[gcc(refs/users/clyon/heads/mve-autovec)] WIP: add new constraint to thumb2_movhi* patterns, and assorted fixes

[Christophe Lyon <clyon@gcc.gnu.org>]

https://gcc.gnu.org/g:99515083dc34ab4473c8801ab0390dbd90a2db6b

commit 99515083dc34ab4473c8801ab0390dbd90a2db6b
Author: Christophe Lyon <christophe.lyon@foss.st.com>
Date:   Tue Nov 16 13:58:33 2021 +0000

    WIP: add new constraint to thumb2_movhi* patterns, and assorted fixes
    
    hack simd_valid_immediate
    
    hack simd_valid_immediate
    
    fix vector-compare-2

Diff:
---
 gcc/config/arm/arm-protos.h   |  1 +
 gcc/config/arm/arm.c          | 29 ++++++++++++++++++++++++++++-
 gcc/config/arm/constraints.md |  6 ++++++
 gcc/config/arm/vfp.md         | 20 ++++++++++++++++----
 4 files changed, 51 insertions(+), 5 deletions(-)

diff --git a/gcc/config/arm/arm-protos.h b/gcc/config/arm/arm-protos.h
index 9e3d71e0c29..66bca03803a 100644
--- a/gcc/config/arm/arm-protos.h
+++ b/gcc/config/arm/arm-protos.h
@@ -101,6 +101,7 @@ extern char *neon_output_shift_immediate (const char *, char, rtx *,
 					  machine_mode, int, bool);
 extern void neon_pairwise_reduce (rtx, rtx, machine_mode,
 				  rtx (*) (rtx, rtx, rtx));
+extern rtx mve_const_bool_vec_to_hi (rtx const_vec);
 extern rtx neon_make_constant (rtx, bool generate = true);
 extern tree arm_builtin_vectorized_function (unsigned int, tree, tree);
 extern void neon_expand_vector_init (rtx, rtx);
diff --git a/gcc/config/arm/arm.c b/gcc/config/arm/arm.c
index bc18d3c888d..3b6a52ce40c 100644
--- a/gcc/config/arm/arm.c
+++ b/gcc/config/arm/arm.c
@@ -12754,7 +12754,8 @@ simd_valid_immediate (rtx op, machine_mode mode, int inverse,
   innersize = GET_MODE_UNIT_SIZE (mode);
 
   /* Only support 128-bit vectors for MVE.  */
-  if (TARGET_HAVE_MVE && (!vector || n_elts * innersize != 16))
+  if (TARGET_HAVE_MVE
+      && (!vector || (GET_MODE_CLASS (mode) == MODE_VECTOR_BOOL) || n_elts * innersize != 16))
     return -1;
 
   /* Vectors of float constants.  */
@@ -13119,6 +13120,30 @@ neon_vdup_constant (rtx vals, bool generate)
   return gen_vec_duplicate (mode, x);
 }
 
+/* FIXME: not sure this is really needed, since it doesn't fix
+   vector-compare-2.c at -O0. */
+rtx
+mve_const_bool_vec_to_hi (rtx const_vec)
+{
+  int n_elts = GET_MODE_NUNITS ( GET_MODE (const_vec));
+  int repeat = 16 / n_elts;
+  int i;
+  int hi_val = 0;
+
+  for (i = 0; i < n_elts; i++)
+    {
+      rtx el = CONST_VECTOR_ELT (const_vec, i);
+      unsigned HOST_WIDE_INT elpart;
+
+      gcc_assert (CONST_INT_P (el));
+      elpart = INTVAL (el);
+
+      for (int j = 0; j < repeat; j++)
+	hi_val |= elpart << (i * repeat + j);
+    }
+  return GEN_INT (hi_val);
+}
+
 /* Return a non-NULL RTX iff VALS, which is a PARALLEL containing only
    constants (for vec_init) or CONST_VECTOR, can be effeciently loaded
    into a register.
@@ -13159,6 +13184,8 @@ neon_make_constant (rtx vals, bool generate)
       && simd_immediate_valid_for_move (const_vec, mode, NULL, NULL))
     /* Load using VMOV.  On Cortex-A8 this takes one cycle.  */
     return const_vec;
+  else if (TARGET_HAVE_MVE && (GET_MODE_CLASS (mode) == MODE_VECTOR_BOOL))
+    return mve_const_bool_vec_to_hi (const_vec);
   else if ((target = neon_vdup_constant (vals, generate)) != NULL_RTX)
     /* Loaded using VDUP.  On Cortex-A8 the VDUP takes one NEON
        pipeline cycle; creating the constant takes one or two ARM
diff --git a/gcc/config/arm/constraints.md b/gcc/config/arm/constraints.md
index a5a19a7ed5b..acc447ba0f0 100644
--- a/gcc/config/arm/constraints.md
+++ b/gcc/config/arm/constraints.md
@@ -312,6 +312,12 @@
  (and (match_code "const_vector")
       (match_test "(TARGET_NEON || TARGET_HAVE_MVE) && op == CONST0_RTX (mode)")))
 
+(define_constraint "DB"
+ "@internal
+  In ARM/Thumb-2 state with MVE a constant vector of booleans."
+ (and (match_code "const_vector")
+      (match_test "TARGET_HAVE_MVE && GET_MODE_CLASS (mode) == MODE_VECTOR_BOOL")))
+
 (define_constraint "Da"
  "@internal
   In ARM/Thumb-2 state a const_int, const_double or const_vector that can
diff --git a/gcc/config/arm/vfp.md b/gcc/config/arm/vfp.md
index b94b54e8a79..de17f5895dd 100644
--- a/gcc/config/arm/vfp.md
+++ b/gcc/config/arm/vfp.md
@@ -76,7 +76,7 @@
    (match_operand:MVE_7_HI 0 "nonimmediate_operand"
     "=rk, r, l, r, m, r, *t, r, *t, Up, r")
    (match_operand:MVE_7_HI 1 "general_operand"
-    "rk, I, Py, n, r, m, r, *t, *t, r, Up"))]
+    "rk, IDB, Py, n, r, m, r, *t, *t, r, Up"))]
  "TARGET_THUMB2 && TARGET_VFP_BASE
   && !TARGET_VFP_FP16INST
   && (register_operand (operands[0], <MODE>mode)
@@ -85,9 +85,14 @@
   switch (which_alternative)
     {
     case 0:
-    case 1:
     case 2:
       return "mov%?\t%0, %1\t%@ movhi";
+    case 1:
+      if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_VECTOR_BOOL)
+        operands[1] = mve_const_bool_vec_to_hi (operands[1]);
+      else
+        operands[1] = gen_lowpart (HImode, operands[1]);
+      return "mov%?\t%0, %1\t%@ movhi";
     case 3:
       return "movw%?\t%0, %L1\t%@ movhi";
     case 4:
@@ -176,17 +181,24 @@
    (match_operand:MVE_7_HI 0 "nonimmediate_operand"
     "=rk, r, l, r, m, r, *t, r, *t, Up, r")
    (match_operand:MVE_7_HI 1 "general_operand"
-    "rk, I, Py, n, r, m, r, *t, *t, r, Up"))]
+    "rk, IDB, Py, n, r, m, r, *t, *t, r, Up"))]
  "TARGET_THUMB2 && (TARGET_VFP_FP16INST || TARGET_HAVE_MVE)
   && (register_operand (operands[0], <MODE>mode)
        || register_operand (operands[1], <MODE>mode))"
 {
+extern rtx mve_const_bool_vec_to_hi (rtx const_vec);
+
   switch (which_alternative)
     {
     case 0:
-    case 1:
     case 2:
       return "mov%?\t%0, %1\t%@ movhi";
+    case 1:
+      if (GET_MODE_CLASS (GET_MODE (operands[1])) == MODE_VECTOR_BOOL)
+        operands[1] = mve_const_bool_vec_to_hi (operands[1]);
+      else
+        operands[1] = gen_lowpart (HImode, operands[1]);
+      return "mov%?\t%0, %1\t%@ movhi";
     case 3:
       return "movw%?\t%0, %L1\t%@ movhi";
     case 4: