RISC-V: Early expand DImode vec_duplicate in RV32 system

An ICE was discovered in recent rounding autovec support:

config/riscv/riscv-v.cc:4314
   65 | }
      | ^
0x1fa5223 riscv_vector::validate_change_or_fail(rtx_def*, rtx_def**,
rtx_def*, bool)
        /home/pli/repos/gcc/222/riscv-gnu-toolchain/gcc/__RISC-V_BUILD/../gcc/config/riscv/riscv-v.cc:4314
0x1fb1aa2 pre_vsetvl::remove_avl_operand()
        /home/pli/repos/gcc/222/riscv-gnu-toolchain/gcc/__RISC-V_BUILD/../gcc/config/riscv/riscv-vsetvl.cc:3342
0x1fb18c1 pre_vsetvl::cleaup()
        /home/pli/repos/gcc/222/riscv-gnu-toolchain/gcc/__RISC-V_BUILD/../gcc/config/riscv/riscv-vsetvl.cc:3308
0x1fb216d pass_vsetvl::lazy_vsetvl()
        /home/pli/repos/gcc/222/riscv-gnu-toolchain/gcc/__RISC-V_BUILD/../gcc/config/riscv/riscv-vsetvl.cc:3480
0x1fb2214 pass_vsetvl::execute(function*)
        /home/pli/repos/gcc/222/riscv-gnu-toolchain/gcc/__RISC-V_BUILD/../gcc/config/riscv/riscv-vsetvl.cc:3504

The root cause is that the RA reload into (set (reg) vec_duplicate:DI). However, it is not valid in RV32 system
since we don't have a single broadcast instruction DI scalar in RV32 system.
We should expand it early for RV32 system.

gcc/ChangeLog:

	* config/riscv/predicates.md: Adapt predicate.
	* config/riscv/riscv-protos.h (can_be_broadcasted_p): New function.
	* config/riscv/riscv-v.cc (can_be_broadcasted_p): Ditto.
	* config/riscv/vector.md (vec_duplicate<mode>): New pattern.
	(*vec_duplicate<mode>): Adapt vec_duplicate insn pattern.

gcc/testsuite/ChangeLog:

	* gcc.target/riscv/rvv/autovec/unop/sew64-rv32.c: New test.

5 files changed, 70 insertions, 9 deletions

diff --git a/gcc/config/riscv/predicates.md b/gcc/config/riscv/predicates.md
index a37d035..90567a8 100644
--- a/gcc/config/riscv/predicates.md
+++ b/gcc/config/riscv/predicates.md
@@ -547,14 +547,7 @@
 
 ;; The scalar operand can be directly broadcast by RVV instructions.
 (define_predicate "direct_broadcast_operand"
-  (and (match_test "!(reload_completed && !FLOAT_MODE_P (GET_MODE (op))
-		&& (register_operand (op, GET_MODE (op)) || CONST_INT_P (op)
-		|| rtx_equal_p (op, CONST0_RTX (GET_MODE (op))))
-		&& maybe_gt (GET_MODE_BITSIZE (GET_MODE (op)), GET_MODE_BITSIZE (Pmode)))")
-    (ior (match_test "rtx_equal_p (op, CONST0_RTX (GET_MODE (op)))")
-         (ior (match_code "const_int,const_poly_int")
-              (ior (match_operand 0 "register_operand")
-                   (match_test "satisfies_constraint_Wdm (op)"))))))
+  (match_test "riscv_vector::can_be_broadcasted_p (op)"))
 
 ;; A CONST_INT operand that has exactly two bits cleared.
 (define_predicate "const_nottwobits_operand"
diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h
index 52d2a2c..85d4f6e 100644
--- a/gcc/config/riscv/riscv-protos.h
+++ b/gcc/config/riscv/riscv-protos.h
@@ -597,6 +597,7 @@ uint8_t get_sew (rtx_insn *);
 enum vlmul_type get_vlmul (rtx_insn *);
 int count_regno_occurrences (rtx_insn *, unsigned int);
 bool imm_avl_p (machine_mode);
+bool can_be_broadcasted_p (rtx);
 }
 
 /* We classify builtin types into two classes:
diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
index eeefda64..0dc2ac4 100644
--- a/gcc/config/riscv/riscv-v.cc
+++ b/gcc/config/riscv/riscv-v.cc
@@ -4362,4 +4362,24 @@ count_regno_occurrences (rtx_insn *rinsn, unsigned int regno)
   return count;
 }
 
+/* Return true if the OP can be directly broadcasted.  */
+bool
+can_be_broadcasted_p (rtx op)
+{
+  machine_mode mode = GET_MODE (op);
+  /* We don't allow RA (register allocation) reload generate
+    (vec_duplicate:DI reg) in RV32 system wheras we allow
+    (vec_duplicate:DI mem) in RV32 system.  */
+  if (!can_create_pseudo_p () && !FLOAT_MODE_P (mode)
+      && maybe_gt (GET_MODE_SIZE (mode), GET_MODE_SIZE (Pmode))
+      && !satisfies_constraint_Wdm (op))
+    return false;
+
+  if (satisfies_constraint_K (op) || register_operand (op, mode)
+      || satisfies_constraint_Wdm (op) || rtx_equal_p (op, CONST0_RTX (mode)))
+    return true;
+
+  return can_create_pseudo_p () && nonmemory_operand (op, mode);
+}
+
 } // namespace riscv_vector
diff --git a/gcc/config/riscv/vector.md b/gcc/config/riscv/vector.md
index 8509c4f..e23f649 100644
--- a/gcc/config/riscv/vector.md
+++ b/gcc/config/riscv/vector.md
@@ -1370,11 +1370,29 @@
 ;; ---- Duplicate Operations
 ;; -----------------------------------------------------------------
 
+(define_expand "vec_duplicate<mode>"
+  [(set (match_operand:V_VLS 0 "register_operand")
+        (vec_duplicate:V_VLS
+          (match_operand:<VEL> 1 "direct_broadcast_operand")))]
+  "TARGET_VECTOR"
+  {
+    /* Early expand DImode broadcast in RV32 system to avoid RA reload
+       generate (set (reg) (vec_duplicate:DI)).  */
+    if (maybe_gt (GET_MODE_SIZE (<VEL>mode), GET_MODE_SIZE (Pmode)))
+      {
+        riscv_vector::emit_vlmax_insn (code_for_pred_broadcast (<MODE>mode),
+				       riscv_vector::UNARY_OP, operands);
+	DONE;
+      }
+    /* Otherwise, allow it fall into general vec_duplicate pattern
+       which allow us to have vv->vx combine optimization in later pass.  */
+  })
+
 ;; According to GCC internal:
 ;; This pattern only handles duplicates of non-constant inputs.
 ;; Constant vectors go through the movm pattern instead.
 ;; So "direct_broadcast_operand" can only be mem or reg, no CONSTANT.
-(define_insn_and_split "vec_duplicate<mode>"
+(define_insn_and_split "*vec_duplicate<mode>"
   [(set (match_operand:V_VLS 0 "register_operand")
         (vec_duplicate:V_VLS
           (match_operand:<VEL> 1 "direct_broadcast_operand")))]
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/sew64-rv32.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/sew64-rv32.c
new file mode 100644
index 0000000..8f813690
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/unop/sew64-rv32.c
@@ -0,0 +1,29 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv32gcv -mabi=ilp32d -O3 -ftree-vectorize -fno-vect-cost-model -ffast-math" } */
+
+#include <stdint-gcc.h>
+#include "test-math.h"
+
+#define ARRAY_SIZE 128
+
+float in[ARRAY_SIZE];
+int64_t out[ARRAY_SIZE];
+int64_t ref[ARRAY_SIZE];
+
+TEST_UNARY_CALL_CVT (float, int64_t, __builtin_llrintf)
+TEST_ASSERT (int64_t)
+
+
+TEST_INIT_CVT (float, __builtin_inf (), int64_t, __builtin_llrintf (__builtin_inff ()), 19)
+TEST_INIT_CVT (float, -__builtin_inf (), int64_t, __builtin_llrintf (-__builtin_inff ()), 20)
+TEST_INIT_CVT (float, __builtin_nanf (""), int64_t, 0x7fffffffffffffff, 21)
+
+int
+main ()
+{
+  RUN_TEST_CVT (float, int64_t, 19, __builtin_llrintf, in, out, ref, ARRAY_SIZE);
+  RUN_TEST_CVT (float, int64_t, 20, __builtin_llrintf, in, out, ref, ARRAY_SIZE);
+  RUN_TEST_CVT (float, int64_t, 21, __builtin_llrintf, in, out, ref, ARRAY_SIZE);
+
+  return 0;
+}