From 65aaf91aefc54d9370538a48eef7f659495b2437 Mon Sep 17 00:00:00 2001
From: David Schleef <ds@schleef.org>
Date: Tue, 6 Jul 2010 12:31:47 -0700
Subject: [PATCH] neon: Use unaligned ops instead of mask register

---
 orc/orcprogram-neon.c | 14 +++++++++++---
 orc/orcrules-neon.c   | 31 +++++++++++++++++++++++++++++++
 2 files changed, 42 insertions(+), 3 deletions(-)

diff --git a/orc/orcprogram-neon.c b/orc/orcprogram-neon.c
index f27b034..b2d0a95 100644
--- a/orc/orcprogram-neon.c
+++ b/orc/orcprogram-neon.c
@@ -220,7 +220,9 @@ orc_compiler_neon_init (OrcCompiler *compiler)
     compiler->loop_shift = loop_shift;
   }
 
-  compiler->need_mask_regs = TRUE;
+  if (0) {
+    compiler->need_mask_regs = TRUE;
+  }
 }
 
 void
@@ -670,7 +672,11 @@ orc_compiler_neon_assemble (OrcCompiler *compiler)
   orc_arm_emit_cmp_imm (compiler, ORC_ARM_IP, 0);
   orc_arm_emit_branch (compiler, ORC_ARM_COND_EQ, 3);
 
-  orc_neon_load_alignment_masks (compiler);
+  if (0) {
+    /* Disable alignment masks for now.  It can easily take all available
+       registers. */
+    orc_neon_load_alignment_masks (compiler);
+  }
 
   orc_arm_emit_label (compiler, 2);
   orc_neon_emit_loop (compiler);
@@ -678,7 +684,9 @@ orc_compiler_neon_assemble (OrcCompiler *compiler)
   orc_arm_emit_cmp_imm (compiler, ORC_ARM_IP, 0);
   orc_arm_emit_branch (compiler, ORC_ARM_COND_NE, 2);
 
-  orc_neon_restore_unalignment (compiler);
+  if (0) {
+    orc_neon_restore_unalignment (compiler);
+  }
 
   orc_arm_emit_label (compiler, 3);
 
diff --git a/orc/orcrules-neon.c b/orc/orcrules-neon.c
index c0e24d9..2135f91 100644
--- a/orc/orcrules-neon.c
+++ b/orc/orcrules-neon.c
@@ -245,6 +245,18 @@ orc_neon_load_vec_unaligned (OrcCompiler *compiler, OrcVariable *var,
 {
   orc_uint32 code;
 
+  ORC_ASM_CODE(compiler,"  vld1.8 %s, [%s]%s\n",
+      orc_neon_reg_name (var->alloc),
+      orc_arm_reg_name (var->ptr_register),
+      update ? "!" : "");
+  code = 0xf420070d;
+  code |= (var->ptr_register&0xf) << 16;
+  code |= ((var->alloc)&0xf) << 12;
+  code |= (((var->alloc)>>4)&0x1) << 22;
+  code |= (!update) << 1;
+  orc_arm_emit (compiler, code);
+#if 0
+  /* used with need_mask_regs */
   ORC_ASM_CODE(compiler,"  vld1.64 %s, [%s]%s\n",
       orc_neon_reg_name (var->aligned_data + 1),
       orc_arm_reg_name (var->ptr_register),
@@ -267,6 +279,7 @@ orc_neon_load_vec_unaligned (OrcCompiler *compiler, OrcVariable *var,
 //orc_neon_emit_mov (compiler, var->alloc, var->mask_alloc);
 
   orc_neon_emit_mov (compiler, var->aligned_data, var->aligned_data + 1);
+#endif
 }
 
 void
@@ -275,6 +288,23 @@ orc_neon_load_halfvec_unaligned (OrcCompiler *compiler, OrcVariable *var,
 {
   orc_uint32 code;
 
+  ORC_ASM_CODE(compiler,"  vld1.8 %s, [%s]\n",
+      orc_neon_reg_name (var->alloc),
+      orc_arm_reg_name (var->ptr_register));
+  code = 0xf420070d;
+  code |= (var->ptr_register&0xf) << 16;
+  code |= ((var->alloc)&0xf) << 12;
+  code |= (((var->alloc)>>4)&0x1) << 22;
+  //code |= (!update) << 1;
+  code |= (1) << 1;
+  orc_arm_emit (compiler, code);
+
+  if (update) {
+    orc_arm_emit_add_imm (compiler, var->ptr_register,
+        var->ptr_register, 4);
+  }
+#if 0
+  /* used with need_mask_regs */
   ORC_ASM_CODE(compiler,"  vld1.32 %s[1], [%s]%s\n",
       orc_neon_reg_name (var->aligned_data),
       orc_arm_reg_name (var->ptr_register),
@@ -296,6 +326,7 @@ orc_neon_load_halfvec_unaligned (OrcCompiler *compiler, OrcVariable *var,
 
   orc_neon_emit_unary (compiler, "vrev64.i32", 0xf3b80000,
       var->aligned_data, var->aligned_data);
+#endif
 }
 
 void
-- 
2.7.4