From c1027afe40865a9d276586a7066debe69549192b Mon Sep 17 00:00:00 2001
From: David Schleef <ds@schleef.org>
Date: Mon, 7 Jun 2010 11:41:35 -0700
Subject: [PATCH] x86: Add alternate rep movs memcpy implementation

---
 orc/orcprogram-sse.c |  6 ++++
 orc/orcx86.c         | 89 ++++++++++++++++++++++++++++++++++++++++++++++++++++
 orc/orcx86.h         |  4 +++
 3 files changed, 99 insertions(+)

diff --git a/orc/orcprogram-sse.c b/orc/orcprogram-sse.c
index e58043a..3490921 100644
--- a/orc/orcprogram-sse.c
+++ b/orc/orcprogram-sse.c
@@ -552,6 +552,12 @@ orc_compiler_sse_assemble (OrcCompiler *compiler)
 {
   int align_var;
 
+  if (0 && orc_x86_assemble_copy_check (compiler)) {
+    /* The rep movs implementation isn't faster most of the time */
+    orc_x86_assemble_copy (compiler);
+    return;
+  }
+
   align_var = get_align_var (compiler);
 
   compiler->vars[align_var].is_aligned = FALSE;
diff --git a/orc/orcx86.c b/orc/orcx86.c
index c410a72..7ca3dfc 100644
--- a/orc/orcx86.c
+++ b/orc/orcx86.c
@@ -698,6 +698,28 @@ void orc_x86_emit_emms (OrcCompiler *compiler)
   *compiler->codeptr++ = 0x77;
 }
 
+void orc_x86_emit_rep_movs (OrcCompiler *compiler, int size)
+{
+  switch (size) {
+    case 1:
+      ORC_ASM_CODE(compiler,"  rep movsb\n");
+      *compiler->codeptr++ = 0xf3;
+      *compiler->codeptr++ = 0xa4;
+      break;
+    case 2:
+      ORC_ASM_CODE(compiler,"  rep movsw\n");
+      *compiler->codeptr++ = 0x66;
+      *compiler->codeptr++ = 0xf3;
+      *compiler->codeptr++ = 0xa5;
+      break;
+    case 4:
+      ORC_ASM_CODE(compiler,"  rep movsl\n");
+      *compiler->codeptr++ = 0xf3;
+      *compiler->codeptr++ = 0xa5;
+      break;
+  }
+}
+
 void
 x86_add_fixup (OrcCompiler *compiler, unsigned char *ptr, int label, int type)
 {
@@ -898,3 +920,70 @@ orc_x86_emit_align (OrcCompiler *compiler)
   }
 }
 
+/* memcpy implementation based on rep movs */
+
+int
+orc_x86_assemble_copy_check (OrcCompiler *compiler)
+{
+  if (compiler->program->n_insns == 1 &&
+      compiler->program->is_2d == FALSE &&
+      (strcmp (compiler->program->insns[0].opcode->name, "copyb") == 0 ||
+      strcmp (compiler->program->insns[0].opcode->name, "copyw") == 0 ||
+      strcmp (compiler->program->insns[0].opcode->name, "copyl") == 0)) {
+    return TRUE;
+  }
+
+  return FALSE;
+}
+
+void
+orc_x86_assemble_copy (OrcCompiler *compiler)
+{
+  OrcInstruction *insn;
+  int shift = 0;
+
+  insn = compiler->program->insns + 0;
+
+  if (strcmp (insn->opcode->name, "copyw") == 0) {
+    shift = 1;
+  } else if (strcmp (insn->opcode->name, "copyl") == 0) {
+    shift = 2;
+  }
+
+  compiler->used_regs[X86_EDI] = TRUE;
+  compiler->used_regs[X86_ESI] = TRUE;
+
+  orc_x86_emit_prologue (compiler);
+
+  orc_x86_emit_mov_memoffset_reg (compiler, 4,
+      (int)ORC_STRUCT_OFFSET(OrcExecutor,arrays[insn->dest_args[0]]),
+      compiler->exec_reg, X86_EDI);
+  orc_x86_emit_mov_memoffset_reg (compiler, 4,
+      (int)ORC_STRUCT_OFFSET(OrcExecutor,arrays[insn->src_args[0]]),
+      compiler->exec_reg, X86_ESI);
+  orc_x86_emit_mov_memoffset_reg (compiler, 4,
+      (int)ORC_STRUCT_OFFSET(OrcExecutor,n), compiler->exec_reg,
+      X86_ECX);
+
+  orc_x86_emit_sar_imm_reg (compiler, 4, 2 - shift, X86_ECX);
+  orc_x86_emit_rep_movs (compiler, 4);
+  if (shift == 0) {
+    orc_x86_emit_mov_memoffset_reg (compiler, 4,
+        (int)ORC_STRUCT_OFFSET(OrcExecutor,n), compiler->exec_reg,
+        X86_ECX);
+    orc_x86_emit_and_imm_reg (compiler, 4, 3, X86_ECX);
+    orc_x86_emit_rep_movs (compiler, 1);
+  }
+  if (shift == 1) {
+    orc_x86_emit_mov_memoffset_reg (compiler, 4,
+        (int)ORC_STRUCT_OFFSET(OrcExecutor,n), compiler->exec_reg,
+        X86_ECX);
+    orc_x86_emit_and_imm_reg (compiler, 4, 1, X86_ECX);
+    orc_x86_emit_rep_movs (compiler, 2);
+  }
+
+  orc_x86_emit_epilogue (compiler);
+
+  orc_x86_do_fixups (compiler);
+}
+
diff --git a/orc/orcx86.h b/orc/orcx86.h
index 0c2f763..82baee3 100644
--- a/orc/orcx86.h
+++ b/orc/orcx86.h
@@ -56,6 +56,7 @@ void orc_x86_emit_cmp_imm_memoffset (OrcCompiler *compiler, int size, int value,
 void orc_x86_emit_test_imm_memoffset (OrcCompiler *compiler, int size, int value,
     int offset, int reg);
 void orc_x86_emit_emms (OrcCompiler *compiler);
+void orc_x86_emit_rep_movs (OrcCompiler *compiler, int size);
 void orc_x86_emit_ret (OrcCompiler *compiler);
 void orc_x86_emit_jle (OrcCompiler *compiler, int label);
 void orc_x86_emit_je (OrcCompiler *compiler, int label);
@@ -72,6 +73,9 @@ void orc_x86_emit_modrm_reg (OrcCompiler *compiler, int reg1, int reg2);
 
 void orc_x86_do_fixups (OrcCompiler *compiler);
 
+int orc_x86_assemble_copy_check (OrcCompiler *compiler);
+void orc_x86_assemble_copy (OrcCompiler *compiler);
+
 
 #endif
 
-- 
2.7.4