The following patch avoids keeping the inherited MEM_ATTRS when
set_mem_attributes_minus_bitpos is called with a variable ARRAY_REF.
The inherited ones may not reflect the correct offset and neither
does the updated alias-set match the inherited MEM_EXPR.  This all
ends up confusing path-based alias-analysis, causing wrong-code.

The fix is to stop not adopting a MEM_EXPR for certain kinds of
expressions and instead handle everything we can.  There's still
the constant kind trees case which I'm too lazy to look into right
now.  I did refrain from adding SSA_NAME there and instead avoided
calling set_mem_attributes_minus_bitpos when debug expression
expansion ended up expanding a SSA definition RHS which should
already have taken care of setting the appropriate MEM_ATTRS.

It also avoids calling set_mem_attributes on the
DECL_INITIAL of a CONST_DECL which seems pointless since there
cannot be a sensible MEM_EXPR derived from that.  We're overwriting
both other possibly useful info, alias-set and alignment immediately
so the following patch simply removes the call instead of making
the function deal with even more (unexpected) trees that are not
memory accesses.

2020-06-23  Richard Biener  <rguenther@suse.de>

	PR middle-end/95493
	PR middle-end/95690
	* cfgexpand.c (expand_debug_expr): Avoid calling
	set_mem_attributes_minus_bitpos when we were expanding
	an SSA name.
	* emit-rtl.c (set_mem_attributes_minus_bitpos): Remove
	ARRAY_REF special-casing, add CONSTRUCTOR to the set of
	special-cases we do not want MEM_EXPRs for.  Assert
	we end up with reasonable MEM_EXPRs.
	* varasm.c (build_constant_desc): Remove set_mem_attributes call.

	* g++.dg/torture/pr95493.C: New testcase.
	* g++.dg/torture/pr95493-1.C: Likewise.
	* gfortran.dg/pr95690.f90: Likewise.