Merge branch 'jv/erts/optimize-cmp'

* jv/erts/optimize-cmp:
  Unify comparison macros in erl_utils.h
  Avoid erts_cmp jump in atom, int and float comparisons

5 files changed, 51 insertions, 39 deletions

diff --git a/erts/emulator/beam/atom.c b/erts/emulator/beam/atom.c
index fd2adac676..099c00bcf6 100644
--- a/erts/emulator/beam/atom.c
+++ b/erts/emulator/beam/atom.c
@@ -176,7 +176,7 @@ atom_alloc(Atom* tmpl)
 
     /*
      * Precompute ordinal value of first 3 bytes + 7 bits.
-     * This is used by utils.c:cmp_atoms().
+     * This is used by utils.c:erts_cmp_atoms().
      * We cannot use the full 32 bits of the first 4 bytes,
      * since we use the sign of the difference between two
      * ordinal values to represent their relative order.
diff --git a/erts/emulator/beam/beam_emu.c b/erts/emulator/beam/beam_emu.c
index d4fb36acc3..10f132abfc 100644
--- a/erts/emulator/beam/beam_emu.c
+++ b/erts/emulator/beam/beam_emu.c
@@ -721,10 +721,10 @@ void** beam_ops;
 #define NotEqualImmed(X, Y, Action) if (X == Y) { Action; }
 #define EqualExact(X, Y, Action) if (!EQ(X,Y)) { Action; }
 #define NotEqualExact(X, Y, Action) if (EQ(X,Y)) { Action; }
-#define Equal(X, Y, Action) if (!CMP_EQ(X,Y)) { Action; }
-#define NotEqual(X, Y, Action) if (!CMP_NE(X,Y)) { Action; }
-#define IsLessThan(X, Y, Action) if (CMP_GE(X, Y)) { Action; }
-#define IsGreaterEqual(X, Y, Action) if (CMP_LT(X, Y)) { Action; }
+#define Equal(X, Y, Action) CMP_EQ_ACTION(X,Y,Action)
+#define NotEqual(X, Y, Action) CMP_NE_ACTION(X,Y,Action)
+#define IsLessThan(X, Y, Action) CMP_LT_ACTION(X,Y,Action)
+#define IsGreaterEqual(X, Y, Action) CMP_GE_ACTION(X,Y,Action)
 
 #define IsFloat(Src, Fail) if (is_not_float(Src)) { Fail; }
 
diff --git a/erts/emulator/beam/erl_bif_op.c b/erts/emulator/beam/erl_bif_op.c
index d53a9e11ca..0f20ded1d6 100644
--- a/erts/emulator/beam/erl_bif_op.c
+++ b/erts/emulator/beam/erl_bif_op.c
@@ -89,22 +89,22 @@ BIF_RETTYPE not_1(BIF_ALIST_1)
 
 BIF_RETTYPE sgt_2(BIF_ALIST_2)
 {
-    BIF_RET(cmp_gt(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
+    BIF_RET(CMP_GT(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
 }
 
 BIF_RETTYPE sge_2(BIF_ALIST_2)
 {
-    BIF_RET(cmp_ge(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
+    BIF_RET(CMP_GE(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
 }
 
 BIF_RETTYPE slt_2(BIF_ALIST_2)
 {
-    BIF_RET(cmp_lt(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
+    BIF_RET(CMP_LT(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
 }
 
 BIF_RETTYPE sle_2(BIF_ALIST_2)
 {
-    BIF_RET(cmp_le(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
+    BIF_RET(CMP_LE(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
 }
 
 BIF_RETTYPE seq_2(BIF_ALIST_2)
@@ -114,7 +114,7 @@ BIF_RETTYPE seq_2(BIF_ALIST_2)
 
 BIF_RETTYPE seqeq_2(BIF_ALIST_2)
 {
-    BIF_RET(cmp_eq(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
+    BIF_RET(CMP_EQ(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
 }
 
 BIF_RETTYPE sneq_2(BIF_ALIST_2)
@@ -124,7 +124,7 @@ BIF_RETTYPE sneq_2(BIF_ALIST_2)
 
 BIF_RETTYPE sneqeq_2(BIF_ALIST_2)
 {
-    BIF_RET(cmp_ne(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
+    BIF_RET(CMP_NE(BIF_ARG_1, BIF_ARG_2) ? am_true : am_false);
 }
 
 BIF_RETTYPE is_atom_1(BIF_ALIST_1)
diff --git a/erts/emulator/beam/erl_utils.h b/erts/emulator/beam/erl_utils.h
index 4058d63eaf..dd7b925bd3 100644
--- a/erts/emulator/beam/erl_utils.h
+++ b/erts/emulator/beam/erl_utils.h
@@ -161,30 +161,42 @@ int eq(Eterm, Eterm);
 
 #define EQ(x,y) (((x) == (y)) || (is_not_both_immed((x),(y)) && eq((x),(y))))
 
+int erts_cmp_atoms(Eterm a, Eterm b);
 Sint erts_cmp(Eterm, Eterm, int, int);
+Sint erts_cmp_compound(Eterm, Eterm, int, int);
 Sint cmp(Eterm a, Eterm b);
 #define CMP(A,B)                         erts_cmp(A,B,0,0)
 #define CMP_TERM(A,B)                    erts_cmp(A,B,1,0)
 #define CMP_EQ_ONLY(A,B)                 erts_cmp(A,B,0,1)
 
-#define cmp_lt(a,b)          (CMP((a),(b)) <  0)
-#define cmp_le(a,b)          (CMP((a),(b)) <= 0)
-#define cmp_eq(a,b)          (CMP_EQ_ONLY((a),(b)) == 0)
-#define cmp_ne(a,b)          (CMP_EQ_ONLY((a),(b)) != 0)
-#define cmp_ge(a,b)          (CMP((a),(b)) >= 0)
-#define cmp_gt(a,b)          (CMP((a),(b)) >  0)
-
-#define cmp_lt_term(a,b)     (CMP_TERM((a),(b)) <  0)
-#define cmp_le_term(a,b)     (CMP_TERM((a),(b)) <= 0)
-#define cmp_ge_term(a,b)     (CMP_TERM((a),(b)) >= 0)
-#define cmp_gt_term(a,b)     (CMP_TERM((a),(b)) >  0)
-
-#define CMP_LT(a,b)          ((a) != (b) && cmp_lt((a),(b)))
-#define CMP_GE(a,b)          ((a) == (b) || cmp_ge((a),(b)))
-#define CMP_EQ(a,b)          ((a) == (b) || cmp_eq((a),(b)))
-#define CMP_NE(a,b)          ((a) != (b) && cmp_ne((a),(b)))
-
-#define CMP_LT_TERM(a,b)     ((a) != (b) && cmp_lt_term((a),(b)))
-#define CMP_GE_TERM(a,b)     ((a) == (b) || cmp_ge_term((a),(b)))
+#define CMP_LT(a,b)          ((a) != (b) && CMP((a),(b)) <  0)
+#define CMP_LE(a,b)          ((a) == (b) || CMP((a),(b)) <= 0)
+#define CMP_EQ(a,b)          ((a) == (b) || CMP_EQ_ONLY((a),(b)) == 0)
+#define CMP_NE(a,b)          ((a) != (b) && CMP_EQ_ONLY((a),(b)) != 0)
+#define CMP_GE(a,b)          ((a) == (b) || CMP((a),(b)) >= 0)
+#define CMP_GT(a,b)          ((a) != (b) && CMP((a),(b)) >  0)
+
+#define CMP_EQ_ACTION(X,Y,Action)	\
+    if ((X) != (Y)) { CMP_SPEC((X),(Y),!=,Action,1); }
+#define CMP_NE_ACTION(X,Y,Action)	\
+    if ((X) == (Y)) { Action; } else { CMP_SPEC((X),(Y),==,Action,1); }
+#define CMP_GE_ACTION(X,Y,Action)	\
+    if ((X) != (Y)) { CMP_SPEC((X),(Y),<,Action,0); }
+#define CMP_LT_ACTION(X,Y,Action)	\
+    if ((X) == (Y)) { Action; } else { CMP_SPEC((X),(Y),>=,Action,0); }
+
+#define CMP_SPEC(X,Y,Op,Action,EqOnly)				\
+    if (is_atom(X) && is_atom(Y)) {				\
+	if (erts_cmp_atoms(X, Y) Op 0) { Action; };		\
+    } else if (is_both_small(X, Y)) {				\
+	if (signed_val(X) Op signed_val(Y)) { Action; };	\
+    } else if (is_float(X) && is_float(Y)) {			\
+        FloatDef af, bf;					\
+        GET_DOUBLE(X, af);					\
+        GET_DOUBLE(Y, bf);					\
+        if (af.fd Op bf.fd) { Action; };			\
+    } else {							\
+	if (erts_cmp_compound(X,Y,0,EqOnly) Op 0) { Action; };	\
+    }
 
 #endif
diff --git a/erts/emulator/beam/utils.c b/erts/emulator/beam/utils.c
index ef851d840d..6aad1ff778 100644
--- a/erts/emulator/beam/utils.c
+++ b/erts/emulator/beam/utils.c
@@ -2991,7 +2991,7 @@ static int cmpbytes(byte *s1, int l1, byte *s2, int l2)
 
 #define float_comp(x,y)    (((x)<(y)) ? -1 : (((x)==(y)) ? 0 : 1))
 
-static int cmp_atoms(Eterm a, Eterm b)
+int erts_cmp_atoms(Eterm a, Eterm b)
 {
     Atom *aa = atom_tab(atom_val(a));
     Atom *bb = atom_tab(atom_val(b));
@@ -3010,12 +3010,12 @@ Sint cmp(Eterm a, Eterm b)
     return erts_cmp(a, b, 0, 0);
 }
 
-static Sint erts_cmp_compound(Eterm a, Eterm b, int exact, int eq_only);
+Sint erts_cmp_compound(Eterm a, Eterm b, int exact, int eq_only);
 
 Sint erts_cmp(Eterm a, Eterm b, int exact, int eq_only)
 {
     if (is_atom(a) && is_atom(b)) {
-        return cmp_atoms(a, b);
+        return erts_cmp_atoms(a, b);
     } else if (is_both_small(a, b)) {
         return (signed_val(a) - signed_val(b));
     } else if (is_float(a) && is_float(b)) {
@@ -3032,7 +3032,7 @@ Sint erts_cmp(Eterm a, Eterm b, int exact, int eq_only)
  * exact = 1 -> term-based compare
  * exact = 0 -> arith-based compare
  */
-static Sint erts_cmp_compound(Eterm a, Eterm b, int exact, int eq_only)
+Sint erts_cmp_compound(Eterm a, Eterm b, int exact, int eq_only)
 {
 #define PSTACK_TYPE struct erts_cmp_hashmap_state
     struct erts_cmp_hashmap_state {
@@ -3089,7 +3089,7 @@ static Sint erts_cmp_compound(Eterm a, Eterm b, int exact, int eq_only)
     do {								\
 	if((AN) != (BN)) {						\
             if((AN)->sysname != (BN)->sysname)				\
-                RETURN_NEQ(cmp_atoms((AN)->sysname, (BN)->sysname));	\
+                RETURN_NEQ(erts_cmp_atoms((AN)->sysname, (BN)->sysname));	\
 	    ASSERT((AN)->creation != (BN)->creation);			\
 	    RETURN_NEQ(((AN)->creation < (BN)->creation) ? -1 : 1);	\
 	}								\
@@ -3107,7 +3107,7 @@ tailrecur_ne:
     /* deal with majority (?) cases by brute-force */
     if (is_atom(a)) {
 	if (is_atom(b)) {
-	    ON_CMP_GOTO(cmp_atoms(a, b));
+	    ON_CMP_GOTO(erts_cmp_atoms(a, b));
 	}
     } else if (is_both_small(a, b)) {
 	ON_CMP_GOTO(signed_val(a) - signed_val(b));
@@ -3341,10 +3341,10 @@ tailrecur_ne:
 		    Export* a_exp = *((Export **) (export_val(a) + 1));
 		    Export* b_exp = *((Export **) (export_val(b) + 1));
 
-		    if ((j = cmp_atoms(a_exp->code[0], b_exp->code[0])) != 0) {
+		    if ((j = erts_cmp_atoms(a_exp->code[0], b_exp->code[0])) != 0) {
 			RETURN_NEQ(j);
 		    }
-		    if ((j = cmp_atoms(a_exp->code[1], b_exp->code[1])) != 0) {
+		    if ((j = erts_cmp_atoms(a_exp->code[1], b_exp->code[1])) != 0) {
 			RETURN_NEQ(j);
 		    }
 		    ON_CMP_GOTO((Sint) a_exp->code[2] - (Sint) b_exp->code[2]);
@@ -3659,7 +3659,7 @@ term_array: /* arrays in 'aa' and 'bb', length in 'i' */
 	b = *bb++;
 	if (!is_same(a, b)) {
 	    if (is_atom(a) && is_atom(b)) {
-		if ((j = cmp_atoms(a, b)) != 0) {
+		if ((j = erts_cmp_atoms(a, b)) != 0) {
 		    goto not_equal;
 		}
 	    } else if (is_both_small(a, b)) {