[Bug tree-optimization/114403] [14 regression] LLVM miscompiled with -O3 -march=znver2 -fno-vect-cost-model since r14-6822-g01f4251b8775c8

["tnfchris at gcc dot gnu.org" <gcc-bugzilla@gcc.gnu.org>]

https://gcc.gnu.org/bugzilla/show_bug.cgi?id=114403

--- Comment #21 from Tamar Christina <tnfchris at gcc dot gnu.org> ---
Created attachment 57932
  --> https://gcc.gnu.org/bugzilla/attachment.cgi?id=57932&action=edit
loop.c

attached reduced testcase that reproduces the issue and also checks the buffer
position and copied values.

As discussed on IRC when peeling for gaps we need to either adjust the upper
bounds of the vector loop or force the vector loop to get to the scalar loop. 
However we already go to the scalar loop, just with the wrong induction value
because we were never supposed to take the main exit.

whether go to the scalar loop depends on
x = (((ptr2 - ptr1) - 16) / 16) + 1
x == (((x - 1) >> 2) << 2)

in this case x == 26, so we do go to the scalar code already, but through the
main exit.

exiting through the main exit assumes you've done all vector iterations, in
this case 6 iterations based on the main exit condition which is first != last.

In this case the inductions values will be set on niters_vector_mult.

so in this case first += 24

But that's wrong since the secondary exit has a known iteration count of 9, due
to (buffer_ptr + store_size) <= buffer_end.

Statement (exit)if (ivtmp_21 != 0)
 is executed at most 8 (bounded by 8) + 1 times in loop 1.

So we will always exit through it as 9 < 24.

that means that when we calculate the upper bounds of the vector loop, we must
add a bias so that in this boundary condition that we do an extra partial
vector iteration.

I think the discussion on IRC went off track for a bit and hopefully this
testcase and the explanation above shows that for all early break and all
epilogue peeling reasons, we must bias up for the upper bound to give the
secondary exits a chance to trigger.

So really do think the correct patch is:

diff --git a/gcc/tree-vect-loop.cc b/gcc/tree-vect-loop.cc
index 4375ebdcb49..0973b952c70 100644
--- a/gcc/tree-vect-loop.cc
+++ b/gcc/tree-vect-loop.cc
@@ -12144,6 +12144,9 @@ vect_transform_loop (loop_vec_info loop_vinfo, gimple
*loop_vectorized_call)
      -min_epilogue_iters to remove iterations that cannot be performed
        by the vector code.  */
   int bias_for_lowest = 1 - min_epilogue_iters;
+  if (LOOP_VINFO_EARLY_BREAKS (loop_vinfo))
+    bias_for_lowest = 1;
+
   int bias_for_assumed = bias_for_lowest;
   int alignment_npeels = LOOP_VINFO_PEELING_FOR_ALIGNMENT (loop_vinfo);
   if (alignment_npeels && LOOP_VINFO_USING_PARTIAL_VECTORS_P (loop_vinfo))

for the reasons described above.  There's no way for us to take the main exit,
which signifies (we've reached the end of all iterations we can possibly do as
vector) and get the correct induction values in this case.