RISC-V: Fix vec_init for simple sequences [PR114028].

For a vec_init (_a, _a, _a, _a) with _a of mode DImode we try to
construct a "superword" of two "_a"s.  This only works for modes < Pmode
when we can "shift and or" both halves into one Pmode register.
This patch disallows the optimization for inner_mode == Pmode and emits
a simple broadcast in such a case.

gcc/ChangeLog:

	PR target/114028

	* config/riscv/riscv-v.cc (rvv_builder::can_duplicate_repeating_sequence_p):
	Return false if inner mode is already Pmode.
	(rvv_builder::is_all_same_sequence): New function.
	(expand_vec_init): Emit broadcast if sequence is all same.

gcc/testsuite/ChangeLog:

	* gcc.target/riscv/rvv/autovec/pr114028.c: New test.

diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
index 0cfbd21ce6f91cf97b278c9d504e20a7d802cd4b..29d58deb99534390f2205b017e09329b1f443441 100644 (file)
--- a/gcc/config/riscv/riscv-v.cc
+++ b/gcc/config/riscv/riscv-v.cc
@@ -443,6 +443,7 @@ public:
   }
 
   bool can_duplicate_repeating_sequence_p ();
+  bool is_repeating_sequence ();
   rtx get_merged_repeating_sequence ();
 
   bool repeating_sequence_use_merge_profitable_p ();
@@ -483,7 +484,8 @@ rvv_builder::can_duplicate_repeating_sequence_p ()
 {
   poly_uint64 new_size = exact_div (full_nelts (), npatterns ());
   unsigned int new_inner_size = m_inner_bits_size * npatterns ();
-  if (!int_mode_for_size (new_inner_size, 0).exists (&m_new_inner_mode)
+  if (m_inner_mode == Pmode
+      || !int_mode_for_size (new_inner_size, 0).exists (&m_new_inner_mode)
       || GET_MODE_SIZE (m_new_inner_mode) > UNITS_PER_WORD
       || !get_vector_mode (m_new_inner_mode, new_size).exists (&m_new_mode))
     return false;
@@ -492,6 +494,18 @@ rvv_builder::can_duplicate_repeating_sequence_p ()
   return nelts_per_pattern () == 1;
 }
 
+/* Return true if the vector is a simple sequence with one pattern and all
+   elements the same.  */
+bool
+rvv_builder::is_repeating_sequence ()
+{
+  if (npatterns () > 1)
+    return false;
+  if (full_nelts ().is_constant ())
+    return repeating_sequence_p (0, full_nelts ().to_constant (), 1);
+  return nelts_per_pattern () == 1;
+}
+
 /* Return true if it is a repeating sequence that using
    merge approach has better codegen than using default
    approach (slide1down).
@@ -2544,6 +2558,15 @@ expand_vec_init (rtx target, rtx vals)
     v.quick_push (XVECEXP (vals, 0, i));
   v.finalize ();
 
+  /* If the sequence is v = { a, a, a, a } just broadcast an element.  */
+  if (v.is_repeating_sequence ())
+    {
+      machine_mode mode = GET_MODE (target);
+      rtx dup = expand_vector_broadcast (mode, v.elt (0));
+      emit_move_insn (target, dup);
+      return;
+    }
+
   if (nelts > 3)
     {
       /* Case 1: Convert v = { a, b, a, b } into v = { ab, ab }.  */

diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr114028.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr114028.c
new file mode 100644 (file)
index 0000000..e41f272
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/pr114028.c
@@ -0,0 +1,25 @@
+/* { dg-do compile }  */
+/* { dg-options "-march=rv64gcv_zvl256b -mabi=lp64d -O3" } */
+
+int a, d = 55003;
+long c = 0, h;
+long e = 1;
+short i;
+
+int
+main ()
+{
+  for (int g = 0; g < 16; g++)
+    {
+      d |= c;
+      short l = d;
+      i = l < 0 || a >> 4 ? d : a;
+      h = i - 8L;
+      e &= h;
+    }
+
+  if (e != 1)
+    __builtin_abort ();
+}
+
+/* { dg-final { scan-assembler-not "vmv\.v\.i\tv\[0-9\],0" } } */