[AArch64] Generate permute patterns using rtx builders

This patch replaces switch statements that call specific generator
functions with code that constructs the rtl pattern directly.
This seemed to scale better to SVE and also seems less error-prone.

As a side-effect, the patch fixes the REV handling for diff==1,
vmode==E_V4HFmode and adds missing support for diff==3,
vmode==E_V4HFmode.

To compensate for the lack of switches that check for specific modes,
the patch makes aarch64_expand_vec_perm_const_1 reject permutes on
single-element vectors (specifically V1DImode).

2017-11-01  Richard Sandiford  <richard.sandiford@linaro.org>
	    Alan Hayward  <alan.hayward@arm.com>
	    David Sherwood  <david.sherwood@arm.com>

gcc/
	* config/aarch64/aarch64.c (aarch64_evpc_trn, aarch64_evpc_uzp)
	(aarch64_evpc_zip, aarch64_evpc_ext, aarch64_evpc_rev)
	(aarch64_evpc_dup): Generate rtl direcly, rather than using
	named expanders.
	(aarch64_expand_vec_perm_const_1): Explicitly check for permutes
	of a single element.
	* config/aarch64/iterators.md: Add a comment above the permute
	unspecs to say that they are generated directly by
	aarch64_expand_vec_perm_const.
	* config/aarch64/aarch64-simd.md: Likewise the permute instructions.

Reviewed-by: James Greenhalgh <james.greenhalgh@arm.com>

Co-Authored-By: Alan Hayward <alan.hayward@arm.com>
Co-Authored-By: David Sherwood <david.sherwood@arm.com>

From-SVN: r254324

4 files changed, 57 insertions, 217 deletions

diff --git a/gcc/ChangeLog b/gcc/ChangeLog
index 5e808e97..dcf1860 100644
--- a/gcc/ChangeLog
+++ b/gcc/ChangeLog
@@ -1,3 +1,18 @@
+2017-11-01  Richard Sandiford  <richard.sandiford@linaro.org>
+	    Alan Hayward  <alan.hayward@arm.com>
+	    David Sherwood  <david.sherwood@arm.com>
+
+	* config/aarch64/aarch64.c (aarch64_evpc_trn, aarch64_evpc_uzp)
+	(aarch64_evpc_zip, aarch64_evpc_ext, aarch64_evpc_rev)
+	(aarch64_evpc_dup): Generate rtl direcly, rather than using
+	named expanders.
+	(aarch64_expand_vec_perm_const_1): Explicitly check for permutes
+	of a single element.
+	* config/aarch64/iterators.md: Add a comment above the permute
+	unspecs to say that they are generated directly by
+	aarch64_expand_vec_perm_const.
+	* config/aarch64/aarch64-simd.md: Likewise the permute instructions.
+
 2017-11-01  Nathan Sidwell  <nathan@acm.org>
 
 	* tree-dump.c (dequeue_and_dump): Use HAS_DECL_ASSEMBLER_NAME_P.
diff --git a/gcc/config/aarch64/aarch64-simd.md b/gcc/config/aarch64/aarch64-simd.md
index 447ee3a..a3600b3 100644
--- a/gcc/config/aarch64/aarch64-simd.md
+++ b/gcc/config/aarch64/aarch64-simd.md
@@ -5369,6 +5369,9 @@
 [(set_attr "type" "multiple")]
 )
 
+;; This instruction's pattern is generated directly by
+;; aarch64_expand_vec_perm_const, so any changes to the pattern would
+;; need corresponding changes there.
 (define_insn "aarch64_<PERMUTE:perm_insn><PERMUTE:perm_hilo><mode>"
   [(set (match_operand:VALL_F16 0 "register_operand" "=w")
 	(unspec:VALL_F16 [(match_operand:VALL_F16 1 "register_operand" "w")
@@ -5379,7 +5382,10 @@
   [(set_attr "type" "neon_permute<q>")]
 )
 
-;; Note immediate (third) operand is lane index not byte index.
+;; This instruction's pattern is generated directly by
+;; aarch64_expand_vec_perm_const, so any changes to the pattern would
+;; need corresponding changes there.  Note that the immediate (third)
+;; operand is a lane index not a byte index.
 (define_insn "aarch64_ext<mode>"
   [(set (match_operand:VALL_F16 0 "register_operand" "=w")
         (unspec:VALL_F16 [(match_operand:VALL_F16 1 "register_operand" "w")
@@ -5395,6 +5401,9 @@
   [(set_attr "type" "neon_ext<q>")]
 )
 
+;; This instruction's pattern is generated directly by
+;; aarch64_expand_vec_perm_const, so any changes to the pattern would
+;; need corresponding changes there.
 (define_insn "aarch64_rev<REVERSE:rev_op><mode>"
   [(set (match_operand:VALL_F16 0 "register_operand" "=w")
 	(unspec:VALL_F16 [(match_operand:VALL_F16 1 "register_operand" "w")]
diff --git a/gcc/config/aarch64/aarch64.c b/gcc/config/aarch64/aarch64.c
index 9118689..69425b7 100644
--- a/gcc/config/aarch64/aarch64.c
+++ b/gcc/config/aarch64/aarch64.c
@@ -13263,7 +13263,6 @@ aarch64_evpc_trn (struct expand_vec_perm_d *d)
 {
   unsigned int i, odd, mask, nelt = d->perm.length ();
   rtx out, in0, in1, x;
-  rtx (*gen) (rtx, rtx, rtx);
   machine_mode vmode = d->vmode;
 
   if (GET_MODE_UNIT_SIZE (vmode) > 8)
@@ -13300,48 +13299,8 @@ aarch64_evpc_trn (struct expand_vec_perm_d *d)
     }
   out = d->target;
 
-  if (odd)
-    {
-      switch (vmode)
-	{
-	case E_V16QImode: gen = gen_aarch64_trn2v16qi; break;
-	case E_V8QImode: gen = gen_aarch64_trn2v8qi; break;
-	case E_V8HImode: gen = gen_aarch64_trn2v8hi; break;
-	case E_V4HImode: gen = gen_aarch64_trn2v4hi; break;
-	case E_V4SImode: gen = gen_aarch64_trn2v4si; break;
-	case E_V2SImode: gen = gen_aarch64_trn2v2si; break;
-	case E_V2DImode: gen = gen_aarch64_trn2v2di; break;
-	case E_V4HFmode: gen = gen_aarch64_trn2v4hf; break;
-	case E_V8HFmode: gen = gen_aarch64_trn2v8hf; break;
-	case E_V4SFmode: gen = gen_aarch64_trn2v4sf; break;
-	case E_V2SFmode: gen = gen_aarch64_trn2v2sf; break;
-	case E_V2DFmode: gen = gen_aarch64_trn2v2df; break;
-	default:
-	  return false;
-	}
-    }
-  else
-    {
-      switch (vmode)
-	{
-	case E_V16QImode: gen = gen_aarch64_trn1v16qi; break;
-	case E_V8QImode: gen = gen_aarch64_trn1v8qi; break;
-	case E_V8HImode: gen = gen_aarch64_trn1v8hi; break;
-	case E_V4HImode: gen = gen_aarch64_trn1v4hi; break;
-	case E_V4SImode: gen = gen_aarch64_trn1v4si; break;
-	case E_V2SImode: gen = gen_aarch64_trn1v2si; break;
-	case E_V2DImode: gen = gen_aarch64_trn1v2di; break;
-	case E_V4HFmode: gen = gen_aarch64_trn1v4hf; break;
-	case E_V8HFmode: gen = gen_aarch64_trn1v8hf; break;
-	case E_V4SFmode: gen = gen_aarch64_trn1v4sf; break;
-	case E_V2SFmode: gen = gen_aarch64_trn1v2sf; break;
-	case E_V2DFmode: gen = gen_aarch64_trn1v2df; break;
-	default:
-	  return false;
-	}
-    }
-
-  emit_insn (gen (out, in0, in1));
+  emit_set_insn (out, gen_rtx_UNSPEC (vmode, gen_rtvec (2, in0, in1),
+				      odd ? UNSPEC_TRN2 : UNSPEC_TRN1));
   return true;
 }
 
@@ -13351,7 +13310,6 @@ aarch64_evpc_uzp (struct expand_vec_perm_d *d)
 {
   unsigned int i, odd, mask, nelt = d->perm.length ();
   rtx out, in0, in1, x;
-  rtx (*gen) (rtx, rtx, rtx);
   machine_mode vmode = d->vmode;
 
   if (GET_MODE_UNIT_SIZE (vmode) > 8)
@@ -13387,48 +13345,8 @@ aarch64_evpc_uzp (struct expand_vec_perm_d *d)
     }
   out = d->target;
 
-  if (odd)
-    {
-      switch (vmode)
-	{
-	case E_V16QImode: gen = gen_aarch64_uzp2v16qi; break;
-	case E_V8QImode: gen = gen_aarch64_uzp2v8qi; break;
-	case E_V8HImode: gen = gen_aarch64_uzp2v8hi; break;
-	case E_V4HImode: gen = gen_aarch64_uzp2v4hi; break;
-	case E_V4SImode: gen = gen_aarch64_uzp2v4si; break;
-	case E_V2SImode: gen = gen_aarch64_uzp2v2si; break;
-	case E_V2DImode: gen = gen_aarch64_uzp2v2di; break;
-	case E_V4HFmode: gen = gen_aarch64_uzp2v4hf; break;
-	case E_V8HFmode: gen = gen_aarch64_uzp2v8hf; break;
-	case E_V4SFmode: gen = gen_aarch64_uzp2v4sf; break;
-	case E_V2SFmode: gen = gen_aarch64_uzp2v2sf; break;
-	case E_V2DFmode: gen = gen_aarch64_uzp2v2df; break;
-	default:
-	  return false;
-	}
-    }
-  else
-    {
-      switch (vmode)
-	{
-	case E_V16QImode: gen = gen_aarch64_uzp1v16qi; break;
-	case E_V8QImode: gen = gen_aarch64_uzp1v8qi; break;
-	case E_V8HImode: gen = gen_aarch64_uzp1v8hi; break;
-	case E_V4HImode: gen = gen_aarch64_uzp1v4hi; break;
-	case E_V4SImode: gen = gen_aarch64_uzp1v4si; break;
-	case E_V2SImode: gen = gen_aarch64_uzp1v2si; break;
-	case E_V2DImode: gen = gen_aarch64_uzp1v2di; break;
-	case E_V4HFmode: gen = gen_aarch64_uzp1v4hf; break;
-	case E_V8HFmode: gen = gen_aarch64_uzp1v8hf; break;
-	case E_V4SFmode: gen = gen_aarch64_uzp1v4sf; break;
-	case E_V2SFmode: gen = gen_aarch64_uzp1v2sf; break;
-	case E_V2DFmode: gen = gen_aarch64_uzp1v2df; break;
-	default:
-	  return false;
-	}
-    }
-
-  emit_insn (gen (out, in0, in1));
+  emit_set_insn (out, gen_rtx_UNSPEC (vmode, gen_rtvec (2, in0, in1),
+				      odd ? UNSPEC_UZP2 : UNSPEC_UZP1));
   return true;
 }
 
@@ -13438,7 +13356,6 @@ aarch64_evpc_zip (struct expand_vec_perm_d *d)
 {
   unsigned int i, high, mask, nelt = d->perm.length ();
   rtx out, in0, in1, x;
-  rtx (*gen) (rtx, rtx, rtx);
   machine_mode vmode = d->vmode;
 
   if (GET_MODE_UNIT_SIZE (vmode) > 8)
@@ -13479,48 +13396,8 @@ aarch64_evpc_zip (struct expand_vec_perm_d *d)
     }
   out = d->target;
 
-  if (high)
-    {
-      switch (vmode)
-	{
-	case E_V16QImode: gen = gen_aarch64_zip2v16qi; break;
-	case E_V8QImode: gen = gen_aarch64_zip2v8qi; break;
-	case E_V8HImode: gen = gen_aarch64_zip2v8hi; break;
-	case E_V4HImode: gen = gen_aarch64_zip2v4hi; break;
-	case E_V4SImode: gen = gen_aarch64_zip2v4si; break;
-	case E_V2SImode: gen = gen_aarch64_zip2v2si; break;
-	case E_V2DImode: gen = gen_aarch64_zip2v2di; break;
-	case E_V4HFmode: gen = gen_aarch64_zip2v4hf; break;
-	case E_V8HFmode: gen = gen_aarch64_zip2v8hf; break;
-	case E_V4SFmode: gen = gen_aarch64_zip2v4sf; break;
-	case E_V2SFmode: gen = gen_aarch64_zip2v2sf; break;
-	case E_V2DFmode: gen = gen_aarch64_zip2v2df; break;
-	default:
-	  return false;
-	}
-    }
-  else
-    {
-      switch (vmode)
-	{
-	case E_V16QImode: gen = gen_aarch64_zip1v16qi; break;
-	case E_V8QImode: gen = gen_aarch64_zip1v8qi; break;
-	case E_V8HImode: gen = gen_aarch64_zip1v8hi; break;
-	case E_V4HImode: gen = gen_aarch64_zip1v4hi; break;
-	case E_V4SImode: gen = gen_aarch64_zip1v4si; break;
-	case E_V2SImode: gen = gen_aarch64_zip1v2si; break;
-	case E_V2DImode: gen = gen_aarch64_zip1v2di; break;
-	case E_V4HFmode: gen = gen_aarch64_zip1v4hf; break;
-	case E_V8HFmode: gen = gen_aarch64_zip1v8hf; break;
-	case E_V4SFmode: gen = gen_aarch64_zip1v4sf; break;
-	case E_V2SFmode: gen = gen_aarch64_zip1v2sf; break;
-	case E_V2DFmode: gen = gen_aarch64_zip1v2df; break;
-	default:
-	  return false;
-	}
-    }
-
-  emit_insn (gen (out, in0, in1));
+  emit_set_insn (out, gen_rtx_UNSPEC (vmode, gen_rtvec (2, in0, in1),
+				      high ? UNSPEC_ZIP2 : UNSPEC_ZIP1));
   return true;
 }
 
@@ -13530,7 +13407,6 @@ static bool
 aarch64_evpc_ext (struct expand_vec_perm_d *d)
 {
   unsigned int i, nelt = d->perm.length ();
-  rtx (*gen) (rtx, rtx, rtx, rtx);
   rtx offset;
 
   unsigned int location = d->perm[0]; /* Always < nelt.  */
@@ -13548,24 +13424,6 @@ aarch64_evpc_ext (struct expand_vec_perm_d *d)
         return false;
     }
 
-  switch (d->vmode)
-    {
-    case E_V16QImode: gen = gen_aarch64_extv16qi; break;
-    case E_V8QImode: gen = gen_aarch64_extv8qi; break;
-    case E_V4HImode: gen = gen_aarch64_extv4hi; break;
-    case E_V8HImode: gen = gen_aarch64_extv8hi; break;
-    case E_V2SImode: gen = gen_aarch64_extv2si; break;
-    case E_V4SImode: gen = gen_aarch64_extv4si; break;
-    case E_V4HFmode: gen = gen_aarch64_extv4hf; break;
-    case E_V8HFmode: gen = gen_aarch64_extv8hf; break;
-    case E_V2SFmode: gen = gen_aarch64_extv2sf; break;
-    case E_V4SFmode: gen = gen_aarch64_extv4sf; break;
-    case E_V2DImode: gen = gen_aarch64_extv2di; break;
-    case E_V2DFmode: gen = gen_aarch64_extv2df; break;
-    default:
-      return false;
-    }
-
   /* Success! */
   if (d->testing_p)
     return true;
@@ -13584,7 +13442,10 @@ aarch64_evpc_ext (struct expand_vec_perm_d *d)
     }
 
   offset = GEN_INT (location);
-  emit_insn (gen (d->target, d->op0, d->op1, offset));
+  emit_set_insn (d->target,
+		 gen_rtx_UNSPEC (d->vmode,
+				 gen_rtvec (3, d->op0, d->op1, offset),
+				 UNSPEC_EXT));
   return true;
 }
 
@@ -13593,55 +13454,21 @@ aarch64_evpc_ext (struct expand_vec_perm_d *d)
 static bool
 aarch64_evpc_rev (struct expand_vec_perm_d *d)
 {
-  unsigned int i, j, diff, nelt = d->perm.length ();
-  rtx (*gen) (rtx, rtx);
+  unsigned int i, j, diff, size, unspec, nelt = d->perm.length ();
 
   if (!d->one_vector_p)
     return false;
 
   diff = d->perm[0];
-  switch (diff)
-    {
-    case 7:
-      switch (d->vmode)
-	{
-	case E_V16QImode: gen = gen_aarch64_rev64v16qi; break;
-	case E_V8QImode: gen = gen_aarch64_rev64v8qi;  break;
-	default:
-	  return false;
-	}
-      break;
-    case 3:
-      switch (d->vmode)
-	{
-	case E_V16QImode: gen = gen_aarch64_rev32v16qi; break;
-	case E_V8QImode: gen = gen_aarch64_rev32v8qi;  break;
-	case E_V8HImode: gen = gen_aarch64_rev64v8hi;  break;
-	case E_V4HImode: gen = gen_aarch64_rev64v4hi;  break;
-	default:
-	  return false;
-	}
-      break;
-    case 1:
-      switch (d->vmode)
-	{
-	case E_V16QImode: gen = gen_aarch64_rev16v16qi; break;
-	case E_V8QImode: gen = gen_aarch64_rev16v8qi;  break;
-	case E_V8HImode: gen = gen_aarch64_rev32v8hi;  break;
-	case E_V4HImode: gen = gen_aarch64_rev32v4hi;  break;
-	case E_V4SImode: gen = gen_aarch64_rev64v4si;  break;
-	case E_V2SImode: gen = gen_aarch64_rev64v2si;  break;
-	case E_V4SFmode: gen = gen_aarch64_rev64v4sf;  break;
-	case E_V2SFmode: gen = gen_aarch64_rev64v2sf;  break;
-	case E_V8HFmode: gen = gen_aarch64_rev64v8hf;  break;
-	case E_V4HFmode: gen = gen_aarch64_rev64v4hf;  break;
-	default:
-	  return false;
-	}
-      break;
-    default:
-      return false;
-    }
+  size = (diff + 1) * GET_MODE_UNIT_SIZE (d->vmode);
+  if (size == 8)
+    unspec = UNSPEC_REV64;
+  else if (size == 4)
+    unspec = UNSPEC_REV32;
+  else if (size == 2)
+    unspec = UNSPEC_REV16;
+  else
+    return false;
 
   for (i = 0; i < nelt ; i += diff + 1)
     for (j = 0; j <= diff; j += 1)
@@ -13660,14 +13487,14 @@ aarch64_evpc_rev (struct expand_vec_perm_d *d)
   if (d->testing_p)
     return true;
 
-  emit_insn (gen (d->target, d->op0));
+  emit_set_insn (d->target, gen_rtx_UNSPEC (d->vmode, gen_rtvec (1, d->op0),
+					    unspec));
   return true;
 }
 
 static bool
 aarch64_evpc_dup (struct expand_vec_perm_d *d)
 {
-  rtx (*gen) (rtx, rtx, rtx);
   rtx out = d->target;
   rtx in0;
   machine_mode vmode = d->vmode;
@@ -13689,25 +13516,9 @@ aarch64_evpc_dup (struct expand_vec_perm_d *d)
   in0 = d->op0;
   lane = GEN_INT (elt); /* The pattern corrects for big-endian.  */
 
-  switch (vmode)
-    {
-    case E_V16QImode: gen = gen_aarch64_dup_lanev16qi; break;
-    case E_V8QImode: gen = gen_aarch64_dup_lanev8qi; break;
-    case E_V8HImode: gen = gen_aarch64_dup_lanev8hi; break;
-    case E_V4HImode: gen = gen_aarch64_dup_lanev4hi; break;
-    case E_V4SImode: gen = gen_aarch64_dup_lanev4si; break;
-    case E_V2SImode: gen = gen_aarch64_dup_lanev2si; break;
-    case E_V2DImode: gen = gen_aarch64_dup_lanev2di; break;
-    case E_V8HFmode: gen = gen_aarch64_dup_lanev8hf; break;
-    case E_V4HFmode: gen = gen_aarch64_dup_lanev4hf; break;
-    case E_V4SFmode: gen = gen_aarch64_dup_lanev4sf; break;
-    case E_V2SFmode: gen = gen_aarch64_dup_lanev2sf; break;
-    case E_V2DFmode: gen = gen_aarch64_dup_lanev2df; break;
-    default:
-      return false;
-    }
-
-  emit_insn (gen (out, in0, lane));
+  rtx parallel = gen_rtx_PARALLEL (vmode, gen_rtvec (1, lane));
+  rtx select = gen_rtx_VEC_SELECT (GET_MODE_INNER (vmode), in0, parallel);
+  emit_set_insn (out, gen_rtx_VEC_DUPLICATE (vmode, select));
   return true;
 }
 
@@ -13760,7 +13571,7 @@ aarch64_expand_vec_perm_const_1 (struct expand_vec_perm_d *d)
       std::swap (d->op0, d->op1);
     }
 
-  if (TARGET_SIMD)
+  if (TARGET_SIMD && nelt > 1)
     {
       if (aarch64_evpc_rev (d))
 	return true;
diff --git a/gcc/config/aarch64/iterators.md b/gcc/config/aarch64/iterators.md
index 68da316..46aa6fd 100644
--- a/gcc/config/aarch64/iterators.md
+++ b/gcc/config/aarch64/iterators.md
@@ -322,16 +322,21 @@
     UNSPEC_TBL		; Used in vector permute patterns.
     UNSPEC_TBX		; Used in vector permute patterns.
     UNSPEC_CONCAT	; Used in vector permute patterns.
+
+    ;; The following permute unspecs are generated directly by
+    ;; aarch64_expand_vec_perm_const, so any changes to the underlying
+    ;; instructions would need a corresponding change there.
     UNSPEC_ZIP1		; Used in vector permute patterns.
     UNSPEC_ZIP2		; Used in vector permute patterns.
     UNSPEC_UZP1		; Used in vector permute patterns.
     UNSPEC_UZP2		; Used in vector permute patterns.
     UNSPEC_TRN1		; Used in vector permute patterns.
     UNSPEC_TRN2		; Used in vector permute patterns.
-    UNSPEC_EXT		; Used in aarch64-simd.md.
+    UNSPEC_EXT		; Used in vector permute patterns.
     UNSPEC_REV64	; Used in vector reverse patterns (permute).
     UNSPEC_REV32	; Used in vector reverse patterns (permute).
     UNSPEC_REV16	; Used in vector reverse patterns (permute).
+
     UNSPEC_AESE		; Used in aarch64-simd.md.
     UNSPEC_AESD         ; Used in aarch64-simd.md.
     UNSPEC_AESMC        ; Used in aarch64-simd.md.