From: David Schleef <ds@schleef.org>
Date: Thu, 8 Jul 2010 23:53:29 +0000 (-0700)
Subject: neon: Add short-array unaligned speedup
X-Git-Tag: orc-0.4.6~10
X-Git-Url: http://review.tizen.org/git/?a=commitdiff_plain;h=8f807687c895ff388b2d69a15cca37766eb2a119;p=platform%2Fupstream%2Forc.git

neon: Add short-array unaligned speedup
---

diff --git a/orc/orcprogram-neon.c b/orc/orcprogram-neon.c
index 38defdd..8161f12 100644
--- a/orc/orcprogram-neon.c
+++ b/orc/orcprogram-neon.c
@@ -595,6 +595,8 @@ get_align_var (OrcCompiler *compiler)
 enum {
   LABEL_ONE_REGION = 1,
   LABEL_ONE_REGION_AFTER,
+  LABEL_REGION0_LOOP,
+  LABEL_REGION0_SKIP,
   LABEL_REGION1_LOOP,
   LABEL_REGION1_SKIP,
   LABEL_REGION2_LOOP_SMALL,
@@ -643,6 +645,42 @@ orc_compiler_neon_assemble (OrcCompiler *compiler)
     orc_arm_emit_label (compiler, LABEL_OUTER_LOOP);
   }
 
+#define ORC_NEON_ALIGNED_DEST_CUTOFF 64
+
+  if (compiler->loop_shift > 0) {
+    orc_arm_emit_load_reg (compiler, ORC_ARM_A3, compiler->exec_reg,
+        (int)ORC_STRUCT_OFFSET(OrcExecutor,n));
+    orc_arm_emit_cmp_imm (compiler, ORC_ARM_A3, ORC_NEON_ALIGNED_DEST_CUTOFF);
+    orc_arm_emit_branch (compiler, ORC_ARM_COND_GT, LABEL_REGION0_SKIP);
+
+    orc_arm_emit_asr_imm (compiler, ORC_ARM_A2, ORC_ARM_A3,
+        compiler->loop_shift);
+    orc_arm_emit_store_reg (compiler, ORC_ARM_A2, compiler->exec_reg,
+        (int)ORC_STRUCT_OFFSET(OrcExecutor,counter2));
+
+    orc_arm_emit_and_imm (compiler, ORC_ARM_A3, ORC_ARM_A3,
+        (1<<compiler->loop_shift)-1);
+    orc_arm_emit_store_reg (compiler, ORC_ARM_A3, compiler->exec_reg,
+        (int)ORC_STRUCT_OFFSET(OrcExecutor,counter3));
+
+    orc_neon_load_constants_inner (compiler);
+    orc_arm_emit_load_reg (compiler, ORC_ARM_IP, compiler->exec_reg,
+        (int)ORC_STRUCT_OFFSET(OrcExecutor,counter2));
+    orc_arm_emit_cmp_imm (compiler, ORC_ARM_IP, 0);
+    orc_arm_emit_branch (compiler, ORC_ARM_COND_EQ, LABEL_REGION2_SKIP);
+
+    compiler->size_region = 0;
+    orc_arm_emit_label (compiler, LABEL_REGION0_LOOP);
+    orc_arm_emit_sub_imm (compiler, ORC_ARM_IP, ORC_ARM_IP, 1, TRUE);
+    orc_neon_emit_loop (compiler, -1);
+    orc_arm_emit_branch (compiler, ORC_ARM_COND_NE, LABEL_REGION0_LOOP);
+    
+
+    orc_arm_emit_branch (compiler, ORC_ARM_COND_AL, LABEL_REGION2_SKIP);
+
+    orc_arm_emit_label (compiler, LABEL_REGION0_SKIP);
+  }
+
   if (compiler->loop_shift > 0) {
     orc_arm_emit_load_imm (compiler, ORC_ARM_IP, 1<<align_shift);
 
@@ -702,8 +740,8 @@ orc_compiler_neon_assemble (OrcCompiler *compiler)
     orc_arm_emit_branch (compiler, ORC_ARM_COND_EQ, LABEL_REGION1_SKIP);
 
     orc_arm_emit_label (compiler, LABEL_REGION1_LOOP);
-    orc_neon_emit_loop (compiler, 0);
     orc_arm_emit_sub_imm (compiler, ORC_ARM_IP, ORC_ARM_IP, 1, TRUE);
+    orc_neon_emit_loop (compiler, -1);
     orc_arm_emit_branch (compiler, ORC_ARM_COND_NE, LABEL_REGION1_LOOP);
     orc_arm_emit_label (compiler, LABEL_REGION1_SKIP);
 
@@ -789,8 +827,8 @@ orc_compiler_neon_assemble (OrcCompiler *compiler)
     orc_arm_emit_branch (compiler, ORC_ARM_COND_EQ, LABEL_REGION3_SKIP);
 
     orc_arm_emit_label (compiler, LABEL_REGION3_LOOP);
-    orc_neon_emit_loop (compiler, 0);
     orc_arm_emit_sub_imm (compiler, ORC_ARM_IP, ORC_ARM_IP, 1, TRUE);
+    orc_neon_emit_loop (compiler, -1);
     orc_arm_emit_branch (compiler, ORC_ARM_COND_NE, LABEL_REGION3_LOOP);
     orc_arm_emit_label (compiler, LABEL_REGION3_SKIP);
 
diff --git a/orc/orcrules-neon.c b/orc/orcrules-neon.c
index a003c3e..98306c0 100644
--- a/orc/orcrules-neon.c
+++ b/orc/orcrules-neon.c
@@ -591,6 +591,20 @@ orc_neon_storeb (OrcCompiler *compiler, int dest, int update, int src1, int is_a
     code |= ((src1>>4)&0x1) << 22;
     code |= (!update) << 1;
     orc_arm_emit (compiler, code);
+  } else if (compiler->loop_shift == 5) {
+    ORC_ASM_CODE(compiler,"  vst1.8 { %s, %s, %s, %s }, [%s]%s\n",
+        orc_neon_reg_name (src1),
+        orc_neon_reg_name (src1+1),
+        orc_neon_reg_name (src1+2),
+        orc_neon_reg_name (src1+3),
+        orc_arm_reg_name (dest),
+        update ? "!" : "");
+    code = 0xf400020d;
+    code |= (dest&0xf) << 16;
+    code |= (src1&0xf) << 12;
+    code |= ((src1>>4)&0x1) << 22;
+    code |= (!update) << 1;
+    orc_arm_emit (compiler, code);
   } else if (is_aligned && compiler->loop_shift == 4) {
     ORC_ASM_CODE(compiler,"  vst1.8 { %s, %s }, [%s,:128]%s\n",
         orc_neon_reg_name (src1),
@@ -603,6 +617,18 @@ orc_neon_storeb (OrcCompiler *compiler, int dest, int update, int src1, int is_a
     code |= ((src1>>4)&0x1) << 22;
     code |= (!update) << 1;
     orc_arm_emit (compiler, code);
+  } else if (compiler->loop_shift == 4) {
+    ORC_ASM_CODE(compiler,"  vst1.8 { %s, %s }, [%s]%s\n",
+        orc_neon_reg_name (src1),
+        orc_neon_reg_name (src1+1),
+        orc_arm_reg_name (dest),
+        update ? "!" : "");
+    code = 0xf4000a0d;
+    code |= (dest&0xf) << 16;
+    code |= (src1&0xf) << 12;
+    code |= ((src1>>4)&0x1) << 22;
+    code |= (!update) << 1;
+    orc_arm_emit (compiler, code);
   } else if (is_aligned && compiler->loop_shift == 3) {
     ORC_ASM_CODE(compiler,"  vst1.8 %s, [%s,:64]%s\n",
         orc_neon_reg_name (src1),