Dynamically allocate temporary registers

diff --git a/orc/orccompiler.c b/orc/orccompiler.c
index a5ffef4..3fa1d04 100644 (file)
--- a/orc/orccompiler.c
+++ b/orc/orccompiler.c
@@ -551,6 +551,45 @@ orc_compiler_assign_rules (OrcCompiler *compiler)
   }
 }
 
+int
+orc_compiler_get_temp_reg (OrcCompiler *compiler)
+{
+  int j;
+
+  for(j=0;j<ORC_N_REGS;j++){
+    compiler->alloc_regs[j] = 0;
+  }
+  for(j=0;j<ORC_N_COMPILER_VARIABLES;j++){
+    if (!compiler->vars[j].alloc) continue;
+
+    ORC_DEBUG("var %d: %d  %d %d", j, compiler->vars[j].alloc,
+        compiler->vars[j].first_use,
+        compiler->vars[j].last_use);
+
+    if (compiler->vars[j].first_use == -1) {
+      compiler->alloc_regs[compiler->vars[j].alloc] = 1;
+    } else if (compiler->vars[j].first_use <= compiler->insn_index &&
+        compiler->vars[j].last_use >= compiler->insn_index) {
+      compiler->alloc_regs[compiler->vars[j].alloc] = 1;
+    }
+  }
+
+  ORC_DEBUG("at insn %d", compiler->insn_index);
+  for(j=0;j<8;j++){
+    ORC_DEBUG("xmm%d: %d %d", j, compiler->valid_regs[ORC_VEC_REG_BASE + j],
+        compiler->alloc_regs[ORC_VEC_REG_BASE + j]);
+  }
+
+  for(j=compiler->min_temp_reg;j<ORC_VEC_REG_BASE+16;j++){
+    if (compiler->valid_regs[j] && !compiler->alloc_regs[j]) {
+      compiler->min_temp_reg = j+1;
+      return j;
+    }
+  }
+
+  return 0;
+}
+
 void
 orc_compiler_rewrite_vars (OrcCompiler *compiler)
 {
@@ -877,6 +916,7 @@ int
 orc_compiler_get_constant (OrcCompiler *compiler, int size, int value)
 {
   int i;
+  int tmp;
 
   if (size < 4) {
     if (size < 2) {
@@ -904,7 +944,8 @@ orc_compiler_get_constant (OrcCompiler *compiler, int size, int value)
   if (compiler->constants[i].alloc_reg != 0) {;
     return compiler->constants[i].alloc_reg;
   }
-  orc_compiler_load_constant (compiler, compiler->tmpreg, size, value);
-  return compiler->tmpreg;
+  tmp = orc_compiler_get_temp_reg (compiler);
+  orc_compiler_load_constant (compiler, tmp, size, value);
+  return tmp;
 }
 

diff --git a/orc/orcprogram-sse.c b/orc/orcprogram-sse.c
index 9f8fe11..58b1e25 100644 (file)
--- a/orc/orcprogram-sse.c
+++ b/orc/orcprogram-sse.c
@@ -826,10 +826,14 @@ orc_sse_emit_loop (OrcCompiler *compiler, int offset, int update)
     insn = compiler->insns + j;
     opcode = insn->opcode;
 
+    compiler->insn_index = j;
+
     if (insn->flags & ORC_INSN_FLAG_INVARIANT) continue;
 
     ORC_ASM_CODE(compiler,"# %d: %s\n", j, insn->opcode->name);
 
+    compiler->min_temp_reg = ORC_VEC_REG_BASE;
+
     rule = insn->rule;
     if (rule && rule->emit) {
       if (!(insn->opcode->flags & (ORC_STATIC_OPCODE_ACCUMULATOR|ORC_STATIC_OPCODE_LOAD|ORC_STATIC_OPCODE_STORE)) &&

diff --git a/orc/orcprogram.h b/orc/orcprogram.h
index b3862ff..97ea55c 100644 (file)
--- a/orc/orcprogram.h
+++ b/orc/orcprogram.h
@@ -431,6 +431,7 @@ struct _OrcCompiler {
   int has_iterator_opcode;
 
   int offset;
+  int min_temp_reg;
 };
 
 #define ORC_SRC_ARG(p,i,n) ((p)->vars[(i)->src_args[(n)]].alloc)
@@ -647,6 +648,7 @@ int orc_program_allocate_register (OrcProgram *program, int is_data);
 void orc_compiler_allocate_codemem (OrcCompiler *compiler);
 int orc_compiler_label_new (OrcCompiler *compiler);
 int orc_compiler_get_constant (OrcCompiler *compiler, int size, int value);
+int orc_compiler_get_temp_reg (OrcCompiler *compiler);
 
 const char *orc_program_get_asm_code (OrcProgram *program);
 const char *orc_target_get_asm_preamble (const char *target);

diff --git a/orc/orcrules-sse.c b/orc/orcrules-sse.c
index 4c6bbc5..f644e69 100644 (file)
--- a/orc/orcrules-sse.c
+++ b/orc/orcrules-sse.c
@@ -482,7 +482,7 @@ sse_rule_accsadubl (OrcCompiler *p, void *user, OrcInstruction *insn)
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp = p->tmpreg;
 #ifndef MMX
-  int tmp2 = X86_XMM7;
+  int tmp2 = orc_compiler_get_temp_reg (p);
 
   if (p->loop_shift <= 2) {
     orc_sse_emit_movdqa (p, src1, tmp);
@@ -515,12 +515,11 @@ sse_rule_signX_ssse3 (OrcCompiler *p, void *user, OrcInstruction *insn)
   int type = ORC_PTR_TO_INT(user);
   int tmpc;
 
+  tmpc = orc_compiler_get_constant (p, 1<<type, 1);
   if (src == dest) {
-    tmpc = orc_compiler_get_constant (p, 1<<type, 1);
     orc_sse_emit_660f (p, names[type], codes[type], src, tmpc);
     orc_sse_emit_movdqa (p, tmpc, dest);
   } else {
-    tmpc = orc_compiler_get_constant (p, 1<<type, 1);
     orc_sse_emit_movdqa (p, tmpc, dest);
     orc_sse_emit_660f (p, names[type], codes[type], src, dest);
   }
@@ -975,11 +974,11 @@ sse_rule_divluw (OrcCompiler *p, void *user, OrcInstruction *insn)
   /* About 5.6 cycles per array member on ginger */
   int src = p->vars[insn->src_args[1]].alloc;
   int dest = p->vars[insn->dest_args[0]].alloc;
-  int a = X86_XMM7;
-  int j = X86_XMM6;
-  int j2 = X86_XMM5;
-  int l = X86_XMM4;
-  int divisor = X86_XMM3;
+  int a = orc_compiler_get_temp_reg (p);
+  int j = orc_compiler_get_temp_reg (p);
+  int j2 = orc_compiler_get_temp_reg (p);
+  int l = orc_compiler_get_temp_reg (p);
+  int divisor = orc_compiler_get_temp_reg (p);
   int tmp;
   int i;
 
@@ -1025,10 +1024,10 @@ sse_rule_divluw (OrcCompiler *p, void *user, OrcInstruction *insn)
   /* About 8.4 cycles per array member on ginger */
   int src = p->vars[insn->src_args[1]].alloc;
   int dest = p->vars[insn->dest_args[0]].alloc;
-  int b = X86_XMM7;
-  int a = X86_XMM6;
-  int k = X86_XMM5;
-  int j = X86_XMM4;
+  int b = orc_compiler_get_temp_reg (p);
+  int a = orc_compiler_get_temp_reg (p);
+  int k = orc_compiler_get_temp_reg (p);
+  int j = orc_compiler_get_temp_reg (p);
   int tmp;
   int i;
 
@@ -1146,7 +1145,7 @@ sse_rule_mullb (OrcCompiler *p, void *user, OrcInstruction *insn)
   int src = p->vars[insn->src_args[1]].alloc;
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp = p->tmpreg;
-  int tmp2 = X86_XMM7;
+  int tmp2 = orc_compiler_get_temp_reg (p);
 
   orc_sse_emit_movdqa (p, dest, tmp);
 
@@ -1169,7 +1168,7 @@ sse_rule_mulhsb (OrcCompiler *p, void *user, OrcInstruction *insn)
   int src = p->vars[insn->src_args[1]].alloc;
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp = p->tmpreg;
-  int tmp2 = X86_XMM7;
+  int tmp2 = orc_compiler_get_temp_reg (p);
 
   orc_sse_emit_movdqa (p, src, tmp);
   orc_sse_emit_movdqa (p, dest, tmp2);
@@ -1197,7 +1196,7 @@ sse_rule_mulhub (OrcCompiler *p, void *user, OrcInstruction *insn)
   int src = p->vars[insn->src_args[1]].alloc;
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp = p->tmpreg;
-  int tmp2 = X86_XMM7;
+  int tmp2 = orc_compiler_get_temp_reg (p);
 
   orc_sse_emit_movdqa (p, src, tmp);
   orc_sse_emit_movdqa (p, dest, tmp2);
@@ -1512,7 +1511,7 @@ sse_emit_load_mask (OrcCompiler *p, unsigned int mask1, unsigned int mask2)
 {
   int tmp = p->tmpreg;
   int gptmp = p->gp_tmpreg;
-  int tmp2 = X86_XMM7;
+  int tmp2 = orc_compiler_get_temp_reg (p);
 
   orc_x86_emit_mov_imm_reg (p, 4, mask1, gptmp);
   orc_x86_emit_mov_reg_sse (p, gptmp, tmp);
@@ -1835,8 +1834,8 @@ sse_rule_addssl_slow (OrcCompiler *p, void *user, OrcInstruction *insn)
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp = p->tmpreg;
 #if 0
-  int tmp2 = X86_XMM7;
-  int tmp3 = X86_XMM6;
+  int tmp2 = orc_compiler_get_temp_reg (p);
+  int tmp3 = orc_compiler_get_temp_reg (p);
 
   orc_sse_emit_movdqa (p, src, tmp);
   orc_sse_emit_pand (p, dest, tmp);
@@ -1866,8 +1865,8 @@ sse_rule_addssl_slow (OrcCompiler *p, void *user, OrcInstruction *insn)
   orc_sse_emit_por (p, tmp2, dest);
 #endif
 
-  int s = X86_XMM7;
-  int t = X86_XMM6;
+  int s = orc_compiler_get_temp_reg (p);
+  int t = orc_compiler_get_temp_reg (p);
 
   /*
      From Tim Terriberry: (slightly faster than above)
@@ -1914,8 +1913,8 @@ sse_rule_subssl_slow (OrcCompiler *p, void *user, OrcInstruction *insn)
   int src = p->vars[insn->src_args[1]].alloc;
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp = p->tmpreg;
-  int tmp2 = X86_XMM7;
-  int tmp3 = X86_XMM6;
+  int tmp2 = orc_compiler_get_temp_reg (p);
+  int tmp3 = orc_compiler_get_temp_reg (p);
 
   tmp = orc_compiler_get_constant (p, 4, 0xffffffff);
   orc_sse_emit_pxor (p, src, tmp);
@@ -1953,7 +1952,7 @@ sse_rule_addusl_slow (OrcCompiler *p, void *user, OrcInstruction *insn)
   int src = p->vars[insn->src_args[1]].alloc;
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp = p->tmpreg;
-  int tmp2 = X86_XMM7;
+  int tmp2 = orc_compiler_get_temp_reg (p);
 
 #if 0
   /* an alternate version.  slower. */
@@ -2000,7 +1999,7 @@ sse_rule_subusl_slow (OrcCompiler *p, void *user, OrcInstruction *insn)
   int src = p->vars[insn->src_args[1]].alloc;
   int dest = p->vars[insn->dest_args[0]].alloc;
   int tmp = p->tmpreg;
-  int tmp2 = X86_XMM7;
+  int tmp2 = orc_compiler_get_temp_reg (p);
 
   orc_sse_emit_movdqa (p, src, tmp2);
   orc_sse_emit_psrld (p, 1, tmp2);