opcode: mark copy opcodes and don't emit mov for them

There is no need to emit a mov operation when the source and dest
registers of a copy operation are different, that's kindof the point of
the copy operation. Saves some useless movs whenever copyX is used.

3 files changed, 6 insertions, 5 deletions

diff --git a/orc/orcopcode.h b/orc/orcopcode.h
index e209f98..63310c9 100644
--- a/orc/orcopcode.h
+++ b/orc/orcopcode.h
@@ -34,6 +34,7 @@ struct _OrcOpcodeSet {
 #define ORC_STATIC_OPCODE_STORE (1<<5)
 #define ORC_STATIC_OPCODE_INVARIANT (1<<6)
 #define ORC_STATIC_OPCODE_ITERATOR (1<<7)
+#define ORC_STATIC_OPCODE_COPY (1<<8)
 
 
 struct _OrcStaticOpcode {
diff --git a/orc/orcopcodes.c b/orc/orcopcodes.c
index eda8610..820f0c9 100644
--- a/orc/orcopcodes.c
+++ b/orc/orcopcodes.c
@@ -296,7 +296,7 @@ static OrcStaticOpcode opcodes[] = {
   { "avgub", 0, { 1 }, { 1, 1 }, emulate_avgub },
   { "cmpeqb", 0, { 1 }, { 1, 1 }, emulate_cmpeqb },
   { "cmpgtsb", 0, { 1 }, { 1, 1 }, emulate_cmpgtsb },
-  { "copyb", 0, { 1 }, { 1 }, emulate_copyb },
+  { "copyb", ORC_STATIC_OPCODE_COPY, { 1 }, { 1 }, emulate_copyb },
   { "loadb", ORC_STATIC_OPCODE_LOAD, { 1 }, { 1 }, emulate_loadb },
   { "loadoffb", ORC_STATIC_OPCODE_LOAD|ORC_STATIC_OPCODE_SCALAR, { 1 }, { 1, 4 }, emulate_loadoffb },
   { "loadupdb", ORC_STATIC_OPCODE_LOAD|ORC_STATIC_OPCODE_ITERATOR, { 1 }, { 1 }, emulate_loadupdb },
@@ -335,7 +335,7 @@ static OrcStaticOpcode opcodes[] = {
   { "avguw", 0, { 2 }, { 2, 2 }, emulate_avguw },
   { "cmpeqw", 0, { 2 }, { 2, 2 }, emulate_cmpeqw },
   { "cmpgtsw", 0, { 2 }, { 2, 2 }, emulate_cmpgtsw },
-  { "copyw", 0, { 2 }, { 2 }, emulate_copyw },
+  { "copyw", ORC_STATIC_OPCODE_COPY, { 2 }, { 2 }, emulate_copyw },
   { "div255w", 0, { 2 }, { 2 }, emulate_div255w },
   { "divluw", 0, { 2 }, { 2, 2 }, emulate_divluw },
   { "loadw", ORC_STATIC_OPCODE_LOAD, { 2 }, { 2 }, emulate_loadw },
@@ -370,7 +370,7 @@ static OrcStaticOpcode opcodes[] = {
   { "avgul", 0, { 4 }, { 4, 4 }, emulate_avgul },
   { "cmpeql", 0, { 4 }, { 4, 4 }, emulate_cmpeql },
   { "cmpgtsl", 0, { 4 }, { 4, 4 }, emulate_cmpgtsl },
-  { "copyl", 0, { 4 }, { 4 }, emulate_copyl },
+  { "copyl", ORC_STATIC_OPCODE_COPY, { 4 }, { 4 }, emulate_copyl },
   { "loadl", ORC_STATIC_OPCODE_LOAD, { 4 }, { 4 }, emulate_loadl },
   { "loadoffl", ORC_STATIC_OPCODE_LOAD|ORC_STATIC_OPCODE_SCALAR, { 4 }, { 4, 4 }, emulate_loadoffl },
   { "loadpl", ORC_STATIC_OPCODE_LOAD|ORC_STATIC_OPCODE_SCALAR|ORC_STATIC_OPCODE_INVARIANT, { 4 }, { 4 }, emulate_loadpl },
@@ -396,7 +396,7 @@ static OrcStaticOpcode opcodes[] = {
   { "loadpq", ORC_STATIC_OPCODE_LOAD|ORC_STATIC_OPCODE_SCALAR|ORC_STATIC_OPCODE_INVARIANT, { 8 }, { 8 }, emulate_loadpq },
   { "storeq", ORC_STATIC_OPCODE_STORE, { 8 }, { 8 }, emulate_storeq },
   { "splatw3q", 0, { 8 }, { 8 }, emulate_splatw3q },
-  { "copyq", 0, { 8 }, { 8 }, emulate_copyq },
+  { "copyq", ORC_STATIC_OPCODE_COPY, { 8 }, { 8 }, emulate_copyq },
   { "cmpeqq", 0, { 8 }, { 8, 8 }, emulate_cmpeqq },
   { "cmpgtsq", 0, { 8 }, { 8, 8 }, emulate_cmpgtsq },
   { "andq", 0, { 8 }, { 8, 8 }, emulate_andq },
diff --git a/orc/orcprogram-sse.c b/orc/orcprogram-sse.c
index abeeff5..51c1536 100644
--- a/orc/orcprogram-sse.c
+++ b/orc/orcprogram-sse.c
@@ -1029,7 +1029,7 @@ orc_sse_emit_loop (OrcCompiler *compiler, int offset, int update)
 
     rule = insn->rule;
     if (rule && rule->emit) {
-      if (!(insn->opcode->flags & (ORC_STATIC_OPCODE_ACCUMULATOR|ORC_STATIC_OPCODE_LOAD|ORC_STATIC_OPCODE_STORE)) &&
+      if (!(insn->opcode->flags & (ORC_STATIC_OPCODE_ACCUMULATOR|ORC_STATIC_OPCODE_LOAD|ORC_STATIC_OPCODE_STORE|ORC_STATIC_OPCODE_COPY)) &&
           compiler->vars[insn->dest_args[0]].alloc !=
           compiler->vars[insn->src_args[0]].alloc) {
 #ifdef MMX