[gcc r12-8976] tree-optimization/107686 - fix bitfield ref through vec_unpack optimization

[Richard Biener <rguenth@gcc.gnu.org>]

https://gcc.gnu.org/g:232305bd0c8dd37b17b815864e730aa630714606

commit r12-8976-g232305bd0c8dd37b17b815864e730aa630714606
Author: Richard Biener <rguenther@suse.de>
Date:   Wed Nov 16 15:27:13 2022 +0100

    tree-optimization/107686 - fix bitfield ref through vec_unpack optimization
    
    The following propely restricts the bitfield access to integral types
    when we look through VEC_UNPACK with the intent to emit a widening
    conversion.
    
            PR tree-optimization/107686
            * tree-ssa-forwprop.cc (optimize_vector_load): Restrict
            VEC_UNPACK support to integral typed bitfield refs.
    
            * gcc.dg/pr107686.c: New testcase.
    
    (cherry picked from commit 246bbdaa5f536b7a199dda9860c473137f40d622)

Diff:
---
 gcc/testsuite/gcc.dg/pr107686.c | 16 ++++++++++++++++
 gcc/tree-ssa-forwprop.cc        |  6 +++++-
 2 files changed, 21 insertions(+), 1 deletion(-)

diff --git a/gcc/testsuite/gcc.dg/pr107686.c b/gcc/testsuite/gcc.dg/pr107686.c
new file mode 100644
index 00000000000..2378103c555
--- /dev/null
+++ b/gcc/testsuite/gcc.dg/pr107686.c
@@ -0,0 +1,16 @@
+/* { dg-do compile { target { dfp && longlong64 } } } */
+/* { dg-options "-O" } */
+/* { dg-additional-options "-mavx2" { target x86_64-*-* i?86-*-* } } */
+
+typedef _Decimal64 __attribute__((__vector_size__ (64))) D;
+typedef __INT32_TYPE__ __attribute__((__vector_size__ (32))) U;
+typedef __INT64_TYPE__ __attribute__((__vector_size__ (64))) V;
+
+U u;
+D d;
+
+void
+foo (void)
+{
+  d = d < (D) __builtin_convertvector (u, V);
+}
diff --git a/gcc/tree-ssa-forwprop.cc b/gcc/tree-ssa-forwprop.cc
index 484491fa1c5..260d5dac83b 100644
--- a/gcc/tree-ssa-forwprop.cc
+++ b/gcc/tree-ssa-forwprop.cc
@@ -3118,7 +3118,11 @@ optimize_vector_load (gimple_stmt_iterator *gsi)
 	      && (def == lhs
 		  || (known_eq (bit_field_size (use_rhs), def_eltsize)
 		      && constant_multiple_p (bit_field_offset (use_rhs),
-					      def_eltsize))))
+					      def_eltsize)
+		      /* We can simulate the VEC_UNPACK_{HI,LO}_EXPR
+			 via a NOP_EXPR only for integral types.
+			 ???  Support VEC_UNPACK_FLOAT_{HI,LO}_EXPR.  */
+		      && INTEGRAL_TYPE_P (TREE_TYPE (use_rhs)))))
 	    {
 	      bf_stmts.safe_push (use_stmt);
 	      continue;