Add swapq opcode

diff --git a/orc/opcodes.h b/orc/opcodes.h
index 80cef73..d9c28c7 100644 (file)
--- a/orc/opcodes.h
+++ b/orc/opcodes.h
@@ -121,6 +121,7 @@ UNARY_LW(select0lw, "(orc_uint32)%s & 0xffff")
 UNARY_LW(select1lw, "((orc_uint32)%s >> 16)&0xffff")
 UNARY_UW(swapw, "ORC_SWAP_W(%s)")
 UNARY_UL(swapl, "ORC_SWAP_L(%s)")
+UNARY_UQ(swapq, "ORC_SWAP_Q(%s)")
 
 BINARY_F(addf, "ORC_DENORMAL(ORC_DENORMAL(%s) + ORC_DENORMAL(%s))")
 BINARY_F(subf, "ORC_DENORMAL(ORC_DENORMAL(%s) - ORC_DENORMAL(%s))")

diff --git a/orc/orcemulateopcodes.c b/orc/orcemulateopcodes.c
index 3bd7175..a0d95eb 100644 (file)
--- a/orc/orcemulateopcodes.c
+++ b/orc/orcemulateopcodes.c
@@ -33,6 +33,7 @@
 #define ORC_CLAMP_UL(x) ORC_CLAMP(x,ORC_UL_MIN,ORC_UL_MAX)
 #define ORC_SWAP_W(x) ((((x)&0xff)<<8) | (((x)&0xff00)>>8))
 #define ORC_SWAP_L(x) ((((x)&0xff)<<24) | (((x)&0xff00)<<8) | (((x)&0xff0000)>>8) | (((x)&0xff000000)>>24))
+#define ORC_SWAP_Q(x) ((((x)&0xffU)<<56) | (((x)&0xff00U)<<40) | (((x)&0xff0000U)<<24) | (((x)&0xff000000U)<<8) | (((x)&0xff00000000U)>>8) | (((x)&0xff0000000000U)>>24) | (((x)&0xff000000000000U)>>40) | (((x)&0xff00000000000000U)>>56))
 #define ORC_PTR_OFFSET(ptr,offset) ((void *)(((unsigned char *)(ptr)) + (offset)))
 #define ORC_MIN_NORMAL (1.1754944909521339405e-38)
 #define ORC_RECAST_INT(x) (((orc_union32)(x)).i)
@@ -3525,6 +3526,29 @@ emulate_swapl (OrcOpcodeExecutor *ex, int offset, int n)
 }
 
 void
+emulate_swapq (OrcOpcodeExecutor *ex, int offset, int n)
+{
+  int i;
+  orc_union64 * ORC_RESTRICT ptr0;
+  const orc_union64 * ORC_RESTRICT ptr4;
+  orc_union64 var32;
+  orc_union64 var33;
+
+  ptr0 = (orc_union64 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union64 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadq */
+    var32 = ptr4[i];
+    /* 1: swapq */
+    var33.i = ORC_SWAP_Q(var32.i);
+    /* 2: storeq */
+    ptr0[i] = var33;
+  }
+
+}
+
+void
 emulate_select0wb (OrcOpcodeExecutor *ex, int offset, int n)
 {
   int i;

diff --git a/orc/orcemulateopcodes.h b/orc/orcemulateopcodes.h
index 5ea3a79..9ab736e 100644 (file)
--- a/orc/orcemulateopcodes.h
+++ b/orc/orcemulateopcodes.h
@@ -143,6 +143,7 @@ void emulate_accl (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_accsadubl (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_swapw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_swapl (OrcOpcodeExecutor *ex, int i, int n);
+void emulate_swapq (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_select0wb (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_select1wb (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_select0lw (OrcOpcodeExecutor *ex, int i, int n);

diff --git a/orc/orcfunctions.c b/orc/orcfunctions.c
index 24e84e2..80dc08f 100644 (file)
--- a/orc/orcfunctions.c
+++ b/orc/orcfunctions.c
@@ -79,6 +79,7 @@ void orc_memset (void * d1, int p1, int n);
 #define ORC_CLAMP_UL(x) ORC_CLAMP(x,ORC_UL_MIN,ORC_UL_MAX)
 #define ORC_SWAP_W(x) ((((x)&0xff)<<8) | (((x)&0xff00)>>8))
 #define ORC_SWAP_L(x) ((((x)&0xff)<<24) | (((x)&0xff00)<<8) | (((x)&0xff0000)>>8) | (((x)&0xff000000)>>24))
+#define ORC_SWAP_Q(x) ((((x)&0xffU)<<56) | (((x)&0xff00U)<<40) | (((x)&0xff0000U)<<24) | (((x)&0xff000000U)<<8) | (((x)&0xff00000000U)>>8) | (((x)&0xff0000000000U)>>24) | (((x)&0xff000000000000U)>>40) | (((x)&0xff00000000000000U)>>56))
 #define ORC_PTR_OFFSET(ptr,offset) ((void *)(((unsigned char *)(ptr)) + (offset)))
 #define ORC_MIN_NORMAL (1.1754944909521339405e-38)
 #define ORC_RECAST_INT(x) (((orc_union32)(x)).i)

diff --git a/orc/orcopcodes.c b/orc/orcopcodes.c
index dd9e766..6c2a0a6 100644 (file)
--- a/orc/orcopcodes.c
+++ b/orc/orcopcodes.c
@@ -442,6 +442,7 @@ static OrcStaticOpcode opcodes[] = {
 
   { "swapw", 0, { 2 }, { 2 }, emulate_swapw },
   { "swapl", 0, { 4 }, { 4 }, emulate_swapl },
+  { "swapq", 0, { 8 }, { 8 }, emulate_swapq },
   { "select0wb", 0, { 1 }, { 2 }, emulate_select0wb },
   { "select1wb", 0, { 1 }, { 2 }, emulate_select1wb },
   { "select0lw", 0, { 2 }, { 4 }, emulate_select0lw },

diff --git a/orc/orcprogram-c.c b/orc/orcprogram-c.c
index ac72f91..5513b40 100644 (file)
--- a/orc/orcprogram-c.c
+++ b/orc/orcprogram-c.c
@@ -101,6 +101,7 @@ orc_target_c_get_asm_preamble (void)
     "#define ORC_CLAMP_UL(x) ORC_CLAMP(x,ORC_UL_MIN,ORC_UL_MAX)\n"
     "#define ORC_SWAP_W(x) ((((x)&0xff)<<8) | (((x)&0xff00)>>8))\n"
     "#define ORC_SWAP_L(x) ((((x)&0xff)<<24) | (((x)&0xff00)<<8) | (((x)&0xff0000)>>8) | (((x)&0xff000000)>>24))\n"
+    "#define ORC_SWAP_Q(x) ((((x)&0xffU)<<56) | (((x)&0xff00U)<<40) | (((x)&0xff0000U)<<24) | (((x)&0xff000000U)<<8) | (((x)&0xff00000000U)>>8) | (((x)&0xff0000000000U)>>24) | (((x)&0xff000000000000U)>>40) | (((x)&0xff00000000000000U)>>56))\n"
     "#define ORC_PTR_OFFSET(ptr,offset) ((void *)(((unsigned char *)(ptr)) + (offset)))\n"
     "#define ORC_MIN_NORMAL (1.1754944909521339405e-38)\n"
     "#define ORC_RECAST_INT(x) (((orc_union32)(x)).i)\n"

diff --git a/orc/orcrules-sse.c b/orc/orcrules-sse.c
index 24043c9..e6bbbea 100644 (file)
--- a/orc/orcrules-sse.c
+++ b/orc/orcrules-sse.c
@@ -1437,6 +1437,27 @@ sse_rule_swapl (OrcCompiler *p, void *user, OrcInstruction *insn)
   orc_sse_emit_por (p, tmp, dest);
 }
 
+static void
+sse_rule_swapq (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  int src = p->vars[insn->src_args[0]].alloc;
+  int dest = p->vars[insn->dest_args[0]].alloc;
+  int tmp = orc_compiler_get_temp_reg (p);
+
+  orc_sse_emit_movdqa (p, src, tmp);
+  orc_sse_emit_psllq (p, 32, tmp);
+  orc_sse_emit_psrlq (p, 32, dest);
+  orc_sse_emit_por (p, tmp, dest);
+  orc_sse_emit_movdqa (p, dest, tmp);
+  orc_sse_emit_pslld (p, 16, tmp);
+  orc_sse_emit_psrld (p, 16, dest);
+  orc_sse_emit_por (p, tmp, dest);
+  orc_sse_emit_movdqa (p, dest, tmp);
+  orc_sse_emit_psllw (p, 8, tmp);
+  orc_sse_emit_psrlw (p, 8, dest);
+  orc_sse_emit_por (p, tmp, dest);
+}
+
 #define LOAD_MASK_IS_SLOW
 #ifndef LOAD_MASK_IS_SLOW
 static void
@@ -2193,6 +2214,7 @@ orc_compiler_sse_register_rules (OrcTarget *target)
   orc_rule_register (rule_set, "absl", sse_rule_absl_slow, NULL);
   orc_rule_register (rule_set, "swapw", sse_rule_swapw, NULL);
   orc_rule_register (rule_set, "swapl", sse_rule_swapl, NULL);
+  orc_rule_register (rule_set, "swapq", sse_rule_swapq, NULL);
   orc_rule_register (rule_set, "splitlw", sse_rule_splitlw, NULL);
   orc_rule_register (rule_set, "splitwb", sse_rule_splitwb, NULL);
   orc_rule_register (rule_set, "avgsl", sse_rule_avgsl, NULL);