arm: elide some cases where the AES erratum workaround is not required.

Some common cases where the AES erratum workaround are not required
are when there are 64- or 128-bit loads from memory, moving a 128-bit
value from core registers, and where a 128-bit constant is being
loaded from a literal pool.  The loads may also be misaligned or
generated via a neon intrinsic function.

gcc/ChangeLog:

	* config/arm/crypto.md (aes_op_protect): Allow moves from core
	registers and from memory.
	(aes_op_protect_misalign_load): New pattern.
	(aes_op_protect_neon_vld1v16qi): New pattern.

1 files changed, 47 insertions, 8 deletions

diff --git a/gcc/config/arm/crypto.md b/gcc/config/arm/crypto.md
index df85735..4c78507 100644
--- a/gcc/config/arm/crypto.md
+++ b/gcc/config/arm/crypto.md
@@ -62,17 +62,56 @@
   [(set_attr "type" "<crypto_type>")]
 )
 
-; Mitigate against AES erratum on Cortex-A57 and Cortex-A72 by performing
-; a 128-bit operation on an operand producer.  This can be eliminated only
-; if we know that the operand was produced by a full-width operation.
-; V16QImode matches <crypto_mode> for the AES instructions.
+;; Mitigate against AES erratum on Cortex-A57 and Cortex-A72 by
+;; performing a 128-bit operation on an operand producer.  This can be
+;; eliminated only if we know that the operand was produced by a
+;; full-width operation.  V16QImode matches <crypto_mode> for the AES
+;; instructions.  Handle some very common cases where the source is
+;; known to be safe (transfers from core registers and memory).
 (define_insn "aes_op_protect"
-  [(set (match_operand:V16QI 0 "register_operand" "=w")
-	(unspec:V16QI [(match_operand:V16QI 1 "register_operand" "0")]
+  [(set (match_operand:V16QI 0 "register_operand" "=w,w,w")
+	(unspec:V16QI [(match_operand:V16QI 1 "general_operand" "w,r,Uni")]
+	 UNSPEC_AES_PROTECT))]
+  "TARGET_CRYPTO && fix_aes_erratum_1742098"
+  {
+    switch (which_alternative)
+      {
+      case 0: return "vmov\t%q0, %q1";
+      case 1: return "vmov\t%e0, %Q1, %R1  @ V16QI\;vmov\t%f0, %J1, %K1";
+      case 2: return output_move_neon (operands);
+      default: gcc_unreachable ();
+      }
+  }
+  [(set_attr "type" "neon_move_q,neon_from_gp_q,neon_load1_4reg")
+   (set_attr "length" "4,8,8")
+   (set_attr "arm_pool_range" "*,*,1020")
+   (set_attr "thumb2_pool_range" "*,*,1018")
+   (set_attr "neg_pool_range" "*,*,996")]
+)
+
+;; Another safe case is when a movmisalign load is used as the source.
+(define_insn "*aes_op_protect_misalign_load"
+  [(set (match_operand:V16QI 0 "s_register_operand" "=w")
+	(unspec:V16QI
+	 [(unspec:V16QI
+	   [(match_operand:V16QI 1 "neon_permissive_struct_operand" "Um")]
+	   UNSPEC_MISALIGNED_ACCESS)]
 	 UNSPEC_AES_PROTECT))]
   "TARGET_CRYPTO && fix_aes_erratum_1742098"
-  "vmov\\t%q0, %q1"
-  [(set_attr "type" "neon_move_q")]
+  "vld1.8\t%{q0}, %A1"
+  [(set_attr "type" "neon_load1_1reg_q")]
+)
+
+;; Similarly for the vld1 intrinsic
+(define_insn "aes_op_protect_neon_vld1v16qi"
+  [(set (match_operand:V16QI 0 "s_register_operand" "=w")
+        (unspec:V16QI
+	 [(unspec:V16QI [(match_operand:V16QI 1 "neon_struct_operand" "Um")]
+           UNSPEC_VLD1)]
+	 UNSPEC_AES_PROTECT))]
+  "TARGET_NEON"
+  "vld1.8\t%h0, %A1"
+  [(set_attr "type" "neon_load1_1reg_q")]
 )
 
 ;; An AESMC operation can feed directly into a subsequent AES