From 4b6619f8d21feabf88bc07323bf51b54c4c77e5c Mon Sep 17 00:00:00 2001
From: David Schleef <ds@schleef.org>
Date: Fri, 6 Aug 2010 18:43:41 -0700
Subject: [PATCH] Add low-level opcodes for compositing

---
 orc/opcodes.h           |   2 +
 orc/orcemulateopcodes.c | 138 ++++++++++++++++++++++++++++++++++++++++++++++++
 orc/orcemulateopcodes.h |   6 +++
 orc/orcopcodes.c        |   7 +++
 orc/orcprogram-c.c      |  54 +++++++++++++++++++
 orc/orcrules-sse.c      |  94 +++++++++++++++++++++++++++++++++
 orc/orcsse.c            |  14 +++++
 orc/orcsse.h            |   1 +
 8 files changed, 316 insertions(+)

diff --git a/orc/opcodes.h b/orc/opcodes.h
index 4db81a8..b324d5d 100644
--- a/orc/opcodes.h
+++ b/orc/opcodes.h
@@ -90,6 +90,7 @@ UNARY_SQ(loadpq, "%s")
 UNARY_BW(convsbw, "%s")
 UNARY_BW(convubw, "(orc_uint8)%s")
 UNARY_WB(convwb, "%s")
+UNARY_WB(convhwb, "((orc_uint16)%s)>>8")
 UNARY_WB(convssswb, "ORC_CLAMP_SB(%s)")
 UNARY_WB(convsuswb, "ORC_CLAMP_UB(%s)")
 UNARY_WB(convusswb, "ORC_CLAMP_SB((orc_uint16)%s)")
@@ -98,6 +99,7 @@ UNARY_WB(convuuswb, "ORC_CLAMP_UB((orc_uint16)%s)")
 UNARY_WL(convswl, "%s")
 UNARY_WL(convuwl, "(orc_uint16)%s")
 UNARY_LW(convlw, "%s")
+UNARY_WB(convhlw, "((orc_uint32)%s)>>16")
 UNARY_LW(convssslw, "ORC_CLAMP_SW(%s)")
 UNARY_LW(convsuslw, "ORC_CLAMP_UW(%s)")
 UNARY_LW(convusslw, "ORC_CLAMP_SW((orc_uint32)%s)")
diff --git a/orc/orcemulateopcodes.c b/orc/orcemulateopcodes.c
index 5865d0d..f8cac0a 100644
--- a/orc/orcemulateopcodes.c
+++ b/orc/orcemulateopcodes.c
@@ -1189,6 +1189,29 @@ emulate_copyw (OrcOpcodeExecutor *ex, int offset, int n)
 }
 
 void
+emulate_div255w (OrcOpcodeExecutor *ex, int offset, int n)
+{
+  int i;
+  orc_int16 * ptr0;
+  const orc_int16 * ptr4;
+  orc_int16 var32;
+  orc_int16 var33;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadw */
+    var32 = ptr4[i];
+    /* 1: div255w */
+    var33 = ((uint16_t)(((orc_uint16)(var32+128)) + (((orc_uint16)(var32+128))>>8)))>>8;
+    /* 2: storew */
+    ptr0[i] = var33;
+  }
+
+}
+
+void
 emulate_loadw (OrcOpcodeExecutor *ex, int offset, int n)
 {
   int i;
@@ -2568,6 +2591,29 @@ emulate_storeq (OrcOpcodeExecutor *ex, int offset, int n)
 }
 
 void
+emulate_splatw0q (OrcOpcodeExecutor *ex, int offset, int n)
+{
+  int i;
+  orc_union64 * ptr0;
+  const orc_union64 * 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: splatw0q */
+    var33.i = ((orc_uint64)(var32.i&0xffff) << 48) | ((orc_uint64)(var32.i&0xffff)<<32) | ((var32.i&0xffff) << 16) | (var32.i&0xffff);
+    /* 2: storeq */
+    ptr0[i] = var33;
+  }
+
+}
+
+void
 emulate_convsbw (OrcOpcodeExecutor *ex, int offset, int n)
 {
   int i;
@@ -2614,6 +2660,52 @@ emulate_convubw (OrcOpcodeExecutor *ex, int offset, int n)
 }
 
 void
+emulate_splatbw (OrcOpcodeExecutor *ex, int offset, int n)
+{
+  int i;
+  orc_int16 * ptr0;
+  const orc_int8 * ptr4;
+  orc_int8 var32;
+  orc_int16 var33;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadb */
+    var32 = ptr4[i];
+    /* 1: splatbw */
+    var33 = ((var32&0xff) << 8) | (var32&0xff);
+    /* 2: storew */
+    ptr0[i] = var33;
+  }
+
+}
+
+void
+emulate_splatbl (OrcOpcodeExecutor *ex, int offset, int n)
+{
+  int i;
+  orc_union32 * ptr0;
+  const orc_int8 * ptr4;
+  orc_int8 var32;
+  orc_union32 var33;
+
+  ptr0 = (orc_union32 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int8 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadb */
+    var32 = ptr4[i];
+    /* 1: splatbl */
+    var33.i = ((var32&0xff) << 24) | ((var32&0xff)<<16) | ((var32&0xff) << 8) | (var32&0xff);
+    /* 2: storel */
+    ptr0[i] = var33;
+  }
+
+}
+
+void
 emulate_convswl (OrcOpcodeExecutor *ex, int offset, int n)
 {
   int i;
@@ -2729,6 +2821,29 @@ emulate_convwb (OrcOpcodeExecutor *ex, int offset, int n)
 }
 
 void
+emulate_convhwb (OrcOpcodeExecutor *ex, int offset, int n)
+{
+  int i;
+  orc_int8 * ptr0;
+  const orc_int16 * ptr4;
+  orc_int16 var32;
+  orc_int8 var33;
+
+  ptr0 = (orc_int8 *)ex->dest_ptrs[0];
+  ptr4 = (orc_int16 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadw */
+    var32 = ptr4[i];
+    /* 1: convhwb */
+    var33 = ((orc_uint16)var32)>>8;
+    /* 2: storeb */
+    ptr0[i] = var33;
+  }
+
+}
+
+void
 emulate_convssswb (OrcOpcodeExecutor *ex, int offset, int n)
 {
   int i;
@@ -2844,6 +2959,29 @@ emulate_convlw (OrcOpcodeExecutor *ex, int offset, int n)
 }
 
 void
+emulate_convhlw (OrcOpcodeExecutor *ex, int offset, int n)
+{
+  int i;
+  orc_int16 * ptr0;
+  const orc_union32 * ptr4;
+  orc_union32 var32;
+  orc_int16 var33;
+
+  ptr0 = (orc_int16 *)ex->dest_ptrs[0];
+  ptr4 = (orc_union32 *)ex->src_ptrs[0];
+
+  for (i = 0; i < n; i++) {
+    /* 0: loadl */
+    var32 = ptr4[i];
+    /* 1: convhlw */
+    var33 = ((orc_uint32)var32.i)>>16;
+    /* 2: storew */
+    ptr0[i] = var33;
+  }
+
+}
+
+void
 emulate_convssslw (OrcOpcodeExecutor *ex, int offset, int n)
 {
   int i;
diff --git a/orc/orcemulateopcodes.h b/orc/orcemulateopcodes.h
index 49d2e2d..354c8b2 100644
--- a/orc/orcemulateopcodes.h
+++ b/orc/orcemulateopcodes.h
@@ -48,6 +48,7 @@ void emulate_avguw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_cmpeqw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_cmpgtsw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_copyw (OrcOpcodeExecutor *ex, int i, int n);
+void emulate_div255w (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_loadw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_loadoffw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_loadpw (OrcOpcodeExecutor *ex, int i, int n);
@@ -102,18 +103,23 @@ void emulate_xorl (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_loadq (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_loadpq (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_storeq (OrcOpcodeExecutor *ex, int i, int n);
+void emulate_splatw0q (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convsbw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convubw (OrcOpcodeExecutor *ex, int i, int n);
+void emulate_splatbw (OrcOpcodeExecutor *ex, int i, int n);
+void emulate_splatbl (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convswl (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convuwl (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convslq (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convulq (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convwb (OrcOpcodeExecutor *ex, int i, int n);
+void emulate_convhwb (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convssswb (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convsuswb (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convusswb (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convuuswb (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convlw (OrcOpcodeExecutor *ex, int i, int n);
+void emulate_convhlw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convssslw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convsuslw (OrcOpcodeExecutor *ex, int i, int n);
 void emulate_convusslw (OrcOpcodeExecutor *ex, int i, int n);
diff --git a/orc/orcopcodes.c b/orc/orcopcodes.c
index b31df2c..18a25d3 100644
--- a/orc/orcopcodes.c
+++ b/orc/orcopcodes.c
@@ -331,6 +331,7 @@ static OrcStaticOpcode opcodes[] = {
   { "cmpeqw", 0, { 2 }, { 2, 2 }, emulate_cmpeqw },
   { "cmpgtsw", 0, { 2 }, { 2, 2 }, emulate_cmpgtsw },
   { "copyw", 0, { 2 }, { 2 }, emulate_copyw },
+  { "div255w", 0, { 2 }, { 2 }, emulate_div255w },
   { "loadw", ORC_STATIC_OPCODE_LOAD, { 2 }, { 2 }, emulate_loadw },
   { "loadoffw", ORC_STATIC_OPCODE_LOAD|ORC_STATIC_OPCODE_SCALAR, { 2 }, { 2, 4 }, emulate_loadoffw },
   { "loadpw", ORC_STATIC_OPCODE_LOAD|ORC_STATIC_OPCODE_SCALAR|ORC_STATIC_OPCODE_INVARIANT, { 2 }, { 2 }, emulate_loadpw },
@@ -388,21 +389,27 @@ static OrcStaticOpcode opcodes[] = {
   { "loadq", ORC_STATIC_OPCODE_LOAD, { 8 }, { 8 }, emulate_loadq },
   { "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 },
+  { "splatw0q", 0, { 8 }, { 8 }, emulate_splatw0q },
 
   { "convsbw", 0, { 2 }, { 1 }, emulate_convsbw },
   { "convubw", 0, { 2 }, { 1 }, emulate_convubw },
+  { "splatbw", 0, { 2 }, { 1 }, emulate_splatbw },
+  { "splatbl", 0, { 4 }, { 1 }, emulate_splatbl },
+
   { "convswl", 0, { 4 }, { 2 }, emulate_convswl },
   { "convuwl", 0, { 4 }, { 2 }, emulate_convuwl },
   { "convslq", 0, { 8 }, { 4 }, emulate_convslq },
   { "convulq", 0, { 8 }, { 4 }, emulate_convulq },
 
   { "convwb", 0, { 1 }, { 2 }, emulate_convwb },
+  { "convhwb", 0, { 1 }, { 2 }, emulate_convhwb },
   { "convssswb", 0, { 1 }, { 2 }, emulate_convssswb },
   { "convsuswb", 0, { 1 }, { 2 }, emulate_convsuswb },
   { "convusswb", 0, { 1 }, { 2 }, emulate_convusswb },
   { "convuuswb", 0, { 1 }, { 2 }, emulate_convuuswb },
 
   { "convlw", 0, { 2 }, { 4 }, emulate_convlw },
+  { "convhlw", 0, { 2 }, { 4 }, emulate_convhlw },
   { "convssslw", 0, { 2 }, { 4 }, emulate_convssslw },
   { "convsuslw", 0, { 2 }, { 4 }, emulate_convsuslw },
   { "convusslw", 0, { 2 }, { 4 }, emulate_convusslw },
diff --git a/orc/orcprogram-c.c b/orc/orcprogram-c.c
index 9fce6ac..09e550a 100644
--- a/orc/orcprogram-c.c
+++ b/orc/orcprogram-c.c
@@ -782,6 +782,56 @@ c_rule_splitwb (OrcCompiler *p, void *user, OrcInstruction *insn)
   ORC_ASM_CODE(p,"    %s = %s & 0xff;\n", dest2, src);
 }
 
+static void
+c_rule_splatbw (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  char dest[20], src[20];
+
+  c_get_name_int (dest, p, insn->dest_args[0]);
+  c_get_name_int (src, p, insn->src_args[0]);
+
+  ORC_ASM_CODE(p,"    %s = ((%s&0xff) << 8) | (%s&0xff);\n", dest, src, src);
+}
+
+static void
+c_rule_splatbl (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  char dest[20], src[20];
+
+  c_get_name_int (dest, p, insn->dest_args[0]);
+  c_get_name_int (src, p, insn->src_args[0]);
+
+  ORC_ASM_CODE(p,
+      "    %s = ((%s&0xff) << 24) | ((%s&0xff)<<16) | ((%s&0xff) << 8) | (%s&0xff);\n",
+      dest, src, src, src, src);
+}
+
+static void
+c_rule_splatw0q (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  char dest[20], src[20];
+
+  c_get_name_int (dest, p, insn->dest_args[0]);
+  c_get_name_int (src, p, insn->src_args[0]);
+
+  ORC_ASM_CODE(p,
+      "    %s = ((orc_uint64)(%s&0xffff) << 48) | ((orc_uint64)(%s&0xffff)<<32) | ((%s&0xffff) << 16) | (%s&0xffff);\n",
+      dest, src, src, src, src);
+}
+
+static void
+c_rule_div255w (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  char dest[20], src[20];
+
+  c_get_name_int (dest, p, insn->dest_args[0]);
+  c_get_name_int (src, p, insn->src_args[0]);
+
+  ORC_ASM_CODE(p,
+      "    %s = ((uint16_t)(((orc_uint16)(%s+128)) + (((orc_uint16)(%s+128))>>8)))>>8;\n",
+      dest, src, src);
+}
+
 static OrcTarget c_target = {
   "c",
   FALSE,
@@ -858,5 +908,9 @@ orc_c_init (void)
   orc_rule_register (rule_set, "accsadubl", c_rule_accsadubl, NULL);
   orc_rule_register (rule_set, "splitlw", c_rule_splitlw, NULL);
   orc_rule_register (rule_set, "splitwb", c_rule_splitwb, NULL);
+  orc_rule_register (rule_set, "splatbw", c_rule_splatbw, NULL);
+  orc_rule_register (rule_set, "splatbl", c_rule_splatbl, NULL);
+  orc_rule_register (rule_set, "splatw0q", c_rule_splatw0q, NULL);
+  orc_rule_register (rule_set, "div255w", c_rule_div255w, NULL);
 }
 
diff --git a/orc/orcrules-sse.c b/orc/orcrules-sse.c
index 0f514b1..8dfde43 100644
--- a/orc/orcrules-sse.c
+++ b/orc/orcrules-sse.c
@@ -802,6 +802,22 @@ sse_rule_convwb (OrcCompiler *p, void *user, OrcInstruction *insn)
 }
 
 static void
+sse_rule_convhwb (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  int src = p->vars[insn->src_args[0]].alloc;
+  int dest = p->vars[insn->dest_args[0]].alloc;
+
+  /* FIXME slow */
+
+  if (dest != src) {
+    orc_sse_emit_movdqa (p, src, dest);
+  }
+
+  orc_sse_emit_psrlw (p, 8, dest);
+  orc_sse_emit_packuswb (p, dest, dest);
+}
+
+static void
 sse_rule_convswl (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
   int src = p->vars[insn->src_args[0]].alloc;
@@ -849,6 +865,22 @@ sse_rule_convlw (OrcCompiler *p, void *user, OrcInstruction *insn)
 }
 
 static void
+sse_rule_convhlw (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  int src = p->vars[insn->src_args[0]].alloc;
+  int dest = p->vars[insn->dest_args[0]].alloc;
+
+  /* FIXME slow */
+
+  if (dest != src) {
+    orc_sse_emit_movdqa (p, src, dest);
+  }
+
+  orc_sse_emit_psrad (p, 16, dest);
+  orc_sse_emit_packssdw (p, dest, dest);
+}
+
+static void
 sse_rule_convssslw (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
   int src = p->vars[insn->src_args[0]].alloc;
@@ -907,6 +939,62 @@ sse_rule_convql (OrcCompiler *p, void *user, OrcInstruction *insn)
 }
 
 static void
+sse_rule_splatw0q (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  int src = p->vars[insn->src_args[0]].alloc;
+  int dest = p->vars[insn->dest_args[0]].alloc;
+
+  if (src != dest) {
+    orc_sse_emit_movdqa (p, src, dest);
+  }
+  orc_sse_emit_pshuflw (p, ORC_SSE_SHUF(0,0,0,0), dest, dest);
+  orc_sse_emit_pshufhw (p, ORC_SSE_SHUF(0,0,0,0), dest, dest);
+}
+
+static void
+sse_rule_splatbw (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  int src = p->vars[insn->src_args[0]].alloc;
+  int dest = p->vars[insn->dest_args[0]].alloc;
+
+  if (src != dest) {
+    orc_sse_emit_movdqa (p, src, dest);
+  }
+  orc_sse_emit_punpcklbw (p, dest, dest);
+}
+
+static void
+sse_rule_splatbl (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  int src = p->vars[insn->src_args[0]].alloc;
+  int dest = p->vars[insn->dest_args[0]].alloc;
+
+  if (src != dest) {
+    orc_sse_emit_movdqa (p, src, dest);
+  }
+  orc_sse_emit_punpcklbw (p, dest, dest);
+  orc_sse_emit_punpcklwd (p, dest, dest);
+}
+
+static void
+sse_rule_div255w (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;
+
+  if (src != dest) {
+    orc_sse_emit_movdqa (p, src, dest);
+  }
+  tmp = orc_compiler_get_constant (p, 2, 0x0080);
+  orc_sse_emit_paddw (p, tmp, dest);
+  orc_sse_emit_movdqa (p, dest, tmp);
+  orc_sse_emit_psrlw (p, 8, tmp);
+  orc_sse_emit_paddw (p, tmp, dest);
+  orc_sse_emit_psrlw (p, 8, dest);
+}
+
+static void
 sse_rule_mulsbw (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
   int src = p->vars[insn->src_args[1]].alloc;
@@ -2102,6 +2190,12 @@ orc_compiler_sse_register_rules (OrcTarget *target)
   orc_rule_register (rule_set, "subssl", sse_rule_subssl_slow, NULL);
   orc_rule_register (rule_set, "addusl", sse_rule_addusl_slow, NULL);
   orc_rule_register (rule_set, "subusl", sse_rule_subusl_slow, NULL);
+  orc_rule_register (rule_set, "convhwb", sse_rule_convhwb, NULL);
+  orc_rule_register (rule_set, "convhlw", sse_rule_convhlw, NULL);
+  orc_rule_register (rule_set, "splatw0q", sse_rule_splatw0q, NULL);
+  orc_rule_register (rule_set, "splatbw", sse_rule_splatbw, NULL);
+  orc_rule_register (rule_set, "splatbl", sse_rule_splatbl, NULL);
+  orc_rule_register (rule_set, "div255w", sse_rule_div255w, NULL);
 
   /* SSE 3 -- no rules */
 
diff --git a/orc/orcsse.c b/orc/orcsse.c
index b6edd27..89f0dfd 100644
--- a/orc/orcsse.c
+++ b/orc/orcsse.c
@@ -110,6 +110,20 @@ orc_sse_emit_pshufd (OrcCompiler *p, int shuf, int src, int dest)
 }
 
 void
+orc_sse_emit_pshufhw (OrcCompiler *p, int shuf, int src, int dest)
+{
+  ORC_ASM_CODE(p,"  pshufhw $0x%04x, %%%s, %%%s\n", shuf,
+      orc_x86_get_regname_sse(src),
+      orc_x86_get_regname_sse(dest));
+  *p->codeptr++ = 0xf3;
+  orc_x86_emit_rex (p, 0, dest, 0, src);
+  *p->codeptr++ = 0x0f;
+  *p->codeptr++ = 0x70;
+  orc_x86_emit_modrm_reg (p, src, dest);
+  *p->codeptr++ = shuf;
+}
+
+void
 orc_sse_emit_pshuflw (OrcCompiler *p, int shuf, int src, int dest)
 {
   ORC_ASM_CODE(p,"  pshuflw $0x%04x, %%%s, %%%s\n", shuf,
diff --git a/orc/orcsse.h b/orc/orcsse.h
index a32ab13..c7583bd 100644
--- a/orc/orcsse.h
+++ b/orc/orcsse.h
@@ -64,6 +64,7 @@ void orc_sse_emit_0f (OrcCompiler *p, const char *insn_name, int code,
     int src, int dest);
 void orc_sse_emit_pshufd (OrcCompiler *p, int shuf, int src, int dest);
 void orc_sse_emit_pshuflw (OrcCompiler *p, int shuf, int src, int dest);
+void orc_sse_emit_pshufhw (OrcCompiler *p, int shuf, int src, int dest);
 void orc_sse_emit_shiftimm (OrcCompiler *p, const char *insn_name,
     int code, int modrm_code, int shift, int reg);
 
-- 
2.7.4