Change meaning of convfl/convdl to be saturated

Some architectures (powerpc) clamp out-of-range values to the
integer range when converting from float to int.  Others (x86)
maps out-of-range values to 0x80000000.  Since we have to choose
one, saturating seems like the better choice.

4 files changed, 73 insertions, 10 deletions

diff --git a/orc/opcodes.h b/orc/opcodes.h
index bdff6a0..821404c 100644
--- a/orc/opcodes.h
+++ b/orc/opcodes.h
@@ -135,7 +135,6 @@ BINARY_F(minf, "ORC_MINF(ORC_DENORMAL(%s),ORC_DENORMAL(%s))")
 BINARY_FL(cmpeqf, "(ORC_DENORMAL(%s) == ORC_DENORMAL(%s)) ? (~0) : 0")
 BINARY_FL(cmpltf, "(ORC_DENORMAL(%s) < ORC_DENORMAL(%s)) ? (~0) : 0")
 BINARY_FL(cmplef, "(ORC_DENORMAL(%s) <= ORC_DENORMAL(%s)) ? (~0) : 0")
-UNARY_FL(convfl, "(int)%s")
 UNARY_LF(convlf, "%s")
 
 BINARY_D(addd, "ORC_DENORMAL_D(ORC_DENORMAL_D(%s) + ORC_DENORMAL_D(%s))")
@@ -148,7 +147,6 @@ BINARY_D(mind, "ORC_MIND(ORC_DENORMAL_D(%s),ORC_DENORMAL_D(%s))")
 BINARY_DQ(cmpeqd, "(ORC_DENORMAL_D(%s) == ORC_DENORMAL_D(%s)) ? (~0ULL) : 0")
 BINARY_DQ(cmpltd, "(ORC_DENORMAL_D(%s) < ORC_DENORMAL_D(%s)) ? (~0ULL) : 0")
 BINARY_DQ(cmpled, "(ORC_DENORMAL_D(%s) <= ORC_DENORMAL_D(%s)) ? (~0ULL) : 0")
-UNARY_DL(convdl, "(int)%s")
 UNARY_LD(convld, "%s")
 UNARY_DF(convdf, "ORC_DENORMAL ((float) %s)")
 UNARY_FD(convfd, "ORC_DENORMAL (%s)")
diff --git a/orc/orcemulateopcodes.c b/orc/orcemulateopcodes.c
index 8ea7c55..8a89ca8 100644
--- a/orc/orcemulateopcodes.c
+++ b/orc/orcemulateopcodes.c
@@ -4102,7 +4102,12 @@ emulate_convfl (OrcOpcodeExecutor *ex, int offset, int n)
     /* 0: loadl */
     var32 = ptr4[i];
     /* 1: convfl */
-    var33.i = (int)var32.f;
+    {
+       int tmp;
+       tmp = (int)var32.f;
+       if (tmp == 0x80000000 && !(var32.i&0x80000000)) tmp = 0x7fffffff;
+       var33.i = tmp;
+    }
     /* 2: storel */
     ptr0[i] = var33;
   }
@@ -4423,7 +4428,12 @@ emulate_convdl (OrcOpcodeExecutor *ex, int offset, int n)
     /* 0: loadq */
     var32 = ptr4[i];
     /* 1: convdl */
-    var33.i = (int)var32.f;
+    {
+       int tmp;
+       tmp = var32.f;
+       if (tmp == 0x80000000 && !(var32.i&0x8000000000000000ULL)) tmp = 0x7fffffff;
+       var33.i = tmp;
+    }
     /* 2: storel */
     ptr0[i] = var33;
   }
diff --git a/orc/orcprogram-c.c b/orc/orcprogram-c.c
index 42d88b6..690219e 100644
--- a/orc/orcprogram-c.c
+++ b/orc/orcprogram-c.c
@@ -1001,6 +1001,40 @@ c_rule_divluw (OrcCompiler *p, void *user, OrcInstruction *insn)
       dest, src2, src1, src2);
 }
 
+static void
+c_rule_convfl (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  char dest[40], src[40], src_i[40];
+
+  c_get_name_int (dest, p, insn, insn->dest_args[0]);
+  c_get_name_float (src, p, insn, insn->src_args[0]);
+  c_get_name_int (src_i, p, insn, insn->src_args[0]);
+
+  ORC_ASM_CODE(p, "    {\n");
+  ORC_ASM_CODE(p,"       int tmp;\n");
+  ORC_ASM_CODE(p,"       tmp = (int)%s;\n", src);
+  ORC_ASM_CODE(p,"       if (tmp == 0x80000000 && !(%s&0x80000000)) tmp = 0x7fffffff;\n", src_i);
+  ORC_ASM_CODE(p,"       %s = tmp;\n", dest);
+  ORC_ASM_CODE(p, "    }\n");
+}
+
+static void
+c_rule_convdl (OrcCompiler *p, void *user, OrcInstruction *insn)
+{
+  char dest[40], src[40], src_i[40];
+
+  c_get_name_int (dest, p, insn, insn->dest_args[0]);
+  c_get_name_float (src, p, insn, insn->src_args[0]);
+  c_get_name_int (src_i, p, insn, insn->src_args[0]);
+
+  ORC_ASM_CODE(p, "    {\n");
+  ORC_ASM_CODE(p,"       int tmp;\n");
+  ORC_ASM_CODE(p,"       tmp = %s;\n", src);
+  ORC_ASM_CODE(p,"       if (tmp == 0x80000000 && !(%s&0x8000000000000000ULL)) tmp = 0x7fffffff;\n", src_i);
+  ORC_ASM_CODE(p,"       %s = tmp;\n", dest);
+  ORC_ASM_CODE(p, "    }\n");
+}
+
 static OrcTarget c_target = {
   "c",
   FALSE,
@@ -1098,5 +1132,7 @@ orc_c_init (void)
   orc_rule_register (rule_set, "splatw3q", c_rule_splatw3q, NULL);
   orc_rule_register (rule_set, "div255w", c_rule_div255w, NULL);
   orc_rule_register (rule_set, "divluw", c_rule_divluw, NULL);
+  orc_rule_register (rule_set, "convfl", c_rule_convfl, NULL);
+  orc_rule_register (rule_set, "convdl", c_rule_convdl, NULL);
 }
 
diff --git a/orc/orcrules-sse.c b/orc/orcrules-sse.c
index b500d02..ad66827 100644
--- a/orc/orcrules-sse.c
+++ b/orc/orcrules-sse.c
@@ -2157,17 +2157,36 @@ sse_rule_cmpled (OrcCompiler *p, void *user, OrcInstruction *insn)
 static void
 sse_rule_convfl (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
-  orc_sse_emit_f30f (p, "cvttps2dq", 0x5b,
-      p->vars[insn->src_args[0]].alloc,
-      p->vars[insn->dest_args[0]].alloc);
+  int src = p->vars[insn->src_args[0]].alloc;
+  int dest = p->vars[insn->dest_args[0]].alloc;
+  int tmpc;
+  int tmp = orc_compiler_get_temp_reg (p);
+  
+  tmpc = orc_compiler_get_temp_constant (p, 4, 0x80000000);
+  orc_sse_emit_movdqa (p, src, tmp);
+  orc_sse_emit_f30f (p, "cvttps2dq", 0x5b, src, dest);
+  orc_sse_emit_psrad (p, 31, tmp);
+  orc_sse_emit_pcmpeqd (p, dest, tmpc);
+  orc_sse_emit_pandn (p, tmpc, tmp);
+  orc_sse_emit_paddd (p, tmp, dest);
+
 }
 
 static void
 sse_rule_convdl (OrcCompiler *p, void *user, OrcInstruction *insn)
 {
-  orc_sse_emit_660f (p, "cvttpd2dq", 0xe6,
-      p->vars[insn->src_args[0]].alloc,
-      p->vars[insn->dest_args[0]].alloc);
+  int src = p->vars[insn->src_args[0]].alloc;
+  int dest = p->vars[insn->dest_args[0]].alloc;
+  int tmpc;
+  int tmp = orc_compiler_get_temp_reg (p);
+  
+  tmpc = orc_compiler_get_temp_constant (p, 4, 0x80000000);
+  orc_sse_emit_pshufd (p, ORC_SSE_SHUF(3,1,3,1), src, tmp);
+  orc_sse_emit_660f (p, "cvttpd2dq", 0xe6, src, dest);
+  orc_sse_emit_psrad (p, 31, tmp);
+  orc_sse_emit_pcmpeqd (p, dest, tmpc);
+  orc_sse_emit_pandn (p, tmpc, tmp);
+  orc_sse_emit_paddd (p, tmp, dest);
 }
 
 static void