. Any reason to keep it open?
Thanks

[jeffhammond]

Please close. Thanks!