user/orc: bump to 0.4.30, migrate to meson

1 files changed, 990 insertions, 0 deletions

diff --git a/user/orc/ppc-fixes.patch b/user/orc/ppc-fixes.patch
new file mode 100644
index 000000000..ab4aecd40
--- /dev/null
+++ b/user/orc/ppc-fixes.patch
@@ -0,0 +1,990 @@
+diff --git a/orc-test/orctest.c b/orc-test/orctest.c
+index 614421b..d9cb9d5 100644
+--- a/orc-test/orctest.c
++++ b/orc-test/orctest.c
+@@ -523,14 +523,14 @@ float_compare (OrcArray *array1, OrcArray *array2, int i, int j)
+       if (isnan(*(float *)ptr1) && isnan(*(float *)ptr2)) return TRUE;
+       if (*(float *)ptr1 == *(float *)ptr2) return TRUE;
+       if ((*(float *)ptr1 < 0.0) == (*(float *)ptr2 < 0.0) &&
+-          abs((orc_int32)(*(orc_uint32 *)ptr1 - *(orc_uint32 *)ptr2) <= 2))
++          abs((orc_int32)(*(orc_uint32 *)ptr1 - *(orc_uint32 *)ptr2)) <= 2)
+         return TRUE;
+       return FALSE;
+     case 8:
+       if (isnan(*(double *)ptr1) && isnan(*(double *)ptr2)) return TRUE;
+       if (*(double *)ptr1 == *(double *)ptr2) return TRUE;
+       if ((*(double *)ptr1 < 0.0) == (*(double *)ptr2 < 0.0) &&
+-          llabs((orc_int64)(*(orc_uint64 *)ptr1 - *(orc_uint64 *)ptr2) <= 2))
++          llabs((orc_int64)(*(orc_uint64 *)ptr1 - *(orc_uint64 *)ptr2)) <= 2)
+         return TRUE;
+       return FALSE;
+   }
+diff --git a/orc/orcpowerpc.c b/orc/orcpowerpc.c
+index e054d07..ed98437 100644
+--- a/orc/orcpowerpc.c
++++ b/orc/orcpowerpc.c
+@@ -208,7 +208,7 @@ powerpc_emit_srawi (OrcCompiler *compiler, int regd, int rega, int shift,
+       powerpc_get_regname(regd),
+       powerpc_get_regname(rega), shift);
+ 
+-  insn = (31<<26) | (powerpc_regnum (regd)<<21) | (powerpc_regnum (rega)<<16);
++  insn = (31<<26) | (powerpc_regnum (rega)<<21) | (powerpc_regnum (regd)<<16);
+   insn |= (shift<<11) | (824<<1) | record;
+ 
+   powerpc_emit (compiler, insn);
+@@ -233,7 +233,7 @@ powerpc_emit_D (OrcCompiler *compiler, const char *name,
+   ORC_ASM_CODE(compiler,"  %s %s, %s, %d\n", name,
+       powerpc_get_regname(regd),
+       powerpc_get_regname(rega), imm);
+-  insn |= (powerpc_regnum (regd)<<21) | (powerpc_regnum (rega)<<16);
++  insn |= (powerpc_regnum (rega)<<21) | (powerpc_regnum (regd)<<16);
+   insn |= imm&0xffff;
+ 
+   powerpc_emit (compiler, insn);
+@@ -543,7 +543,7 @@ powerpc_load_constant (OrcCompiler *p, int i, int reg)
+   {
+     unsigned int insn;
+ 
+-    ORC_ASM_CODE(p,"  addi %s, %s, %db - %s\n",
++    ORC_ASM_CODE(p,"  addi %s, %s, %df - %s\n",
+         powerpc_get_regname(greg),
+         powerpc_get_regname(greg), p->constants[i].label, p->program->name);
+     insn = (14<<26) | (powerpc_regnum (greg)<<21) | (powerpc_regnum (greg)<<16);
+@@ -740,3 +740,42 @@ void powerpc_emit_label (OrcCompiler *compiler, int label)
+   powerpc_add_label (compiler, compiler->codeptr, label);
+ }
+ 
++void
++powerpc_load_align (OrcCompiler *compiler, int vector_reg, int offset_reg, int src_reg)
++{
++  if (IS_POWERPC_BE (compiler)) {
++    ORC_ASM_CODE(compiler,"  lvsl %s, %s, %s\n",
++        powerpc_get_regname (vector_reg),
++        offset_reg == 0 ? "0" : powerpc_get_regname (offset_reg),
++        powerpc_get_regname (src_reg));
++    powerpc_emit_X (compiler, 0x7c00000c, powerpc_regnum(vector_reg),
++        offset_reg == 0 ? 0 : powerpc_regnum(offset_reg), powerpc_regnum(src_reg));
++  } else {
++    ORC_ASM_CODE(compiler,"  lvsr %s, %s, %s\n",
++        powerpc_get_regname (vector_reg),
++        offset_reg == 0 ? "0" : powerpc_get_regname (offset_reg),
++        powerpc_get_regname (src_reg));
++    powerpc_emit_X (compiler, 0x7c00004c, powerpc_regnum(vector_reg),
++        offset_reg == 0 ? 0 : powerpc_regnum(offset_reg), powerpc_regnum(src_reg));
++  }
++}
++
++void
++powerpc_store_align (OrcCompiler *compiler, int vector_reg, int offset_reg, int src_reg)
++{
++  if (IS_POWERPC_BE (compiler)) {
++    ORC_ASM_CODE(compiler,"  lvsr %s, %s, %s\n",
++        powerpc_get_regname (vector_reg),
++        offset_reg == 0 ? "0" : powerpc_get_regname (offset_reg),
++        powerpc_get_regname (src_reg));
++    powerpc_emit_X (compiler, 0x7c00004c, powerpc_regnum(vector_reg),
++        offset_reg == 0 ? 0 : powerpc_regnum(offset_reg), powerpc_regnum(src_reg));
++  } else {
++    ORC_ASM_CODE(compiler,"  lvsl %s, %s, %s\n",
++        powerpc_get_regname (vector_reg),
++        offset_reg == 0 ? "0" : powerpc_get_regname (offset_reg),
++        powerpc_get_regname (src_reg));
++    powerpc_emit_X (compiler, 0x7c00000c, powerpc_regnum(vector_reg),
++        offset_reg == 0 ? 0 : powerpc_regnum(offset_reg), powerpc_regnum(src_reg));
++  }
++}
+diff --git a/orc/orcpowerpc.h b/orc/orcpowerpc.h
+index e6b315a..ba824b2 100644
+--- a/orc/orcpowerpc.h
++++ b/orc/orcpowerpc.h
+@@ -138,6 +138,9 @@ void powerpc_emit_full_constants (OrcCompiler* p);
+ void powerpc_emit_load_address (OrcCompiler* compiler, int regd, int rega, int imm);
+ void powerpc_load_constant (OrcCompiler* p, int i, int reg);
+ 
++void powerpc_load_align (OrcCompiler *compiler, int vector_reg, int offset_reg, int src_reg);
++void powerpc_store_align (OrcCompiler *compiler, int vector_reg, int offset_reg, int src_reg);
++
+ /* instructions */
+ #define powerpc_emit_vand(p,a,b,c)         powerpc_emit_VX_2 (p, "vand", 0x10000404, a, b, c)
+ #define powerpc_emit_vandc(p,a,b,c)        powerpc_emit_VX_2 (p, "vandc", 0x10000444, a, b, c)
+@@ -159,8 +162,10 @@ void powerpc_load_constant (OrcCompiler* p, int i, int reg);
+ 
+ #define powerpc_emit_vmrghb(p,a,b,c)       powerpc_emit_VX_2 (p, "vmrghb", 0x1000000c, a, b, c)
+ #define powerpc_emit_vmrghh(p,a,b,c)       powerpc_emit_VX_2 (p, "vmrghh", 0x1000004c, a, b, c)
++#define powerpc_emit_vmrghw(p,a,b,c)       powerpc_emit_VX_2 (p, "vmrghw", 0x1000008c, a, b, c)
+ #define powerpc_emit_vmrglb(p,a,b,c)       powerpc_emit_VX_2 (p, "vmrglb", 0x1000010c, a, b, c)
+ #define powerpc_emit_vmrglh(p,a,b,c)       powerpc_emit_VX_2 (p, "vmrglh", 0x1000014c, a, b, c)
++#define powerpc_emit_vmrglw(p,a,b,c)       powerpc_emit_VX_2 (p, "vmrglw", 0x1000018c, a, b, c)
+ 
+ #define powerpc_emit_vpkshss(p,a,b,c)      powerpc_emit_VX_2 (p, "vpkshss", 0x1000018e, a, b, c)
+ #define powerpc_emit_vpkshus(p,a,b,c)      powerpc_emit_VX_2 (p, "vpkshus", 0x1000010e, a, b, c)
+diff --git a/orc/orcprogram-altivec.c b/orc/orcprogram-altivec.c
+index 659115e..808b4e1 100644
+--- a/orc/orcprogram-altivec.c
++++ b/orc/orcprogram-altivec.c
+@@ -18,6 +18,8 @@ static void orc_compiler_powerpc_init (OrcCompiler *compiler);
+ static unsigned int orc_compiler_powerpc_get_default_flags (void);
+ static void orc_compiler_powerpc_assemble (OrcCompiler *compiler);
+ static const char* powerpc_get_flag_name (int shift);
++static int orc_powerpc_assemble_copy_check (OrcCompiler *compiler);
++static void orc_powerpc_assemble_copy (OrcCompiler *compiler);
+ 
+ 
+ static void
+@@ -199,8 +201,8 @@ powerpc_load_constants_outer (OrcCompiler *compiler)
+     }
+   }
+ 
+-  orc_compiler_emit_invariants (compiler);
+-
++  /* Load constants first, as they may be used by invariants */
++  ORC_ASM_CODE(compiler,"# load constants\n");
+   for(i=0;i<compiler->n_constants;i++) {
+     if (compiler->constants[i].is_long &&
+         !compiler->constants[i].alloc_reg) {
+@@ -212,6 +214,9 @@ powerpc_load_constants_outer (OrcCompiler *compiler)
+       }
+     }
+   }
++
++  ORC_ASM_CODE(compiler,"# load invariants\n");
++  orc_compiler_emit_invariants (compiler);
+ }
+ 
+ static void
+@@ -219,6 +224,7 @@ powerpc_load_inner_constants (OrcCompiler *compiler)
+ {
+   int i;
+ 
++  ORC_ASM_CODE(compiler,"# load inner constants\n");
+   for(i=0;i<ORC_N_COMPILER_VARIABLES;i++){
+     if (compiler->vars[i].name == NULL) continue;
+     switch (compiler->vars[i].vartype) {
+@@ -305,6 +311,11 @@ orc_compiler_powerpc_assemble (OrcCompiler *compiler)
+   int label_leave;
+   int set_vscr = FALSE;
+ 
++  if (orc_powerpc_assemble_copy_check (compiler)) {
++    orc_powerpc_assemble_copy (compiler);
++    return;
++  }
++
+   label_outer_loop_start = orc_compiler_label_new (compiler);
+   label_loop_start = orc_compiler_label_new (compiler);
+   label_leave = orc_compiler_label_new (compiler);
+@@ -312,6 +323,8 @@ orc_compiler_powerpc_assemble (OrcCompiler *compiler)
+   {
+     int i;
+ 
++    /* Emit invariants also to check for constants */
++    orc_compiler_emit_invariants (compiler);
+     orc_powerpc_emit_loop (compiler, 0);
+ 
+     compiler->codeptr = compiler->code;
+@@ -486,3 +499,199 @@ orc_compiler_powerpc_assemble (OrcCompiler *compiler)
+   powerpc_do_fixups (compiler);
+ }
+ 
++static
++int orc_powerpc_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 ||
++      strcmp (compiler->program->insns[0].opcode->name, "copyq") == 0) &&
++      (compiler->program->insns[0].flags &
++          (ORC_INSTRUCTION_FLAG_X2 | ORC_INSTRUCTION_FLAG_X4)) == 0 &&
++      compiler->program->n_param_vars == 0 &&
++      compiler->program->n_const_vars == 0) {
++    /* TODO: add param & const support if this turns out to be faster */
++    return TRUE;
++  }
++
++  return FALSE;
++}
++
++static
++void orc_powerpc_assemply_copy_loop (OrcCompiler *compiler, int size,
++    int shift, int next_label)
++{
++  const int src = POWERPC_R5;
++  const int dst = POWERPC_R6;
++  const int count = POWERPC_R7;
++  const int tmp = POWERPC_R0;
++  const int vtmp = POWERPC_V0;
++  const int vperm = POWERPC_V1;
++  int label_copy;
++
++  label_copy = orc_compiler_label_new (compiler);
++
++  ORC_ASM_CODE(compiler,"  cmplwi %s, %d\n",
++      powerpc_get_regname(count), size);
++  powerpc_emit(compiler, 0x28000000|(powerpc_regnum(count)<<16)|(size&0xffff));
++
++  ORC_ASM_CODE(compiler,"  blt %d%c\n", next_label,
++      (compiler->labels[next_label]!=NULL) ? 'b' : 'f');
++  powerpc_add_fixup (compiler, 0, compiler->codeptr, next_label);
++  powerpc_emit (compiler, 0x41800000);
++
++  powerpc_emit_D(compiler, "andi.", 0x70000000, tmp, src, size-1);
++  ORC_ASM_CODE(compiler,"  bgt %d%c\n", next_label,
++      (compiler->labels[next_label]!=NULL) ? 'b' : 'f');
++  powerpc_add_fixup (compiler, 0, compiler->codeptr, next_label);
++  powerpc_emit (compiler, 0x41810000);
++
++  powerpc_emit_D(compiler, "andi.", 0x70000000, tmp, dst, size-1);
++  ORC_ASM_CODE(compiler,"  bgt %d%c\n", next_label,
++      (compiler->labels[next_label]!=NULL) ? 'b' : 'f');
++  powerpc_add_fixup (compiler, 0, compiler->codeptr, next_label);
++  powerpc_emit (compiler, 0x41810000);
++
++  powerpc_emit_srawi (compiler, tmp, count, shift, 0);
++
++  ORC_ASM_CODE (compiler, "  mtctr %s\n", powerpc_get_regname(tmp));
++  powerpc_emit (compiler, 0x7c0903a6 | (powerpc_regnum (tmp)<<21));
++
++  powerpc_emit_label (compiler, label_copy);
++  if (size == 16) {
++    ORC_ASM_CODE(compiler,"  lvx %s, 0, %s\n",
++        powerpc_get_regname (vtmp),
++        powerpc_get_regname (src));
++    powerpc_emit_X (compiler, 0x7c0000ce, powerpc_regnum(vtmp),
++        0, powerpc_regnum(src));
++    ORC_ASM_CODE(compiler,"  stvx %s, 0, %s\n",
++        powerpc_get_regname (vtmp),
++        powerpc_get_regname (dst));
++    powerpc_emit_X (compiler, 0x7c0001ce,
++        powerpc_regnum(vtmp),
++        0, powerpc_regnum(dst));
++  } else {
++    switch (size) {
++      case 1:
++        ORC_ASM_CODE(compiler,"  lvebx %s, 0, %s\n",
++            powerpc_get_regname (vtmp),
++            powerpc_get_regname (src));
++        powerpc_emit_X (compiler, 0x7c00000e, powerpc_regnum(vtmp),
++            0, powerpc_regnum(src));
++        break;
++      case 2:
++        ORC_ASM_CODE(compiler,"  lvehx %s, 0, %s\n",
++            powerpc_get_regname (vtmp),
++            powerpc_get_regname (src));
++        powerpc_emit_X (compiler, 0x7c00004e, powerpc_regnum(vtmp),
++            0, powerpc_regnum(src));
++        break;
++      case 4:
++        ORC_ASM_CODE(compiler,"  lvewx %s, 0, %s\n",
++            powerpc_get_regname (vtmp),
++            powerpc_get_regname (src));
++        powerpc_emit_X (compiler, 0x7c00008e, powerpc_regnum(vtmp),
++            0, powerpc_regnum(src));
++        break;
++    }
++    powerpc_load_align (compiler, vperm, 0, src);
++    powerpc_emit_vperm (compiler, vtmp, vtmp, vtmp, vperm);
++    powerpc_store_align (compiler, vperm, 0, dst);
++    powerpc_emit_vperm (compiler, vtmp, vtmp, vtmp, vperm);
++    switch (size) {
++      case 1:
++        ORC_ASM_CODE(compiler,"  stvebx %s, 0, %s\n",
++            powerpc_get_regname (vtmp),
++            powerpc_get_regname (dst));
++        powerpc_emit_X (compiler, 0x7c00010e,
++            powerpc_regnum(vtmp),
++            0, powerpc_regnum(dst));
++        break;
++      case 2:
++        ORC_ASM_CODE(compiler,"  stvehx %s, 0, %s\n",
++            powerpc_get_regname (vtmp),
++            powerpc_get_regname (dst));
++        powerpc_emit_X (compiler, 0x7c00014e,
++            powerpc_regnum(vtmp),
++            0, powerpc_regnum(dst));
++        break;
++      case 4:
++	ORC_ASM_CODE(compiler,"  stvewx %s, 0, %s\n",
++            powerpc_get_regname (vtmp),
++            powerpc_get_regname (dst));
++        powerpc_emit_X (compiler, 0x7c00018e,
++            powerpc_regnum(vtmp),
++            0, powerpc_regnum(dst));
++        break;
++    }
++  }
++
++  powerpc_emit_addi (compiler, src, src, size);
++  powerpc_emit_addi (compiler, dst, dst, size);
++  powerpc_emit_addi (compiler, count, count, -size);
++  powerpc_emit_bne (compiler, label_copy);
++
++  powerpc_emit_label (compiler, next_label);
++}
++
++static
++void orc_powerpc_assemble_copy (OrcCompiler *compiler)
++{
++  const int src = POWERPC_R5;
++  const int dst = POWERPC_R6;
++  const int count = POWERPC_V7;
++  OrcInstruction *insn;
++  int shift = 0;
++  int label_word;
++  int label_halfword;
++  int label_byte;
++  int label_done;
++
++  insn = compiler->program->insns + 0;
++
++  if (strcmp (insn->opcode->name, "copyw") == 0) {
++    shift = 1;
++  } else if (strcmp (insn->opcode->name, "copyl") == 0) {
++    shift = 2;
++  } else if (strcmp (insn->opcode->name, "copyq") == 0) {
++    shift = 3;
++  }
++
++  label_word = orc_compiler_label_new (compiler);
++  label_halfword = orc_compiler_label_new (compiler);
++  label_byte = orc_compiler_label_new (compiler);
++  label_done = orc_compiler_label_new (compiler);
++
++  powerpc_emit_prologue (compiler);
++
++  powerpc_emit_load_address (compiler,
++      dst,
++      POWERPC_R3,
++      (int)ORC_STRUCT_OFFSET(OrcExecutor, arrays[insn->dest_args[0]]));
++  powerpc_emit_load_address (compiler,
++      src,
++      POWERPC_R3,
++      (int)ORC_STRUCT_OFFSET(OrcExecutor, arrays[insn->src_args[0]]));
++  powerpc_emit_lwz (compiler, count, POWERPC_R3,
++      (int)ORC_STRUCT_OFFSET(OrcExecutor, n));
++
++  powerpc_emit_addi (compiler, POWERPC_R0, 0, shift);
++  ORC_ASM_CODE(compiler, "  slw %s, %s, %s\n",
++      powerpc_get_regname(count),
++      powerpc_get_regname(count),
++      powerpc_get_regname(POWERPC_R0));
++  powerpc_emit (compiler, (31<<26) |
++      (powerpc_regnum (count)<<21) |
++      (powerpc_regnum (count)<<16) |
++      (powerpc_regnum (POWERPC_R0)<<11) | (24<<1));
++
++  orc_powerpc_assemply_copy_loop (compiler, 16, 4, label_word);
++  orc_powerpc_assemply_copy_loop (compiler, 4, 2, label_halfword);
++  orc_powerpc_assemply_copy_loop (compiler, 2, 1, label_byte);
++  orc_powerpc_assemply_copy_loop (compiler, 1, 0, label_done);
++
++  powerpc_emit_epilogue (compiler);
++  powerpc_do_fixups (compiler);
++}
+diff --git a/orc/orcrules-altivec.c b/orc/orcrules-altivec.c
+index fa40ac3..7729b69 100644
+--- a/orc/orcrules-altivec.c
++++ b/orc/orcrules-altivec.c
+@@ -15,26 +15,6 @@
+ 
+ /* rules */
+ 
+-static void
+-powerpc_load_align (OrcCompiler *compiler, int vector_reg, int offset_reg, int src_reg)
+-{
+-  if (IS_POWERPC_BE (compiler)) {
+-    ORC_ASM_CODE(compiler,"  lvsl %s, %s, %s\n",
+-      powerpc_get_regname (vector_reg),
+-      offset_reg == 0 ? "0" : powerpc_get_regname (offset_reg),
+-      powerpc_get_regname (src_reg));
+-    powerpc_emit_X (compiler, 0x7c00000c, powerpc_regnum(vector_reg),
+-      offset_reg == 0 ? 0 : powerpc_regnum(offset_reg), powerpc_regnum(src_reg));
+-  } else {
+-    ORC_ASM_CODE(compiler,"  lvsr %s, %s, %s\n",
+-      powerpc_get_regname (vector_reg),
+-      offset_reg == 0 ? "0" : powerpc_get_regname (offset_reg),
+-      powerpc_get_regname (src_reg));
+-    powerpc_emit_X (compiler, 0x7c00004c, powerpc_regnum(vector_reg),
+-      offset_reg == 0 ? 0 : powerpc_regnum(offset_reg), powerpc_regnum(src_reg));
+-  }
+-}
+-
+ static inline int
+ powerpc_select_value (OrcCompiler *compiler, int be_value, int le_value)
+ {
+@@ -322,19 +302,7 @@ powerpc_rule_storeX (OrcCompiler *compiler, void *user, OrcInstruction *insn)
+   int perm = orc_compiler_get_temp_reg (compiler);
+   int tmp = orc_compiler_get_temp_reg (compiler);
+ 
+-  if (IS_POWERPC_BE (compiler)) {
+-    ORC_ASM_CODE(compiler,"  lvsr %s, 0, %s\n",
+-        powerpc_get_regname (perm),
+-        powerpc_get_regname (dest->ptr_register));
+-    powerpc_emit_X (compiler, 0x7c00004c, powerpc_regnum(perm),
+-        0, powerpc_regnum(dest->ptr_register));
+-  } else {
+-    ORC_ASM_CODE(compiler,"  lvsl %s, 0, %s\n",
+-        powerpc_get_regname (perm),
+-        powerpc_get_regname (dest->ptr_register));
+-    powerpc_emit_X (compiler, 0x7c00000c, powerpc_regnum(perm),
+-        0, powerpc_regnum(dest->ptr_register));
+-  }
++  powerpc_store_align (compiler, perm, 0, dest->ptr_register);
+   powerpc_emit_vperm (compiler, tmp, src->alloc, src->alloc, perm);
+ 
+   switch (size) {
+@@ -370,8 +338,7 @@ powerpc_rule_storeX (OrcCompiler *compiler, void *user, OrcInstruction *insn)
+           powerpc_regnum(tmp),
+           0, powerpc_regnum(dest->ptr_register));
+ 
+-      powerpc_emit_D (compiler, "addi", 0x38000000, compiler->gp_tmpreg,
+-          0, 4);
++      powerpc_emit_addi (compiler, compiler->gp_tmpreg, 0, 4);
+ 
+       ORC_ASM_CODE(compiler,"  stvewx %s, %s, %s\n",
+           powerpc_get_regname (tmp),
+@@ -397,6 +364,48 @@ powerpc_rule_storeX (OrcCompiler *compiler, void *user, OrcInstruction *insn)
+   }
+ }
+ 
++static void
++powerpc_denormalize_sp_full(OrcCompiler *p, int reg, int zero, int mask)
++{
++  int tmp = p->tmpreg;
++  powerpc_emit_VX_2(p, "xxland", 0xf0000417, tmp, reg, mask);
++  powerpc_emit_VX_2(p, "vcmpequw", 0x10000086, tmp, tmp, zero);
++  powerpc_emit_VA(p, "xxsel", 0xf000003f, tmp, reg, zero, tmp);
++  powerpc_emit_VX_2(p, "xvcpsgnsp", 0xf0000687, reg, reg, tmp);
++}
++
++static void
++powerpc_denormalize_sp_zero(OrcCompiler *p, int reg, int zero)
++{
++  int mask = powerpc_get_constant_full (p, 0x7f800000, 0x7f800000,
++      0x7f800000, 0x7f800000);
++  powerpc_denormalize_sp_full(p, reg, zero, mask);
++}
++
++static void
++powerpc_denormalize_sp(OrcCompiler *p, int reg)
++{
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0);
++  powerpc_denormalize_sp_zero(p, reg, zero);
++}
++
++static void
++powerpc_denormalize_dp_full(OrcCompiler *p, int reg, int zero, int mask)
++{
++  int tmp = p->tmpreg;
++  powerpc_emit_VX_2(p, "xxland", 0xf0000417, tmp, reg, mask);
++  powerpc_emit_VX_2(p, "vcmpequd", 0x100000c7, tmp, tmp, zero);
++  powerpc_emit_VA(p, "xxsel", 0xf000003f, tmp, reg, zero, tmp);
++  powerpc_emit_VX_2(p, "xvcpsgndp", 0xf0000787, reg, reg, tmp);
++}
++
++static void
++powerpc_denormalize_dp_zero(OrcCompiler *p, int reg, int zero)
++{
++  int mask = powerpc_get_constant_full (p, 0x7ff00000, 0x00000000,
++      0x7ff00000, 0x00000000);
++  powerpc_denormalize_dp_full(p, reg, zero, mask);
++}
+ 
+ 
+ #define RULE(name, opcode, code) \
+@@ -409,6 +418,33 @@ powerpc_rule_ ## name (OrcCompiler *p, void *user, OrcInstruction *insn) \
+   powerpc_emit_VX_2 (p, opcode, code , dest, src1, src2);\
+ }
+ 
++#define RULE_DP(name, opcode, code) \
++static void \
++powerpc_rule_ ## name (OrcCompiler *p, void *user, OrcInstruction *insn) \
++{ \
++  int src1 = ORC_SRC_ARG (p, insn, 0); \
++  int src2 = ORC_SRC_ARG (p, insn, 1); \
++  int dest = ORC_DEST_ARG (p, insn, 0); \
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0); \
++  powerpc_denormalize_dp_zero (p, src1, zero); \
++  powerpc_denormalize_dp_zero (p, src2, zero); \
++  powerpc_emit_VX_2 (p, opcode, code , dest, src1, src2); \
++  powerpc_denormalize_dp_zero (p, dest, zero); \
++}
++
++#define RULE_DP_SRC(name, opcode, code) \
++static void \
++powerpc_rule_ ## name (OrcCompiler *p, void *user, OrcInstruction *insn) \
++{ \
++  int src1 = ORC_SRC_ARG (p, insn, 0); \
++  int src2 = ORC_SRC_ARG (p, insn, 1); \
++  int dest = ORC_DEST_ARG (p, insn, 0); \
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0); \
++  powerpc_denormalize_dp_zero (p, src1, zero); \
++  powerpc_denormalize_dp_zero (p, src2, zero); \
++  powerpc_emit_VX_2 (p, opcode, code , dest, src1, src2); \
++}
++
+ #define RULE_SHIFT(name, opcode, code) \
+ static void \
+ powerpc_rule_ ## name (OrcCompiler *p, void *user, OrcInstruction *insn) \
+@@ -503,13 +539,11 @@ RULE(maxf, "vmaxfp", 0x1000040a)
+ RULE(minf, "vminfp", 0x1000044a)
+ RULE(cmpeqf, "vcmpeqfp", 0x100000c6)
+ 
+-RULE(addd, "xvadddp", 0xf0000307)
+-RULE(subd, "xvsubdp", 0xf0000347)
+-RULE(muld, "xvmuldp", 0xf0000387)
+-RULE(divd, "xvdivdp", 0xf00003c7)
+-RULE(mind, "xvmindp", 0xf0000747)
+-RULE(maxd, "xvmaxdp", 0xf0000707)
+-RULE(cmpeqd, "xvcmpeqdp", 0xf000031f)
++RULE_DP(addd, "xvadddp", 0xf0000307)
++RULE_DP(subd, "xvsubdp", 0xf0000347)
++RULE_DP(muld, "xvmuldp", 0xf0000387)
++RULE_DP(divd, "xvdivdp", 0xf00003c7)
++RULE_DP_SRC(cmpeqd, "xvcmpeqdp", 0xf000031f)
+ 
+ RULE(addq, "vaddudm", 0x100000c0)
+ RULE(subq, "vsubudm", 0x100004c0)
+@@ -541,13 +575,12 @@ powerpc_rule_mullb (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
++  int perm = powerpc_get_constant_full(p, 0x11011303, 0x15051707,
++      0x19091b0b, 0x1d0d1f0f);
+ 
+-  if (IS_POWERPC_BE (p)) {
+-    powerpc_emit_vmulesb (p, dest, src1, src2);
+-    powerpc_emit_vsldoi (p, dest, dest, dest, 1);
+-  } else {
+-    powerpc_emit_vmulosb (p, dest, src1, src2);
+-  }
++  powerpc_emit_vmulesb (p, p->tmpreg, src1, src2);
++  powerpc_emit_vmulosb (p, dest, src1, src2);
++  powerpc_emit_vperm(p, dest, dest, p->tmpreg, perm);
+ }
+ 
+ static void
+@@ -556,15 +589,12 @@ powerpc_rule_mulhsb (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
++  int perm = powerpc_get_constant_full(p, 0x10001202, 0x14041606,
++      0x18081a0a, 0x1c0c1e0e);
+ 
+-  if (IS_POWERPC_BE (p)) {
+-    powerpc_emit_vmulesb (p, dest, src1, src2);
+-  } else {
+-    int tmp = powerpc_get_constant (p, ORC_CONST_SPLAT_B, 8);
+-
+-    powerpc_emit_vmulosb (p, dest, src1, src2);
+-    powerpc_emit_VX_2 (p, "vsro", 0x1000044c, dest, src1, tmp);
+-  }
++  powerpc_emit_vmulesb (p, p->tmpreg, src1, src2);
++  powerpc_emit_vmulosb (p, dest, src1, src2);
++  powerpc_emit_vperm(p, dest, dest, p->tmpreg, perm);
+ }
+ 
+ static void
+@@ -573,15 +603,12 @@ powerpc_rule_mulhub (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
++  int perm = powerpc_get_constant_full(p, 0x10001202, 0x14041606,
++      0x18081a0a, 0x1c0c1e0e);
+ 
+-  if (IS_POWERPC_BE (p)) {
+-    powerpc_emit_vmuleub (p, dest, src1, src2);
+-  } else {
+-    int tmp = powerpc_get_constant (p, ORC_CONST_SPLAT_B, 8);
+-
+-    powerpc_emit_vmuloub (p, dest, src1, src2);
+-    powerpc_emit_VX_2 (p, "vsro", 0x1000044c, dest, dest, tmp);
+-  }
++  powerpc_emit_vmuleub (p, p->tmpreg, src1, src2);
++  powerpc_emit_vmuloub (p, dest, src1, src2);
++  powerpc_emit_vperm(p, dest, dest, p->tmpreg, perm);
+ }
+ 
+ static void
+@@ -590,16 +617,12 @@ powerpc_rule_mulhsw (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
++  int perm = powerpc_get_constant_full(p, 0x10110001, 0x14150405,
++      0x18190809, 0x1c1d0c0d);
+ 
+-  if (IS_POWERPC_BE (p)) {
+-    powerpc_emit_vmulesh (p, dest, src1, src2);
+-  } else {
+-    int tmp = powerpc_get_constant (p, ORC_CONST_SPLAT_B, 8);
+-
+-    powerpc_emit_vmulosh (p, dest, src1, src2);
+-    powerpc_emit_VX_2 (p, "vsro", 0x1000044c, dest, dest, tmp);
+-    powerpc_emit_VX_2 (p, "vsro", 0x1000044c, dest, dest, tmp);
+-  }
++  powerpc_emit_vmulesh (p, p->tmpreg, src1, src2);
++  powerpc_emit_vmulosh (p, dest, src1, src2);
++  powerpc_emit_vperm(p, dest, dest, p->tmpreg, perm);
+ }
+ 
+ static void
+@@ -608,16 +631,12 @@ powerpc_rule_mulhuw (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
++  int perm = powerpc_get_constant_full(p, 0x10110001, 0x14150405,
++      0x18190809, 0x1c1d0c0d);
+ 
+-  if (IS_POWERPC_BE (p)) {
+-    powerpc_emit_vmuleuh (p, dest, src1, src2);
+-  } else {
+-    int tmp = powerpc_get_constant (p, ORC_CONST_SPLAT_B, 8);
+-
+-    powerpc_emit_vmulouh (p, dest, src1, src2);
+-    powerpc_emit_VX_2 (p, "vsro", 0x1000044c, dest, dest, tmp);
+-    powerpc_emit_VX_2 (p, "vsro", 0x1000044c, dest, dest, tmp);
+-  }
++  powerpc_emit_vmuleuh(p, p->tmpreg, src1, src2);
++  powerpc_emit_vmulouh(p, dest, src1, src2);
++  powerpc_emit_vperm(p, dest, dest, p->tmpreg, perm);
+ }
+ 
+ static void
+@@ -765,10 +784,12 @@ powerpc_rule_mulsbw (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
+ 
++  powerpc_emit_vmulesb (p, p->tmpreg, src1, src2);
++  powerpc_emit_vmulosb (p, dest, src1, src2);
+   if (IS_POWERPC_BE (p)) {
+-    powerpc_emit_vmulesb (p, dest, src1, src2);
++    powerpc_emit_vmrghh (p, dest, p->tmpreg, dest);
+   } else {
+-    powerpc_emit_vmulosb (p, dest, src1, src2);
++    powerpc_emit_vmrglh (p, dest, p->tmpreg, dest);
+   }
+ }
+ 
+@@ -779,10 +800,12 @@ powerpc_rule_mulubw (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
+ 
++  powerpc_emit_vmuleub (p, p->tmpreg, src1, src2);
++  powerpc_emit_vmuloub (p, dest, src1, src2);
+   if (IS_POWERPC_BE (p)) {
+-    powerpc_emit_vmuleub (p, dest, src1, src2);
++    powerpc_emit_vmrghh (p, dest, p->tmpreg, dest);
+   } else {
+-    powerpc_emit_vmuloub (p, dest, src1, src2);
++    powerpc_emit_vmrglh (p, dest, p->tmpreg, dest);
+   }
+ }
+ 
+@@ -793,10 +816,12 @@ powerpc_rule_mulswl (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
+ 
++  powerpc_emit_vmulesh (p, p->tmpreg, src1, src2);
++  powerpc_emit_vmulosh (p, dest, src1, src2);
+   if (IS_POWERPC_BE (p)) {
+-    powerpc_emit_vmulesh (p, dest, src1, src2);
++    powerpc_emit_vmrghw (p, dest, p->tmpreg, dest);
+   } else {
+-    powerpc_emit_vmulosh (p, dest, src1, src2);
++    powerpc_emit_vmrglw (p, dest, p->tmpreg, dest);
+   }
+ }
+ 
+@@ -807,10 +832,12 @@ powerpc_rule_muluwl (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
+ 
++  powerpc_emit_vmuleuh (p, p->tmpreg, src1, src2);
++  powerpc_emit_vmulouh (p, dest, src1, src2);
+   if (IS_POWERPC_BE (p)) {
+-    powerpc_emit_vmuleuh (p, dest, src1, src2);
++    powerpc_emit_vmrghw (p, dest, p->tmpreg, dest);
+   } else {
+-    powerpc_emit_vmulouh (p, dest, src1, src2);
++    powerpc_emit_vmrglw (p, dest, p->tmpreg, dest);
+   }
+ }
+ 
+@@ -848,17 +875,13 @@ powerpc_rule_mulll (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
++  int tmp = orc_compiler_get_temp_reg(p);
++  int perm = powerpc_get_constant_full(p, 0x14151617, 0x04050607,
++      0x1c1d1e1f, 0x0c0d0e0f);
+ 
+-  if (IS_POWERPC_BE (p)) {
+-    int perm;
+-
+-    powerpc_emit_vmuleuw (p, dest, src1, src2);
+-    perm = powerpc_get_constant_full (p, 0x04050607, 0x04050607,
+-        0x0c0d0e0f, 0x0c0d0e0f);
+-    powerpc_emit_vperm (p, dest, dest, dest, perm);
+-  } else {
+-    powerpc_emit_vmulouw (p, dest, src1, src2);
+-  }
++  powerpc_emit_vmulouw (p, tmp, src1, src2);
++  powerpc_emit_vmuleuw (p, dest, src1, src2);
++  powerpc_emit_vperm (p, dest, tmp, dest, perm);
+ }
+ 
+ static void
+@@ -1428,20 +1451,11 @@ powerpc_rule_divf (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
+-  int y = orc_compiler_get_temp_reg (p);
+-  int t = orc_compiler_get_temp_reg (p);
+-  int c1;
+-  int c0;
+-
+-  c1 = powerpc_get_constant (p, ORC_CONST_SPLAT_L, 0x3f800000); /* 1.0 */
+-
+-  powerpc_emit_VX_db (p, "vrefp", 0x1000010a, y, src2);
+-
+-  powerpc_emit_VA_acb (p, "vnmsubfp", 0x1000002f, t, y, c1, src2);
+-  powerpc_emit_VA_acb (p, "vmaddfp", 0x1000002e, y, y, y, t);
+-
+-  c0 = powerpc_get_constant (p, ORC_CONST_SPLAT_L, 0x00000000); /* 0.0 */
+-  powerpc_emit_VA_acb (p, "vmaddfp", 0x1000002e, dest, y, c0, src1);
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0);
++  powerpc_denormalize_sp_zero (p, src1, zero);
++  powerpc_denormalize_sp_zero (p, src2, zero);
++  powerpc_emit_VX_2(p, "xvdivsp", 0xf00002c7, dest, src1, src2);
++  powerpc_denormalize_sp_zero (p, dest, zero);
+ }
+ 
+ static void
+@@ -1521,8 +1535,10 @@ powerpc_rule_sqrtf(OrcCompiler* p, void* user, OrcInstruction* insn)
+ {
+   int src1 = ORC_SRC_ARG(p, insn, 0);
+   int dest = ORC_DEST_ARG(p, insn, 0);
+-
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0);
++  powerpc_denormalize_sp_zero (p, src1, zero);
+   powerpc_emit_VX_db(p, "xvsqrtsp", 0xf000022f, dest, src1);
++  powerpc_denormalize_sp_zero (p, dest, zero);
+ }
+ 
+ static void
+@@ -1530,18 +1546,65 @@ powerpc_rule_sqrtd(OrcCompiler* p, void* user, OrcInstruction* insn)
+ {
+   int src1 = ORC_SRC_ARG(p, insn, 0);
+   int dest = ORC_DEST_ARG(p, insn, 0);
+-
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0);
++  powerpc_denormalize_dp_zero (p, src1, zero);
+   powerpc_emit_VX_db(p, "xvsqrtdp", 0xf000032f, dest, src1);
++  powerpc_denormalize_dp_zero (p, dest, zero);
+ }
+ 
+ static void
+-powerpc_rule_cmpltd (OrcCompiler *p, void *user, OrcInstruction *insn)
++powerpc_rule_mind (OrcCompiler *p, void *user, OrcInstruction *insn)
+ {
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0);
++  int mask = powerpc_get_constant_full (p, 0x7ff80000, 0x00000000,
++      0x7ff80000, 0x00000000);
++  int tmp2 = orc_compiler_get_temp_reg(p);
++  int tmp = p->tmpreg;
++  powerpc_denormalize_dp_zero (p, src1, zero);
++  powerpc_denormalize_dp_zero (p, src2, zero);
++  powerpc_emit_VX_2(p, "xvmindp", 0xf0000747, tmp2, src2, src1);
++  // Handle QNaN
++  powerpc_emit_VX_2(p, "xxland", 0xf0000417, tmp, src1, mask);
++  powerpc_emit_VX_2(p, "vcmpequd", 0x100000c7, tmp, tmp, mask);
++  powerpc_emit_VA(p, "xxsel", 0xf000003f, dest, tmp2, src1, tmp);
++}
+ 
+-//  powerpc_emit_VXR (p, "vcmpgtfp", 0x100002c6, dest, src2, src1, FALSE);
++static void
++powerpc_rule_maxd (OrcCompiler *p, void *user, OrcInstruction *insn)
++{
++  int src1 = ORC_SRC_ARG (p, insn, 0);
++  int src2 = ORC_SRC_ARG (p, insn, 1);
++  int dest = ORC_DEST_ARG (p, insn, 0);
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0);
++  int mask = powerpc_get_constant_full (p, 0x7ff80000, 0x00000000,
++      0x7ff80000, 0x00000000);
++  int tmp2 = orc_compiler_get_temp_reg(p);
++  int tmp3 = orc_compiler_get_temp_reg(p);
++  int tmp = p->tmpreg;
++  powerpc_denormalize_dp_zero (p, src1, zero);
++  powerpc_denormalize_dp_zero (p, src2, zero);
++  powerpc_emit_VX_2(p, "xvmaxdp", 0xf0000707, tmp2, src2, src1);
++  // Handle QNaN
++  powerpc_emit_VX_2(p, "xxland", 0xf0000417, tmp, src2, mask);
++  powerpc_emit_VX_2(p, "vcmpequd", 0x100000c7, tmp, tmp, mask);
++  powerpc_emit_VA(p, "xxsel", 0xf000003f, tmp3, tmp2, src2, tmp);
++  powerpc_emit_VX_2(p, "xxland", 0xf0000417, tmp, src1, mask);
++  powerpc_emit_VX_2(p, "vcmpequd", 0x100000c7, tmp, tmp, mask);
++  powerpc_emit_VA(p, "xxsel", 0xf000003f, dest, tmp3, src1, tmp);
++}
++
++static void
++powerpc_rule_cmpltd (OrcCompiler *p, void *user, OrcInstruction *insn)
++{
++  int src1 = ORC_SRC_ARG (p, insn, 0);
++  int src2 = ORC_SRC_ARG (p, insn, 1);
++  int dest = ORC_DEST_ARG (p, insn, 0);
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0);
++  powerpc_denormalize_dp_zero (p, src1, zero);
++  powerpc_denormalize_dp_zero (p, src2, zero);
+   powerpc_emit_VX_2(p, "xvcmpgtdp", 0xf000035f, dest, src2, src1);
+ }
+ 
+@@ -1551,8 +1614,9 @@ powerpc_rule_cmpled (OrcCompiler *p, void *user, OrcInstruction *insn)
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int src2 = ORC_SRC_ARG (p, insn, 1);
+   int dest = ORC_DEST_ARG (p, insn, 0);
+-
+-//  powerpc_emit_VXR (p, "vcmpgefp", 0x100001c6, dest, src2, src1, FALSE);
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0);
++  powerpc_denormalize_dp_zero (p, src1, zero);
++  powerpc_denormalize_dp_zero (p, src2, zero);
+   powerpc_emit_VX_2(p, "xvcmpgedp", 0xf000039f, dest, src2, src1);
+ }
+ 
+@@ -1561,11 +1625,13 @@ powerpc_rule_convld (OrcCompiler *p, void *user, OrcInstruction *insn)
+ {
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int dest = ORC_DEST_ARG (p, insn, 0);
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0);
+ 
+   if (IS_POWERPC_LE (p)) {
+     powerpc_emit_vsldoi(p, src1, src1, src1, 4);
+   }
+   powerpc_emit_VX_db (p, "xvcvsxddp", 0xf00003e3, dest, src1);
++  powerpc_denormalize_dp_zero (p, dest, zero);
+ }
+ 
+ static void
+@@ -1573,8 +1639,19 @@ powerpc_rule_convdl (OrcCompiler *p, void *user, OrcInstruction *insn)
+ {
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int dest = ORC_DEST_ARG (p, insn, 0);
++  int mask = powerpc_get_constant_full (p, 0xfff00000, 0x00000000,
++      0xfff00000, 0x00000000);
++  int nan = powerpc_get_constant_full (p, 0x7ff00000, 0x00000000,
++      0x7ff00000, 0x00000000);
++  int maxint = powerpc_get_constant_full (p, 0x7fffffff, 0x7fffffff,
++      0x7fffffff, 0x7fffffff);
++  int tmp = p->tmpreg;
+ 
++  powerpc_emit_VX_2(p, "xxland", 0xf0000417, tmp, src1, mask);
++  powerpc_emit_VX_2(p, "vcmpequd", 0x100000c7, tmp, tmp, nan);
+   powerpc_emit_VX_db (p, "xvcvdpsxws", 0xf0000363, dest, src1);
++  powerpc_emit_VA(p, "xxsel", 0xf000003f, dest, dest, maxint, tmp);
++
+   if (IS_POWERPC_LE (p)) {
+     int tmp = powerpc_get_constant (p, ORC_CONST_SPLAT_B, 32);
+     powerpc_emit_VX_2 (p, "vsro", 0x1000044c, dest, dest, tmp);
+@@ -1586,11 +1663,23 @@ powerpc_rule_convfd (OrcCompiler *p, void *user, OrcInstruction *insn)
+ {
+   int src1 = ORC_SRC_ARG (p, insn, 0);
+   int dest = ORC_DEST_ARG (p, insn, 0);
++  int tmp2 = orc_compiler_get_temp_reg(p);
++  int zero = powerpc_get_constant (p, ORC_CONST_ZERO, 0);
++  int mask = powerpc_get_constant_full (p, 0x7f800000, 0x00000000,
++      0x7f800000, 0x00000000);
++  int tmp = p->tmpreg;
+ 
+   if (IS_POWERPC_LE (p)) {
+-    powerpc_emit_vsldoi(p, src1, src1, src1, 4);
++    powerpc_emit_vsldoi(p, tmp2, src1, src1, 4);
++  } else {
++    powerpc_emit_vor(p, tmp2, src1, src1);
+   }
+-  powerpc_emit_VX_db (p, "xvcvspdp", 0xf0000727, dest, src1);
++  powerpc_emit_VX_db (p, "xvcvspdp", 0xf0000727, dest, tmp2);
++
++  powerpc_emit_VX_2(p, "xxland", 0xf0000417, tmp, tmp2, mask);
++  powerpc_emit_VX_2(p, "vcmpequd", 0x100000c7, tmp, tmp, zero);
++  powerpc_emit_VA(p, "xxsel", 0xf000003f, tmp, dest, zero, tmp);
++  powerpc_emit_VX_2(p, "xvcpsgndp", 0xf0000787, dest, dest, tmp);
+ }
+ 
+ static void
+@@ -1604,6 +1693,7 @@ powerpc_rule_convdf (OrcCompiler *p, void *user, OrcInstruction *insn)
+     int tmp = powerpc_get_constant (p, ORC_CONST_SPLAT_B, 32);
+     powerpc_emit_VX_2 (p, "vsro", 0x1000044c, dest, dest, tmp);
+   }
++  powerpc_denormalize_sp(p, dest);
+ }
+ 
+ 
+@@ -1757,7 +1847,6 @@ orc_compiler_powerpc_register_rules (OrcTarget *target)
+   REG(cmplef);
+   REG(cmpltf);
+   REG(mulf);
+-  REG(divf);
+   REG(convfl);
+   REG(convlf);
+ 
+@@ -1789,6 +1878,7 @@ orc_compiler_powerpc_register_rules (OrcTarget *target)
+ 
+   rule_set = orc_rule_set_new(orc_opcode_set_get("sys"), target, ORC_TARGET_POWERPC_VSX);
+ 
++  REG(divf);
+   REG(sqrtf);
+   REG(addd);
+   REG(subd);
+diff --git a/orc/orctarget.h b/orc/orctarget.h
+index 642bfe3..53000b3 100644
+--- a/orc/orctarget.h
++++ b/orc/orctarget.h
+@@ -19,7 +19,7 @@ enum {
+   ORC_TARGET_FAST_DENORMAL = (1<<31)
+ };
+ 
+-enum {
++typedef enum {
+   ORC_TARGET_POWERPC_64BIT = (1<<0),
+   ORC_TARGET_POWERPC_LE = (1<<1),
+   ORC_TARGET_POWERPC_ALTIVEC = (1<<2),
+diff --git a/testsuite/exec_parse.c b/testsuite/exec_parse.c
+index 0858413..859ad4e 100644
+--- a/testsuite/exec_parse.c
++++ b/testsuite/exec_parse.c
+@@ -20,6 +20,7 @@ main (int argc, char *argv[])
+   char *code;
+   int n;
+   int i;
++  int ret;
+   OrcProgram **programs;
+   const char *filename = NULL;
+ 
+@@ -45,7 +46,11 @@ main (int argc, char *argv[])
+ 
+   for(i=0;i<n;i++){
+     if (verbose) printf("%s\n", programs[i]->name);
+-    orc_test_compare_output_full (programs[i], 0);
++    ret = orc_test_compare_output_full (programs[i], 0);
++    if (!ret) {
++      printf("failed %s\n", programs[i]->name);
++      error = TRUE;
++    }
+     orc_program_free (programs[i]);
+   }
+ 
+diff --git a/testsuite/memcpy_speed.c b/testsuite/memcpy_speed.c
+index a1a69f1..4cf1c39 100644
+--- a/testsuite/memcpy_speed.c
++++ b/testsuite/memcpy_speed.c
+@@ -102,7 +102,7 @@ main(int argc, char *argv[])
+     max = 140;
+   }
+ 
+-  for(i=0;i<max;i++){
++  for(i=0;i<max;i+=2){
+     double x = i*0.1 + 6.0;
+     int size = pow(2.0, x);
+