[Bug tree-optimization/92283] [10 Regression] 454.calculix miscomparison since r276645 with -O2 -march=znver2

[rguenth at gcc dot gnu.org]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=92283

Richard Biener <rguenth at gcc dot gnu.org> changed:

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

--- Comment #18 from Richard Biener <rguenth at gcc dot gnu.org> ---
The fortran loop is

   real*8  vold(0:3,*),v(0:3,*)
...

            if((iperturb.ne.0).and.(nmethod.eq.1)) then
               vmax=0.d0
               do i=1,nk
                  do j=jmin,jmax
                     vmax=max(vmax,dabs(v(j,i)))
c                     write(*,*) i,j,vold(j,i),v(j,i),vmax
                     v(j,i)=v(j,i)+vold(j,i)
                  enddo
               enddo
            endif

and the .gimple diff

@@ -2536,21 +2970,29 @@
   vmax__lsm.757_3905 = [results.f:130:0] *vmax_2002(D);
   [results.f:130:0] _31 = (integer(kind=8)) i.120_4019;
   [results.f:130:0] _32 = _31 * 4;
+  ivtmp.773_566 = (sizetype) jmin_1722;

   <bb 20> [local count: 4326176]:
-  # j.121_4017 = PHI <[results.f:129:0] jmin_1722(19), [results.f:129:0]
_39(670)>
+  # ivtmp.773_3273 = PHI <ivtmp.773_566(19), ivtmp.773_1666(670)>
   # vmax__lsm.757_1845 = PHI <vmax__lsm.757_3905(19), [results.f:130:0]
M.64_2048(670)>
-  [results.f:130:0] _33 = (integer(kind=8)) j.121_4017;
-  [results.f:130:0] _34 = _32 + _33;
-  [results.f:130:0] _35 = _34 + -4;
-  [results.f:130:0] _36 = [results.f:130:0] [results.f:130:0]
(*v_2012(D))[_35];
+  _2429 = (sizetype) _32;
+  _565 = _2429 * 8;
+  _4606 = v_2012(D) + _565;
+  [results.f:130:0] _36 = MEM[base: _4606, index: ivtmp.773_3273, step: 8,
offset: -32B];
   [results.f:130:0] _2047 = ABS_EXPR <_36>;
   [results.f:130:0] M.64_2048 = MAX_EXPR <vmax__lsm.757_1845, _2047>;
-  [results.f:132:0] _37 = [results.f:132:0] [results.f:132:0]
(*vold_2011(D))[_35];
+  _4529 = (sizetype) _32;
+  _558 = _4529 * 8;
+  _4481 = vold_2011(D) + _558;
+  [results.f:132:0] _37 = MEM[base: _4481, index: ivtmp.773_3273, step: 8,
offset: -32B];
   [results.f:132:0] _38 = _36 + _37;
-  [results.f:132:0] [results.f:132:0] [results.f:132:0] (*v_2012(D))[_35] =
_38;
-  [results.f:129:0] _39 = j.121_4017 + 1;
-  [results.f:129:0] if (_39 > jmax_1721)
+  _564 = (sizetype) _32;
+  _4573 = _564 * 8;
+  _560 = v_2012(D) + _4573;
+  [results.f:132:0] MEM[base: _560, index: ivtmp.773_3273, step: 8, offset:
-32B] = _38;
+  [results.f:129:0] ivtmp.773_1666 = ivtmp.773_3273 + 1;
+  _3693 = (integer(kind=4)) ivtmp.773_1666;
+  [results.f:129:0] if (_3693 > jmax_1721)
     goto <bb 21>; [11.00%]
   else
     goto <bb 670>; [89.00%]

but as far as I can see the IVOPTs transform is correct, assembly differences
are too large to swallow though (trying to disable scheduling).