diff --git a/usr.sbin/pmcstudy/pmcstudy.8 b/usr.sbin/pmcstudy/pmcstudy.8
index ebe63c40301..7c038975dd7 100644
--- a/usr.sbin/pmcstudy/pmcstudy.8
+++ b/usr.sbin/pmcstudy/pmcstudy.8
@@ -32,7 +32,7 @@
 .Nd Perform various studies on a system's overall PMCs.
 .Sh SYNOPSIS
 .Nm
-.Oo Fl i Ar inputfile | Fl T | Fl v | Fl m Ar max | Fl e exp | Fl Ar E | Fl h | fl H Oc
+.Oo Fl i Ar inputfile | Fl A |  Fl T | Fl v | Fl m Ar max | Fl e exp | Fl Ar E | Fl h | fl H Oc
 .Nm
 .Fl i Ar inputfile
 .Nm
@@ -128,6 +128,8 @@ test like
 "UOPS_RETIRED.RETIRE_SLOTS / (4 * CPU_CLK_UNHALTED.THREAD_P)".
 .It Fl L
 This option will list all known PMCs and their abbreviation (%NNN).
+.It Fl A
+Run all canned tests.
 .El
 .Sh SEE ALSO
 .Xr pmc 3 ,
diff --git a/usr.sbin/pmcstudy/pmcstudy.c b/usr.sbin/pmcstudy/pmcstudy.c
index 06ded9723fc..16c9f514833 100644
--- a/usr.sbin/pmcstudy/pmcstudy.c
+++ b/usr.sbin/pmcstudy/pmcstudy.c
@@ -38,6 +38,9 @@
 #include "eval_expr.h"
 __FBSDID("$FreeBSD$");
 
+static int max_pmc_counters = 1;
+static int run_all = 0;
+
 #define MAX_COUNTER_SLOTS 1024
 #define MAX_NLEN 64
 #define MAX_CPU 64
@@ -45,20 +48,20 @@ static int verbose = 0;
 
 extern char **environ;
 extern struct expression *master_exp;
-struct expression *master_exp = NULL;
+struct expression *master_exp=NULL;
 
 #define PMC_INITIAL_ALLOC 512
 extern char **valid_pmcs;
 char **valid_pmcs = NULL;
 extern int valid_pmc_cnt;
-int valid_pmc_cnt = 0;
+int valid_pmc_cnt=0;
 extern int pmc_allocated_cnt;
-int pmc_allocated_cnt = 0;
+int pmc_allocated_cnt=0;
 
 /*
  * The following two varients on popen and pclose with
  * the cavet that they get you the PID so that you
- * can supply it to pclose so it can send a SIGTERM
+ * can supply it to pclose so it can send a SIGTERM 
  *  to the process.
  */
 static FILE *
@@ -75,7 +78,7 @@ my_popen(const char *command, const char *dir, pid_t *p_pid)
 	if ((strcmp(dir, "r") != 0) &&
 	    (strcmp(dir, "w") != 0)) {
 		errno = EINVAL;
-		return (NULL);
+		return(NULL);
 	}
 	if (pipe(pdesin) < 0)
 		return (NULL);
@@ -94,14 +97,14 @@ my_popen(const char *command, const char *dir, pid_t *p_pid)
 	argv[3] = NULL;
 
 	switch (pid = fork()) {
-	case -1:		/* Error. */
+	case -1:			/* Error. */
 		(void)close(pdesin[0]);
 		(void)close(pdesin[1]);
 		(void)close(pdesout[0]);
 		(void)close(pdesout[1]);
 		return (NULL);
 		/* NOTREACHED */
-	case 0:		/* Child. */
+	case 0:				/* Child. */
 		/* Close out un-used sides */
 		(void)close(pdesin[1]);
 		(void)close(pdesout[0]);
@@ -129,8 +132,8 @@ my_popen(const char *command, const char *dir, pid_t *p_pid)
 		(void)close(pdesin[0]);
 		(void)close(pdesout[0]);
 		(void)close(pdesout[1]);
-		return (io_out);
-	} else {
+		return(io_out);
+ 	} else {
 		/* Prepare the input stream */
 		io_in = fdopen(pdesout[0], "r");
 		(void)close(pdesout[1]);
@@ -146,7 +149,7 @@ my_popen(const char *command, const char *dir, pid_t *p_pid)
  *	if already `pclosed', or waitpid returns an error.
  */
 static void
-my_pclose(FILE * io, pid_t the_pid)
+my_pclose(FILE *io, pid_t the_pid)
 {
 	int pstat;
 	pid_t pid;
@@ -164,41 +167,41 @@ my_pclose(FILE * io, pid_t the_pid)
 
 struct counters {
 	struct counters *next_cpu;
-	char counter_name[MAX_NLEN];	/* Name of counter */
-	int cpu;		/* CPU we are on */
-	int pos;		/* Index we are filling to. */
+	char counter_name[MAX_NLEN];		/* Name of counter */
+	int cpu;				/* CPU we are on */
+	int pos;				/* Index we are filling to. */
 	uint64_t vals[MAX_COUNTER_SLOTS];	/* Last 64 entries */
-	uint64_t sum;		/* Summary of entries */
+	uint64_t sum;				/* Summary of entries */
 };
 
 extern struct counters *glob_cpu[MAX_CPU];
 struct counters *glob_cpu[MAX_CPU];
 
 extern struct counters *cnts;
-struct counters *cnts = NULL;
+struct counters *cnts=NULL;
 
 extern int ncnts;
-int ncnts = 0;
+int ncnts=0;
 
-extern int (*expression) (struct counters *, int);
-int (*expression) (struct counters *, int);
+extern int (*expression)(struct counters *, int);
+int (*expression)(struct counters *, int);
 
-static const char *threshold = NULL;
+static const char *threshold=NULL;
 static const char *command;
 
 struct cpu_entry {
 	const char *name;
 	const char *thresh;
 	const char *command;
-	int (*func) (struct counters *, int);
+	int (*func)(struct counters *, int);
+	int counters_required;
 };
 
-
 struct cpu_type {
 	char cputype[32];
 	int number;
 	struct cpu_entry *ents;
-	void (*explain) (const char *name);
+	void (*explain)(const char *name);
 };
 extern struct cpu_type the_cpu;
 struct cpu_type the_cpu;
@@ -207,7 +210,6 @@ static void
 explain_name_sb(const char *name)
 {
 	const char *mythresh;
-
 	if (strcmp(name, "allocstall1") == 0) {
 		printf("Examine PARTIAL_RAT_STALLS.SLOW_LEA_WINDOW / CPU_CLK_UNHALTED.THREAD_P\n");
 		mythresh = "thresh > .05";
@@ -218,10 +220,10 @@ explain_name_sb(const char *name)
 		printf("Examine (20 * BR_MISP_RETIRED.ALL_BRANCHES)/CPU_CLK_UNHALTED.THREAD_P\n");
 		mythresh = "thresh >= .2";
 	} else if (strcmp(name, "splitload") == 0) {
-		printf("Examine MEM_UOP_RETIRED.SPLIT_LOADS * 5) / CPU_CLK_UNHALTED.THREAD_P\n");
+		printf("Examine MEM_UOPS_RETIRED.SPLIT_LOADS * 5) / CPU_CLK_UNHALTED.THREAD_P\n");
 		mythresh = "thresh >= .1";
 	} else if (strcmp(name, "splitstore") == 0) {
-		printf("Examine MEM_UOP_RETIRED.SPLIT_STORES / MEM_UOP_RETIRED.ALL_STORES\n");
+		printf("Examine MEM_UOPS_RETIRED.SPLIT_STORES / MEM_UOPS_RETIRED.ALL_STORES\n");
 		mythresh = "thresh >= .01";
 	} else if (strcmp(name, "contested") == 0) {
 		printf("Examine (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM * 60) / CPU_CLK_UNHALTED.THREAD_P\n");
@@ -279,7 +281,7 @@ explain_name_sb(const char *name)
 	} else {
 		printf("Unknown name:%s\n", name);
 		mythresh = "unknown entry";
-	}
+        }
 	printf("If the value printed is %s we may have the ability to improve performance\n", mythresh);
 }
 
@@ -287,7 +289,6 @@ static void
 explain_name_ib(const char *name)
 {
 	const char *mythresh;
-
 	if (strcmp(name, "br_miss") == 0) {
 		printf("Examine ((BR_MISP_RETIRED.ALL_BRANCHES /(BR_MISP_RETIRED.ALL_BRANCHES +\n");
 		printf("         MACHINE_CLEAR.COUNT) * ((UOPS_ISSUED.ANY - UOPS_RETIRED.RETIRE_SLOTS + 4 * INT_MISC.RECOVERY_CYCLES)\n");
@@ -307,7 +308,7 @@ explain_name_ib(const char *name)
 		mythresh = "thresh >= .2";
 	} else if (strcmp(name, "itlbmiss") == 0) {
 		printf("Examine ITLB_MISSES.WALK_DURATION / CPU_CLK_UNHALTED.THREAD_P\n");
-		mythresh = "thresh > .05";
+		mythresh = "thresh > .05"; 
 	} else if (strcmp(name, "icachemiss") == 0) {
 		printf("Examine (ICACHE.IFETCH_STALL - ITLB_MISSES.WALK_DURATION)/ CPU_CLK_UNHALTED.THREAD_P\n");
 		mythresh = "thresh > .05";
@@ -325,7 +326,7 @@ explain_name_ib(const char *name)
 		printf("         LD_BLOCKS.NO_SR)/CPU_CLK_UNHALTED.THREAD_P\n");
 		mythresh = "thresh >= .1";
 	} else if (strcmp(name, "splitstore") == 0) {
-		printf("Examine MEM_UOP_RETIRED.SPLIT_STORES / MEM_UOP_RETIRED.ALL_STORES\n");
+		printf("Examine MEM_UOPS_RETIRED.SPLIT_STORES / MEM_UOPS_RETIRED.ALL_STORES\n");
 		mythresh = "thresh >= .01";
 	} else if (strcmp(name, "aliasing_4k") == 0) {
 		printf("Examine (LD_BLOCKS_PARTIAL.ADDRESS_ALIAS * 5) / CPU_CLK_UNHALTED.THREAD_P\n");
@@ -362,7 +363,7 @@ explain_name_ib(const char *name)
 	} else {
 		printf("Unknown name:%s\n", name);
 		mythresh = "unknown entry";
-	}
+        }
 	printf("If the value printed is %s we may have the ability to improve performance\n", mythresh);
 }
 
@@ -371,7 +372,6 @@ static void
 explain_name_has(const char *name)
 {
 	const char *mythresh;
-
 	if (strcmp(name, "eff1") == 0) {
 		printf("Examine (UOPS_RETIRED.RETIRE_SLOTS)/(4 *CPU_CLK_UNHALTED.THREAD_P)\n");
 		mythresh = "thresh < .75";
@@ -380,7 +380,7 @@ explain_name_has(const char *name)
 		mythresh = "thresh > 1.0";
 	} else if (strcmp(name, "itlbmiss") == 0) {
 		printf("Examine ITLB_MISSES.WALK_DURATION / CPU_CLK_UNHALTED.THREAD_P\n");
-		mythresh = "thresh > .05";
+		mythresh = "thresh > .05"; 
 	} else if (strcmp(name, "icachemiss") == 0) {
 		printf("Examine (36 * ICACHE.MISSES)/ CPU_CLK_UNHALTED.THREAD_P\n");
 		mythresh = "thresh > .05";
@@ -406,10 +406,10 @@ explain_name_has(const char *name)
 		printf("Examine (LD_BLOCKS_STORE_FORWARD * 13) / CPU_CLK_UNHALTED.THREAD_P\n");
 		mythresh = "thresh >= .05";
 	} else if (strcmp(name, "splitload") == 0) {
-		printf("Examine  (MEM_UOP_RETIRED.SPLIT_LOADS * 5) / CPU_CLK_UNHALTED.THREAD_P\n");
+		printf("Examine  (MEM_UOPS_RETIRED.SPLIT_LOADS * 5) / CPU_CLK_UNHALTED.THREAD_P\n");
 		mythresh = "thresh >= .1";
 	} else if (strcmp(name, "splitstore") == 0) {
-		printf("Examine MEM_UOP_RETIRED.SPLIT_STORES / MEM_UOP_RETIRED.ALL_STORES\n");
+		printf("Examine MEM_UOPS_RETIRED.SPLIT_STORES / MEM_UOPS_RETIRED.ALL_STORES\n");
 		mythresh = "thresh >= .01";
 	} else if (strcmp(name, "aliasing_4k") == 0) {
 		printf("Examine (LD_BLOCKS_PARTIAL.ADDRESS_ALIAS * 5) / CPU_CLK_UNHALTED.THREAD_P\n");
@@ -442,10 +442,12 @@ explain_name_has(const char *name)
 	} else {
 		printf("Unknown name:%s\n", name);
 		mythresh = "unknown entry";
-	}
+        }
 	printf("If the value printed is %s we may have the ability to improve performance\n", mythresh);
 }
 
+
+
 static struct counters *
 find_counter(struct counters *base, const char *name)
 {
@@ -454,16 +456,16 @@ find_counter(struct counters *base, const char *name)
 
 	at = base;
 	len = strlen(name);
-	while (at) {
+	while(at) {
 		if (strncmp(at->counter_name, name, len) == 0) {
-			return (at);
+			return(at);
 		}
 		at = at->next_cpu;
 	}
 	printf("Can't find counter %s\n", name);
 	printf("We have:\n");
 	at = base;
-	while (at) {
+	while(at) {
 		printf("- %s\n", at->counter_name);
 		at = at->next_cpu;
 	}
@@ -478,7 +480,6 @@ allocstall1(struct counters *cpu, int pos)
 	struct counters *partial;
 	struct counters *unhalt;
 	double un, par, res;
-
 	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
 	partial = find_counter(cpu, "PARTIAL_RAT_STALLS.SLOW_LEA_WINDOW");
 	if (pos != -1) {
@@ -488,9 +489,9 @@ allocstall1(struct counters *cpu, int pos)
 		par = partial->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = par / un;
+	res = par/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
@@ -501,7 +502,6 @@ allocstall2(struct counters *cpu, int pos)
 	struct counters *partial;
 	struct counters *unhalt;
 	double un, par, res;
-
 	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
 	partial = find_counter(cpu, "PARTIAL_RAT_STALLS.FLAGS_MERGE_UOP");
 	if (pos != -1) {
@@ -511,9 +511,9 @@ allocstall2(struct counters *cpu, int pos)
 		par = partial->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = par / un;
+	res = par/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
@@ -522,14 +522,12 @@ br_mispredict(struct counters *cpu, int pos)
 	struct counters *brctr;
 	struct counters *unhalt;
 	int ret;
-
 /*  3  - (20 * BR_MISP_RETIRED.ALL_BRANCHES)/CPU_CLK_UNHALTED.THREAD_P (thresh >= .2) */
 	double br, un, con, res;
-
 	con = 20.0;
-
+	
 	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
-	brctr = find_counter(cpu, "BR_MISP_RETIRED.ALL_BRANCHES");
+        brctr = find_counter(cpu, "BR_MISP_RETIRED.ALL_BRANCHES");
 	if (pos != -1) {
 		br = brctr->vals[pos] * 1.0;
 		un = unhalt->vals[pos] * 1.0;
@@ -537,9 +535,9 @@ br_mispredict(struct counters *cpu, int pos)
 		br = brctr->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (con * br) / un;
-	ret = printf("%1.3f", res);
-	return (ret);
+	res = (con * br)/un;
+ 	ret = printf("%1.3f", res);
+	return(ret);
 }
 
 static int
@@ -549,25 +547,22 @@ br_mispredictib(struct counters *cpu, int pos)
 	struct counters *unhalt;
 	struct counters *clear, *clear2, *clear3;
 	struct counters *uops;
-	struct counters *recv;
+	struct counters *recv;	
 	struct counters *iss;
-
 /*	  "pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s BR_MISP_RETIRED.ALL_BRANCHES -s MACHINE_CLEARS.MEMORY_ORDERING -s MACHINE_CLEARS.SMC -s MACHINE_CLEARS.MASKMOV -s UOPS_ISSUED.ANY -s UOPS_RETIRED.RETIRE_SLOTS -s INT_MISC.RECOVERY_CYCLES -w 1",*/
 	int ret;
-
-	/*
-	 * (BR_MISP_RETIRED.ALL_BRANCHES / (BR_MISP_RETIRED.ALL_BRANCHES +
-	 * MACHINE_CLEAR.COUNT) * ((UOPS_ISSUED.ANY -
-	 * UOPS_RETIRED.RETIRE_SLOTS + 4 * INT_MISC.RECOVERY_CYCLES) / (4 *
-	 * CPU_CLK_UNHALTED.THREAD)))
-	 * 
+        /*  
+	 * (BR_MISP_RETIRED.ALL_BRANCHES / 
+	 *         (BR_MISP_RETIRED.ALL_BRANCHES +
+	 *          MACHINE_CLEAR.COUNT) * 
+	 *	   ((UOPS_ISSUED.ANY - UOPS_RETIRED.RETIRE_SLOTS + 4 * INT_MISC.RECOVERY_CYCLES) / (4 * CPU_CLK_UNHALTED.THREAD)))
+	 *
 	 */
 	double br, cl, cl2, cl3, uo, re, un, con, res, is;
-
 	con = 4.0;
-
+	
 	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
-	brctr = find_counter(cpu, "BR_MISP_RETIRED.ALL_BRANCHES");
+        brctr = find_counter(cpu, "BR_MISP_RETIRED.ALL_BRANCHES");
 	clear = find_counter(cpu, "MACHINE_CLEARS.MEMORY_ORDERING");
 	clear2 = find_counter(cpu, "MACHINE_CLEARS.SMC");
 	clear3 = find_counter(cpu, "MACHINE_CLEARS.MASKMOV");
@@ -593,11 +588,12 @@ br_mispredictib(struct counters *cpu, int pos)
 		is = iss->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (br / (br + cl + cl2 + cl3) * ((is - uo + con * re) / (con * un)));
-	ret = printf("%1.3f", res);
-	return (ret);
+	res = (br/(br + cl + cl2 + cl3) * ((is - uo + con * re) / (con * un)));
+ 	ret = printf("%1.3f", res);
+	return(ret);
 }
 
+
 static int
 br_mispredict_broad(struct counters *cpu, int pos)
 {
@@ -611,9 +607,9 @@ br_mispredict_broad(struct counters *cpu, int pos)
 	double br, cl, uo, uo_r, re, con, un, res;
 
 	con = 4.0;
-
+	
 	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
-	brctr = find_counter(cpu, "BR_MISP_RETIRED.ALL_BRANCHES");
+        brctr = find_counter(cpu, "BR_MISP_RETIRED.ALL_BRANCHES");
 	clear = find_counter(cpu, "MACHINE_CLEARS.CYCLES");
 	uops = find_counter(cpu, "UOPS_ISSUED.ANY");
 	uops_ret = find_counter(cpu, "UOPS_RETIRED.RETIRE_SLOTS");
@@ -635,8 +631,8 @@ br_mispredict_broad(struct counters *cpu, int pos)
 		re = recv->sum * 1.0;
 	}
 	res = br / (br + cl) * (uo - uo_r + con * re) / (un * con);
-	ret = printf("%1.3f", res);
-	return (ret);
+ 	ret = printf("%1.3f", res);
+	return(ret);
 }
 
 static int
@@ -647,12 +643,9 @@ splitloadib(struct counters *cpu, int pos)
 	struct counters *l1d, *ldblock;
 	struct counters *unhalt;
 	double un, memd, res, l1, ldb;
-
-	/*
-	 * ((L1D_PEND_MISS.PENDING / MEM_LOAD_UOPS_RETIRED.L1_MISS) *
-	 * LD_BLOCKS.NO_SR) / CPU_CLK_UNHALTED.THREAD_P "pmcstat -s
-	 * CPU_CLK_UNHALTED.THREAD_P -s L1D_PEND_MISS.PENDING -s
-	 * MEM_LOAD_UOPS_RETIRED.L1_MISS -s LD_BLOCKS.NO_SR -w 1",
+        /*  
+	 * ((L1D_PEND_MISS.PENDING / MEM_LOAD_UOPS_RETIRED.L1_MISS) * LD_BLOCKS.NO_SR) / CPU_CLK_UNHALTED.THREAD_P
+	 * "pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s L1D_PEND_MISS.PENDING -s MEM_LOAD_UOPS_RETIRED.L1_MISS -s LD_BLOCKS.NO_SR -w 1",
 	 */
 
 	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
@@ -670,11 +663,12 @@ splitloadib(struct counters *cpu, int pos)
 		ldb = ldblock->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = ((l1 / memd) * ldb) / un;
+	res = ((l1 / memd) * ldb)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
+
 static int
 splitload(struct counters *cpu, int pos)
 {
@@ -682,7 +676,31 @@ splitload(struct counters *cpu, int pos)
 	struct counters *mem;
 	struct counters *unhalt;
 	double con, un, memd, res;
+/*  4  - (MEM_UOPS_RETIRED.SPLIT_LOADS * 5) / CPU_CLK_UNHALTED.THREAD_P (thresh >= .1)*/
 
+	con = 5.0;
+	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
+	mem = find_counter(cpu, "MEM_UOPS_RETIRED.SPLIT_LOADS");
+	if (pos != -1) {
+		memd = mem->vals[pos] * 1.0;
+		un = unhalt->vals[pos] * 1.0;
+	} else {
+		memd = mem->sum * 1.0;
+		un = unhalt->sum * 1.0;
+	}
+	res = (memd * con)/un;
+	ret = printf("%1.3f", res);
+	return(ret);
+}
+
+
+static int
+splitload_sb(struct counters *cpu, int pos)
+{
+	int ret;
+	struct counters *mem;
+	struct counters *unhalt;
+	double con, un, memd, res;
 /*  4  - (MEM_UOP_RETIRED.SPLIT_LOADS * 5) / CPU_CLK_UNHALTED.THREAD_P (thresh >= .1)*/
 
 	con = 5.0;
@@ -695,23 +713,20 @@ splitload(struct counters *cpu, int pos)
 		memd = mem->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (memd * con) / un;
+	res = (memd * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
+
 static int
-splitstore(struct counters *cpu, int pos)
+splitstore_sb(struct counters *cpu, int pos)
 {
-	/*
-	 * 5  - MEM_UOP_RETIRED.SPLIT_STORES / MEM_UOP_RETIRED.ALL_STORES
-	 * (thresh > 0.01)
-	 */
+        /*  5  - MEM_UOP_RETIRED.SPLIT_STORES / MEM_UOP_RETIRED.ALL_STORES (thresh > 0.01) */
 	int ret;
 	struct counters *mem_split;
 	struct counters *mem_stores;
 	double memsplit, memstore, res;
-
 	mem_split = find_counter(cpu, "MEM_UOP_RETIRED.SPLIT_STORES");
 	mem_stores = find_counter(cpu, "MEM_UOP_RETIRED.ALL_STORES");
 	if (pos != -1) {
@@ -721,19 +736,40 @@ splitstore(struct counters *cpu, int pos)
 		memsplit = mem_split->sum * 1.0;
 		memstore = mem_stores->sum * 1.0;
 	}
-	res = memsplit / memstore;
+	res = memsplit/memstore;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
+}
+
+
+
+static int
+splitstore(struct counters *cpu, int pos)
+{
+        /*  5  - MEM_UOPS_RETIRED.SPLIT_STORES / MEM_UOPS_RETIRED.ALL_STORES (thresh > 0.01) */
+	int ret;
+	struct counters *mem_split;
+	struct counters *mem_stores;
+	double memsplit, memstore, res;
+	mem_split = find_counter(cpu, "MEM_UOPS_RETIRED.SPLIT_STORES");
+	mem_stores = find_counter(cpu, "MEM_UOPS_RETIRED.ALL_STORES");
+	if (pos != -1) {
+		memsplit = mem_split->vals[pos] * 1.0;
+		memstore = mem_stores->vals[pos] * 1.0;
+	} else {
+		memsplit = mem_split->sum * 1.0;
+		memstore = mem_stores->sum * 1.0;
+	}
+	res = memsplit/memstore;
+	ret = printf("%1.3f", res);
+	return(ret);
 }
 
 
 static int
 contested(struct counters *cpu, int pos)
 {
-	/*
-	 * 6  - (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM * 60) /
-	 * CPU_CLK_UNHALTED.THREAD_P (thresh >.05)
-	 */
+        /*  6  - (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM * 60) / CPU_CLK_UNHALTED.THREAD_P (thresh >.05) */
 	int ret;
 	struct counters *mem;
 	struct counters *unhalt;
@@ -749,18 +785,15 @@ contested(struct counters *cpu, int pos)
 		memd = mem->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (memd * con) / un;
+	res = (memd * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 contested_has(struct counters *cpu, int pos)
 {
-	/*
-	 * 6  - (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM * 84) /
-	 * CPU_CLK_UNHALTED.THREAD_P (thresh >.05)
-	 */
+        /*  6  - (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM * 84) / CPU_CLK_UNHALTED.THREAD_P (thresh >.05) */
 	int ret;
 	struct counters *mem;
 	struct counters *unhalt;
@@ -776,18 +809,15 @@ contested_has(struct counters *cpu, int pos)
 		memd = mem->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (memd * con) / un;
+	res = (memd * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 contestedbroad(struct counters *cpu, int pos)
 {
-	/*
-	 * 6  - (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM * 84) /
-	 * CPU_CLK_UNHALTED.THREAD_P (thresh >.05)
-	 */
+        /*  6  - (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM * 84) / CPU_CLK_UNHALTED.THREAD_P (thresh >.05) */
 	int ret;
 	struct counters *mem;
 	struct counters *mem2;
@@ -797,7 +827,7 @@ contestedbroad(struct counters *cpu, int pos)
 	con = 84.0;
 	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
 	mem = find_counter(cpu, "MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM");
-	mem2 = find_counter(cpu, "MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_MISS");
+	mem2 = find_counter(cpu,"MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_MISS");
 
 	if (pos != -1) {
 		memd = mem->vals[pos] * 1.0;
@@ -808,19 +838,16 @@ contestedbroad(struct counters *cpu, int pos)
 		memtoo = mem2->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = ((memd * con) + memtoo) / un;
+	res = ((memd * con) + memtoo)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 
 static int
 blockstoreforward(struct counters *cpu, int pos)
 {
-	/*
-	 * 7  - (LD_BLOCKS_STORE_FORWARD * 13) / CPU_CLK_UNHALTED.THREAD_P
-	 * (thresh >= .05)
-	 */
+        /*  7  - (LD_BLOCKS_STORE_FORWARD * 13) / CPU_CLK_UNHALTED.THREAD_P (thresh >= .05)*/
 	int ret;
 	struct counters *ldb;
 	struct counters *unhalt;
@@ -836,19 +863,17 @@ blockstoreforward(struct counters *cpu, int pos)
 		ld = ldb->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (ld * con) / un;
+	res = (ld * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 cache2(struct counters *cpu, int pos)
 {
-	/*
-	 * ** Suspect *** 8  - ((MEM_LOAD_RETIRED.L3_HIT * 26) +
-	 * (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT * 43) +
-	 * (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM * 60)) /
-	 * CPU_CLK_UNHALTED.THREAD_P (thresh >.2)
+	/* ** Suspect ***
+	 *  8  - ((MEM_LOAD_RETIRED.L3_HIT * 26) + (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT * 43) +
+	 *        (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM * 60)) / CPU_CLK_UNHALTED.THREAD_P (thresh >.2)
 	 */
 	int ret;
 	struct counters *mem1, *mem2, *mem3;
@@ -874,17 +899,16 @@ cache2(struct counters *cpu, int pos)
 		me_3 = mem3->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = ((me_1 * con1) + (me_2 * con2) + (me_3 * con3)) / un;
+	res = ((me_1 * con1) + (me_2 * con2) + (me_3 * con3))/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 datasharing(struct counters *cpu, int pos)
 {
-	/*
-	 * (MEM_LOAD_UOPS_L3_HIT_RETIRED.XSNP_HIT * 43)/
-	 * CPU_CLK_UNHALTED.THREAD_P (thresh >.2)
+	/* 
+	 * (MEM_LOAD_UOPS_L3_HIT_RETIRED.XSNP_HIT * 43)/ CPU_CLK_UNHALTED.THREAD_P (thresh >.2)
 	 */
 	int ret;
 	struct counters *mem;
@@ -901,9 +925,9 @@ datasharing(struct counters *cpu, int pos)
 		me = mem->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (me * con) / un;
+	res = (me * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 
 }
 
@@ -911,9 +935,8 @@ datasharing(struct counters *cpu, int pos)
 static int
 datasharing_has(struct counters *cpu, int pos)
 {
-	/*
-	 * (MEM_LOAD_UOPS_L3_HIT_RETIRED.XSNP_HIT * 43)/
-	 * CPU_CLK_UNHALTED.THREAD_P (thresh >.2)
+	/* 
+	 * (MEM_LOAD_UOPS_L3_HIT_RETIRED.XSNP_HIT * 43)/ CPU_CLK_UNHALTED.THREAD_P (thresh >.2)
 	 */
 	int ret;
 	struct counters *mem;
@@ -930,9 +953,9 @@ datasharing_has(struct counters *cpu, int pos)
 		me = mem->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (me * con) / un;
+	res = (me * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 
 }
 
@@ -940,9 +963,8 @@ datasharing_has(struct counters *cpu, int pos)
 static int
 cache2ib(struct counters *cpu, int pos)
 {
-	/*
-	 * (29 * MEM_LOAD_UOPS_RETIRED.LLC_HIT / CPU_CLK_UNHALTED.THREAD_P
-	 * (thresh >.2)
+        /*
+	 *  (29 * MEM_LOAD_UOPS_RETIRED.LLC_HIT / CPU_CLK_UNHALTED.THREAD_P (thresh >.2)
 	 */
 	int ret;
 	struct counters *mem;
@@ -959,9 +981,9 @@ cache2ib(struct counters *cpu, int pos)
 		me = mem->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (con * me) / un;
+	res = (con * me)/un; 
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
@@ -969,9 +991,9 @@ cache2has(struct counters *cpu, int pos)
 {
 	/*
 	 * Examine ((MEM_LOAD_UOPS_RETIRED.LLC_HIT * 36) + \
-	 * (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT * 72) +
-	 * (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM * 84)) /
-	 * CPU_CLK_UNHALTED.THREAD_P
+	 *          (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT * 72) +
+	 *          (MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM * 84))
+	 *           / CPU_CLK_UNHALTED.THREAD_P
 	 */
 	int ret;
 	struct counters *mem1, *mem2, *mem3;
@@ -996,17 +1018,17 @@ cache2has(struct counters *cpu, int pos)
 		me3 = mem3->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = ((me1 * con1) + (me2 * con2) + (me3 * con3)) / un;
+	res = ((me1 * con1) + (me2 * con2) + (me3 * con3))/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
+
 static int
 cache2broad(struct counters *cpu, int pos)
 {
-	/*
-	 * (29 * MEM_LOAD_UOPS_RETIRED.LLC_HIT / CPU_CLK_UNHALTED.THREAD_P
-	 * (thresh >.2)
+        /*
+	 *  (29 * MEM_LOAD_UOPS_RETIRED.LLC_HIT / CPU_CLK_UNHALTED.THREAD_P (thresh >.2)
 	 */
 	int ret;
 	struct counters *mem;
@@ -1023,19 +1045,16 @@ cache2broad(struct counters *cpu, int pos)
 		me = mem->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (con * me) / un;
+	res = (con * me)/un; 
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 
 static int
 cache1(struct counters *cpu, int pos)
 {
-	/*
-	 * 9  - (MEM_LOAD_UOPS_MISC_RETIRED.LLC_MISS * 180) /
-	 * CPU_CLK_UNHALTED.THREAD_P (thresh >= .2)
-	 */
+	/*  9  - (MEM_LOAD_UOPS_MISC_RETIRED.LLC_MISS * 180) / CPU_CLK_UNHALTED.THREAD_P (thresh >= .2) */
 	int ret;
 	struct counters *mem;
 	struct counters *unhalt;
@@ -1051,18 +1070,15 @@ cache1(struct counters *cpu, int pos)
 		me = mem->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (me * con) / un;
+	res = (me * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 cache1ib(struct counters *cpu, int pos)
 {
-	/*
-	 * 9  - (MEM_LOAD_UOPS_L3_MISS_RETIRED.LCOAL_DRAM * 180) /
-	 * CPU_CLK_UNHALTED.THREAD_P (thresh >= .2)
-	 */
+	/*  9  - (MEM_LOAD_UOPS_L3_MISS_RETIRED.LCOAL_DRAM * 180) / CPU_CLK_UNHALTED.THREAD_P (thresh >= .2) */
 	int ret;
 	struct counters *mem;
 	struct counters *unhalt;
@@ -1078,19 +1094,16 @@ cache1ib(struct counters *cpu, int pos)
 		me = mem->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (me * con) / un;
+	res = (me * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 
 static int
 cache1broad(struct counters *cpu, int pos)
 {
-	/*
-	 * 9  - (MEM_LOAD_UOPS_L3_MISS_RETIRED.LCOAL_DRAM * 180) /
-	 * CPU_CLK_UNHALTED.THREAD_P (thresh >= .2)
-	 */
+	/*  9  - (MEM_LOAD_UOPS_L3_MISS_RETIRED.LCOAL_DRAM * 180) / CPU_CLK_UNHALTED.THREAD_P (thresh >= .2) */
 	int ret;
 	struct counters *mem;
 	struct counters *unhalt;
@@ -1106,20 +1119,16 @@ cache1broad(struct counters *cpu, int pos)
 		me = mem->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (me * con) / un;
+	res = (me * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 
 static int
 dtlb_missload(struct counters *cpu, int pos)
 {
-	/*
-	 * 10  - ((DTLB_LOAD_MISSES.STLB_HIT * 7) +
-	 * DTLB_LOAD_MISSES.WALK_DURATION) / CPU_CLK_UNHALTED.THREAD_P (t
-	 * >=.1)
-	 */
+	/* 10  - ((DTLB_LOAD_MISSES.STLB_HIT * 7) + DTLB_LOAD_MISSES.WALK_DURATION) / CPU_CLK_UNHALTED.THREAD_P (t >=.1) */
 	int ret;
 	struct counters *dtlb_m, *dtlb_d;
 	struct counters *unhalt;
@@ -1138,40 +1147,39 @@ dtlb_missload(struct counters *cpu, int pos)
 		d2 = dtlb_d->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = ((d1 * con) + d2) / un;
+	res = ((d1 * con) + d2)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 dtlb_missstore(struct counters *cpu, int pos)
 {
-	/*
-	 * ((DTLB_STORE_MISSES.STLB_HIT * 7) +
-	 * DTLB_STORE_MISSES.WALK_DURATION) / CPU_CLK_UNHALTED.THREAD_P (t
-	 * >= .1)
+        /* 
+	 * ((DTLB_STORE_MISSES.STLB_HIT * 7) + DTLB_STORE_MISSES.WALK_DURATION) / 
+	 * CPU_CLK_UNHALTED.THREAD_P (t >= .1) 
 	 */
-	int ret;
-	struct counters *dtsb_m, *dtsb_d;
-	struct counters *unhalt;
-	double con, un, d1, d2, res;
+        int ret;
+        struct counters *dtsb_m, *dtsb_d;
+        struct counters *unhalt;
+        double con, un, d1, d2, res;
 
-	con = 7.0;
-	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
-	dtsb_m = find_counter(cpu, "DTLB_STORE_MISSES.STLB_HIT");
-	dtsb_d = find_counter(cpu, "DTLB_STORE_MISSES.WALK_DURATION");
-	if (pos != -1) {
-		d1 = dtsb_m->vals[pos] * 1.0;
-		d2 = dtsb_d->vals[pos] * 1.0;
-		un = unhalt->vals[pos] * 1.0;
-	} else {
-		d1 = dtsb_m->sum * 1.0;
-		d2 = dtsb_d->sum * 1.0;
-		un = unhalt->sum * 1.0;
-	}
-	res = ((d1 * con) + d2) / un;
-	ret = printf("%1.3f", res);
-	return (ret);
+        con = 7.0;
+        unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
+        dtsb_m = find_counter(cpu, "DTLB_STORE_MISSES.STLB_HIT");
+        dtsb_d = find_counter(cpu, "DTLB_STORE_MISSES.WALK_DURATION");
+        if (pos != -1) {
+                d1 = dtsb_m->vals[pos] * 1.0;
+                d2 = dtsb_d->vals[pos] * 1.0;
+                un = unhalt->vals[pos] * 1.0;
+        } else {
+                d1 = dtsb_m->sum * 1.0;
+                d2 = dtsb_d->sum * 1.0;
+                un = unhalt->sum * 1.0;
+        }
+        res = ((d1 * con) + d2)/un;
+        ret = printf("%1.3f", res);
+        return(ret);
 }
 
 static int
@@ -1192,19 +1200,16 @@ itlb_miss(struct counters *cpu, int pos)
 		d1 = itlb->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = d1 / un;
+	res = d1/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 
 static int
 itlb_miss_broad(struct counters *cpu, int pos)
 {
-	/*
-	 * (7 * ITLB_MISSES.STLB_HIT_4K + ITLB_MISSES.WALK_DURATION) /
-	 * CPU_CLK_UNTHREAD_P
-	 */
+	/* (7 * ITLB_MISSES.STLB_HIT_4K + ITLB_MISSES.WALK_DURATION) / CPU_CLK_UNTHREAD_P   */
 	int ret;
 	struct counters *itlb;
 	struct counters *unhalt;
@@ -1223,19 +1228,16 @@ itlb_miss_broad(struct counters *cpu, int pos)
 		un = unhalt->sum * 1.0;
 		k = four_k->sum * 1.0;
 	}
-	res = (7.0 * k + d1) / un;
+	res = (7.0 * k + d1)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 
 static int
 icache_miss(struct counters *cpu, int pos)
 {
-	/*
-	 * (ICACHE.IFETCH_STALL - ITLB_MISSES.WALK_DURATION) /
-	 * CPU_CLK_UNHALTED.THREAD_P IB
-	 */
+	/* (ICACHE.IFETCH_STALL - ITLB_MISSES.WALK_DURATION) / CPU_CLK_UNHALTED.THREAD_P IB */
 
 	int ret;
 	struct counters *itlb, *icache;
@@ -1254,9 +1256,9 @@ icache_miss(struct counters *cpu, int pos)
 		ic = icache->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (ic - d1) / un;
+	res = (ic-d1)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 
 }
 
@@ -1280,16 +1282,16 @@ icache_miss_has(struct counters *cpu, int pos)
 		ic = icache->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (con * ic) / un;
+	res = (con * ic)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 
 }
 
 static int
 lcp_stall(struct counters *cpu, int pos)
 {
-	/* ILD_STALL.LCP/CPU_CLK_UNHALTED.THREAD_P IB */
+         /* ILD_STALL.LCP/CPU_CLK_UNHALTED.THREAD_P IB */
 	int ret;
 	struct counters *ild;
 	struct counters *unhalt;
@@ -1304,9 +1306,9 @@ lcp_stall(struct counters *cpu, int pos)
 		d1 = ild->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = d1 / un;
+	res = d1/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 
 }
 
@@ -1314,10 +1316,7 @@ lcp_stall(struct counters *cpu, int pos)
 static int
 frontendstall(struct counters *cpu, int pos)
 {
-	/*
-	 * 12  -  IDQ_UOPS_NOT_DELIVERED.CORE / (CPU_CLK_UNHALTED.THREAD_P *
-	 * 4) (thresh >= .15)
-	 */
+      /* 12  -  IDQ_UOPS_NOT_DELIVERED.CORE / (CPU_CLK_UNHALTED.THREAD_P * 4) (thresh >= .15) */
 	int ret;
 	struct counters *idq;
 	struct counters *unhalt;
@@ -1333,20 +1332,17 @@ frontendstall(struct counters *cpu, int pos)
 		id = idq->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = id / (un * con);
+	res = id/(un * con);
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 clears(struct counters *cpu, int pos)
 {
-	/*
-	 * 13  - ((MACHINE_CLEARS.MEMORY_ORDERING + MACHINE_CLEARS.SMC +
-	 * MACHINE_CLEARS.MASKMOV ) * 100 )  / CPU_CLK_UNHALTED.THREAD_P
-	 * (thresh  >= .02)
-	 */
-
+	/* 13  - ((MACHINE_CLEARS.MEMORY_ORDERING + MACHINE_CLEARS.SMC + MACHINE_CLEARS.MASKMOV ) * 100 )  
+	 *         / CPU_CLK_UNHALTED.THREAD_P (thresh  >= .02)*/
+	
 	int ret;
 	struct counters *clr1, *clr2, *clr3;
 	struct counters *unhalt;
@@ -1357,7 +1353,7 @@ clears(struct counters *cpu, int pos)
 	clr1 = find_counter(cpu, "MACHINE_CLEARS.MEMORY_ORDERING");
 	clr2 = find_counter(cpu, "MACHINE_CLEARS.SMC");
 	clr3 = find_counter(cpu, "MACHINE_CLEARS.MASKMOV");
-
+	
 	if (pos != -1) {
 		cl1 = clr1->vals[pos] * 1.0;
 		cl2 = clr2->vals[pos] * 1.0;
@@ -1369,11 +1365,13 @@ clears(struct counters *cpu, int pos)
 		cl3 = clr3->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = ((cl1 + cl2 + cl3) * con) / un;
+	res = ((cl1 + cl2 + cl3) * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
+
+
 static int
 clears_broad(struct counters *cpu, int pos)
 {
@@ -1402,11 +1400,15 @@ clears_broad(struct counters *cpu, int pos)
 		un = unhalt->sum * 1.0;
 	}
 	/* Formula not listed but extrapulated to add the cy ?? */
-	res = ((cl1 + cl2 + cl3 + cy) * con) / un;
+	res = ((cl1 + cl2 + cl3 + cy) * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
+
+
+
+
 static int
 microassist(struct counters *cpu, int pos)
 {
@@ -1426,11 +1428,12 @@ microassist(struct counters *cpu, int pos)
 		id = idq->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = id / (un * con);
+	res = id/(un * con);
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
+
 static int
 microassist_broad(struct counters *cpu, int pos)
 {
@@ -1457,19 +1460,17 @@ microassist_broad(struct counters *cpu, int pos)
 		uoi = uopiss->sum * 1.0;
 		uor = uopret->sum * 1.0;
 	}
-	res = (uor / uoi) * (id / (un * con));
+	res = (uor/uoi) * (id/(un * con));
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
+
 static int
 aliasing(struct counters *cpu, int pos)
 {
-	/*
-	 * 15  - (LD_BLOCKS_PARTIAL.ADDRESS_ALIAS * 5) /
-	 * CPU_CLK_UNHALTED.THREAD_P (thresh > .1)
-	 */
-	int ret;
+	/* 15  - (LD_BLOCKS_PARTIAL.ADDRESS_ALIAS * 5) / CPU_CLK_UNHALTED.THREAD_P (thresh > .1) */
+	int ret;	
 	struct counters *ld;
 	struct counters *unhalt;
 	double un, lds, con, res;
@@ -1484,19 +1485,16 @@ aliasing(struct counters *cpu, int pos)
 		lds = ld->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (lds * con) / un;
+	res = (lds * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 aliasing_broad(struct counters *cpu, int pos)
 {
-	/*
-	 * 15  - (LD_BLOCKS_PARTIAL.ADDRESS_ALIAS * 5) /
-	 * CPU_CLK_UNHALTED.THREAD_P (thresh > .1)
-	 */
-	int ret;
+	/* 15  - (LD_BLOCKS_PARTIAL.ADDRESS_ALIAS * 5) / CPU_CLK_UNHALTED.THREAD_P (thresh > .1) */
+	int ret;	
 	struct counters *ld;
 	struct counters *unhalt;
 	double un, lds, con, res;
@@ -1511,9 +1509,9 @@ aliasing_broad(struct counters *cpu, int pos)
 		lds = ld->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (lds * con) / un;
+	res = (lds * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 
@@ -1521,7 +1519,7 @@ static int
 fpassists(struct counters *cpu, int pos)
 {
 	/* 16  - FP_ASSIST.ANY/INST_RETIRED.ANY_P */
-	int ret;
+	int ret;	
 	struct counters *fp;
 	struct counters *inst;
 	double un, fpd, res;
@@ -1535,19 +1533,16 @@ fpassists(struct counters *cpu, int pos)
 		fpd = fp->sum * 1.0;
 		un = inst->sum * 1.0;
 	}
-	res = fpd / un;
+	res = fpd/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 otherassistavx(struct counters *cpu, int pos)
 {
-	/*
-	 * 17  - (OTHER_ASSISTS.AVX_TO_SSE * 75)/CPU_CLK_UNHALTED.THREAD_P
-	 * thresh  .1
-	 */
-	int ret;
+	/* 17  - (OTHER_ASSISTS.AVX_TO_SSE * 75)/CPU_CLK_UNHALTED.THREAD_P thresh  .1*/
+	int ret;	
 	struct counters *oth;
 	struct counters *unhalt;
 	double un, ot, con, res;
@@ -1562,24 +1557,21 @@ otherassistavx(struct counters *cpu, int pos)
 		ot = oth->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (ot * con) / un;
+	res = (ot * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 otherassistsse(struct counters *cpu, int pos)
 {
 
-	int ret;
+	int ret;	
 	struct counters *oth;
 	struct counters *unhalt;
 	double un, ot, con, res;
 
-	/*
-	 * 18     (OTHER_ASSISTS.SSE_TO_AVX * 75)/CPU_CLK_UNHALTED.THREAD_P
-	 * thresh .1
-	 */
+	/* 18     (OTHER_ASSISTS.SSE_TO_AVX * 75)/CPU_CLK_UNHALTED.THREAD_P  thresh .1*/
 	con = 75.0;
 	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
 	oth = find_counter(cpu, "OTHER_ASSISTS.SSE_TO_AVX");
@@ -1590,24 +1582,21 @@ otherassistsse(struct counters *cpu, int pos)
 		ot = oth->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = (ot * con) / un;
+	res = (ot * con)/un;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 efficiency1(struct counters *cpu, int pos)
 {
 
-	int ret;
+	int ret;	
 	struct counters *uops;
 	struct counters *unhalt;
 	double un, ot, con, res;
 
-	/*
-	 * 19 (UOPS_RETIRED.RETIRE_SLOTS/(4*CPU_CLK_UNHALTED.THREAD_P) look
-	 * if thresh < .9
-	 */
+        /* 19 (UOPS_RETIRED.RETIRE_SLOTS/(4*CPU_CLK_UNHALTED.THREAD_P) look if thresh < .9*/
 	con = 4.0;
 	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
 	uops = find_counter(cpu, "UOPS_RETIRED.RETIRE_SLOTS");
@@ -1618,24 +1607,21 @@ efficiency1(struct counters *cpu, int pos)
 		ot = uops->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = ot / (con * un);
+	res = ot/(con * un);
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 static int
 efficiency2(struct counters *cpu, int pos)
 {
 
-	int ret;
+	int ret;	
 	struct counters *uops;
 	struct counters *unhalt;
 	double un, ot, res;
 
-	/*
-	 * 20  - CPU_CLK_UNHALTED.THREAD_P/INST_RETIRED.ANY_P good if > 1.
-	 * (comp factor)
-	 */
+        /* 20  - CPU_CLK_UNHALTED.THREAD_P/INST_RETIRED.ANY_P good if > 1. (comp factor)*/
 	unhalt = find_counter(cpu, "CPU_CLK_UNHALTED.THREAD_P");
 	uops = find_counter(cpu, "INST_RETIRED.ANY_P");
 	if (pos != -1) {
@@ -1645,205 +1631,206 @@ efficiency2(struct counters *cpu, int pos)
 		ot = uops->sum * 1.0;
 		un = unhalt->sum * 1.0;
 	}
-	res = un / ot;
+	res = un/ot;
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
-#define SANDY_BRIDGE_COUNT 20
+#define SANDY_BRIDGE_COUNT 20	
 static struct cpu_entry sandy_bridge[SANDY_BRIDGE_COUNT] = {
-	 /*01*/ {"allocstall1", "thresh > .05",
-		"pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s PARTIAL_RAT_STALLS.SLOW_LEA_WINDOW -w 1",
-	allocstall1},
-	 /*02*/ {"allocstall2", "thresh > .05",
-		"pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s PARTIAL_RAT_STALLS.FLAGS_MERGE_UOP_CYCLES -w 1",
-	allocstall2},
-	 /*03*/ {"br_miss", "thresh >= .2",
-		"pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s BR_MISP_RETIRED.ALL_BRANCHES -w 1",
-	br_mispredict},
-	 /*04*/ {"splitload", "thresh >= .1",
-		"pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s MEM_UOP_RETIRED.SPLIT_LOADS -w 1",
-	splitload},
-	 /*05*/ {"splitstore", "thresh >= .01",
-		"pmcstat -s MEM_UOP_RETIRED.SPLIT_STORES -s MEM_UOP_RETIRED.ALL_STORES -w 1",
-	splitstore},
-	 /*06*/ {"contested", "thresh >= .05",
-		"pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM  -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	contested},
-	 /*07*/ {"blockstorefwd", "thresh >= .05",
-		"pmcstat -s LD_BLOCKS_STORE_FORWARD -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	blockstoreforward},
-	 /*08*/ {"cache2", "thresh >= .2",
-		"pmcstat -s MEM_LOAD_UOPS_RETIRED.LLC_HIT -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	cache2},
-	 /*09*/ {"cache1", "thresh >= .2",
-		"pmcstat -s MEM_LOAD_UOPS_MISC_RETIRED.LLC_MISS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	cache1},
-	 /*10*/ {"dtlbmissload", "thresh >= .1",
-		"pmcstat -s DTLB_LOAD_MISSES.STLB_HIT -s DTLB_LOAD_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	dtlb_missload},
-	 /*11*/ {"dtlbmissstore", "thresh >= .05",
-		"pmcstat -s DTLB_STORE_MISSES.STLB_HIT -s DTLB_STORE_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	dtlb_missstore},
-	 /*12*/ {"frontendstall", "thresh >= .15",
-		"pmcstat -s IDQ_UOPS_NOT_DELIVERED.CORE -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	frontendstall},
-	 /*13*/ {"clears", "thresh >= .02",
-		"pmcstat -s MACHINE_CLEARS.MEMORY_ORDERING -s MACHINE_CLEARS.SMC -s MACHINE_CLEARS.MASKMOV -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	clears},
-	 /*14*/ {"microassist", "thresh >= .05",
-		"pmcstat -s IDQ.MS_UOPS,cmask=1 -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	microassist},
-	 /*15*/ {"aliasing_4k", "thresh >= .1",
-		"pmcstat -s LD_BLOCKS_PARTIAL.ADDRESS_ALIAS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	aliasing},
-	 /*16*/ {"fpassist", "look for a excessive value",
-		"pmcstat -s FP_ASSIST.ANY -s INST_RETIRED.ANY_P -w 1",
-	fpassists},
-	 /*17*/ {"otherassistavx", "look for a excessive value",
-		"pmcstat -s OTHER_ASSISTS.AVX_TO_SSE -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	otherassistavx},
-	 /*18*/ {"otherassistsse", "look for a excessive value",
-		"pmcstat -s OTHER_ASSISTS.SSE_TO_AVX -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	otherassistsse},
-	 /*19*/ {"eff1", "thresh < .9",
-		"pmcstat -s UOPS_RETIRED.RETIRE_SLOTS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	efficiency1},
-	 /*20*/ {"eff2", "thresh > 1.0",
-		"pmcstat -s INST_RETIRED.ANY_P -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	efficiency2},
+/*01*/	{ "allocstall1", "thresh > .05", 
+	  "pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s PARTIAL_RAT_STALLS.SLOW_LEA_WINDOW -w 1",
+	  allocstall1, 2 },
+/* -- not defined for SB right (partial-rat_stalls) 02*/
+        { "allocstall2", "thresh > .05", 
+	  "pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s PARTIAL_RAT_STALLS.FLAGS_MERGE_UOP -w 1",
+	  allocstall2, 2 },
+/*03*/	{ "br_miss", "thresh >= .2", 
+	  "pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s BR_MISP_RETIRED.ALL_BRANCHES -w 1",
+	  br_mispredict, 2 },
+/*04*/	{ "splitload", "thresh >= .1", 
+	  "pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s MEM_UOP_RETIRED.SPLIT_LOADS -w 1",
+	  splitload_sb, 2 },
+/* 05*/	{ "splitstore", "thresh >= .01", 
+	  "pmcstat -s MEM_UOP_RETIRED.SPLIT_STORES -s MEM_UOP_RETIRED.ALL_STORES -w 1",
+	  splitstore_sb, 2 },
+/*06*/	{ "contested", "thresh >= .05", 
+	  "pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM  -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  contested, 2 },
+/*07*/	{ "blockstorefwd", "thresh >= .05", 
+	  "pmcstat -s LD_BLOCKS_STORE_FORWARD -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  blockstoreforward, 2 },
+/*08*/	{ "cache2", "thresh >= .2", 
+	  "pmcstat -s MEM_LOAD_UOPS_RETIRED.LLC_HIT -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  cache2, 4 },
+/*09*/	{ "cache1", "thresh >= .2", 
+	  "pmcstat -s MEM_LOAD_UOPS_MISC_RETIRED.LLC_MISS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  cache1, 2 },
+/*10*/	{ "dtlbmissload", "thresh >= .1", 
+	  "pmcstat -s DTLB_LOAD_MISSES.STLB_HIT -s DTLB_LOAD_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  dtlb_missload, 3 },
+/*11*/	{ "dtlbmissstore", "thresh >= .05", 
+	  "pmcstat -s DTLB_STORE_MISSES.STLB_HIT -s DTLB_STORE_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  dtlb_missstore, 3 },
+/*12*/	{ "frontendstall", "thresh >= .15", 
+	  "pmcstat -s IDQ_UOPS_NOT_DELIVERED.CORE -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  frontendstall, 2 },
+/*13*/	{ "clears", "thresh >= .02", 
+	  "pmcstat -s MACHINE_CLEARS.MEMORY_ORDERING -s MACHINE_CLEARS.SMC -s MACHINE_CLEARS.MASKMOV -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  clears, 4 },
+/*14*/	{ "microassist", "thresh >= .05", 
+	  "pmcstat -s IDQ.MS_UOPS,cmask=1 -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  microassist, 2 },
+/*15*/	{ "aliasing_4k", "thresh >= .1", 
+	  "pmcstat -s LD_BLOCKS_PARTIAL.ADDRESS_ALIAS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  aliasing, 2 },
+/*16*/	{ "fpassist", "look for a excessive value", 
+	  "pmcstat -s FP_ASSIST.ANY -s INST_RETIRED.ANY_P -w 1",
+	  fpassists, 2 },
+/*17*/	{ "otherassistavx", "look for a excessive value", 
+	  "pmcstat -s OTHER_ASSISTS.AVX_TO_SSE -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  otherassistavx, 2},
+/*18*/	{ "otherassistsse", "look for a excessive value", 
+	  "pmcstat -s OTHER_ASSISTS.SSE_TO_AVX -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  otherassistsse, 2 },
+/*19*/	{ "eff1", "thresh < .9", 
+	  "pmcstat -s UOPS_RETIRED.RETIRE_SLOTS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  efficiency1, 2 },
+/*20*/	{ "eff2", "thresh > 1.0", 
+	  "pmcstat -s INST_RETIRED.ANY_P -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  efficiency2, 2 },
 };
 
 
 #define IVY_BRIDGE_COUNT 21
 static struct cpu_entry ivy_bridge[IVY_BRIDGE_COUNT] = {
-	 /*1*/ {"eff1", "thresh < .75",
-		"pmcstat -s UOPS_RETIRED.RETIRE_SLOTS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	efficiency1},
-	 /*2*/ {"eff2", "thresh > 1.0",
-		"pmcstat -s INST_RETIRED.ANY_P -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	efficiency2},
-	 /*3*/ {"itlbmiss", "thresh > .05",
-		"pmcstat -s ITLB_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	itlb_miss},
-	 /*4*/ {"icachemiss", "thresh > .05",
-		"pmcstat -s ICACHE.IFETCH_STALL -s ITLB_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	icache_miss},
-	 /*5*/ {"lcpstall", "thresh > .05",
-		"pmcstat -s ILD_STALL.LCP -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	lcp_stall},
-	 /*6*/ {"cache1", "thresh >= .2",
-		"pmcstat -s MEM_LOAD_UOPS_LLC_MISS_RETIRED.LOCAL_DRAM -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	cache1ib},
-	 /*7*/ {"cache2", "thresh >= .2",
-		"pmcstat -s MEM_LOAD_UOPS_RETIRED.LLC_HIT -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	cache2ib},
-	 /*8*/ {"contested", "thresh >= .05",
-		"pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM  -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	contested},
-	 /*9*/ {"datashare", "thresh >= .05",
-		"pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	datasharing},
-	 /*10*/ {"blockstorefwd", "thresh >= .05",
-		"pmcstat -s LD_BLOCKS_STORE_FORWARD -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	blockstoreforward},
-	 /*11*/ {"splitload", "thresh >= .1",
-		"pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s L1D_PEND_MISS.PENDING -s MEM_LOAD_UOPS_RETIRED.L1_MISS -s LD_BLOCKS.NO_SR -w 1",
-	splitloadib},
-	 /*12*/ {"splitstore", "thresh >= .01",
-		"pmcstat -s MEM_UOP_RETIRED.SPLIT_STORES -s MEM_UOP_RETIRED.ALL_STORES -w 1",
-	splitstore},
-	 /*13*/ {"aliasing_4k", "thresh >= .1",
-		"pmcstat -s LD_BLOCKS_PARTIAL.ADDRESS_ALIAS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	aliasing},
-	 /*14*/ {"dtlbmissload", "thresh >= .1",
-		"pmcstat -s DTLB_LOAD_MISSES.STLB_HIT -s DTLB_LOAD_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	dtlb_missload},
-	 /*15*/ {"dtlbmissstore", "thresh >= .05",
-		"pmcstat -s DTLB_STORE_MISSES.STLB_HIT -s DTLB_STORE_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	dtlb_missstore},
-	 /*16*/ {"br_miss", "thresh >= .2",
-		"pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s BR_MISP_RETIRED.ALL_BRANCHES -s MACHINE_CLEARS.MEMORY_ORDERING -s MACHINE_CLEARS.SMC -s MACHINE_CLEARS.MASKMOV -s UOPS_ISSUED.ANY -s UOPS_RETIRED.RETIRE_SLOTS -s INT_MISC.RECOVERY_CYCLES -w 1",
-	br_mispredictib},
-	 /*17*/ {"clears", "thresh >= .02",
-		"pmcstat -s MACHINE_CLEARS.MEMORY_ORDERING -s MACHINE_CLEARS.SMC -s MACHINE_CLEARS.MASKMOV -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	clears},
-	 /*18*/ {"microassist", "thresh >= .05",
-		"pmcstat -s IDQ.MS_UOPS,cmask=1 -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	microassist},
-	 /*19*/ {"fpassist", "look for a excessive value",
-		"pmcstat -s FP_ASSIST.ANY -s INST_RETIRED.ANY_P -w 1",
-	fpassists},
-	 /*20*/ {"otherassistavx", "look for a excessive value",
-		"pmcstat -s OTHER_ASSISTS.AVX_TO_SSE -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	otherassistavx},
-	 /*21*/ {"otherassistsse", "look for a excessive value",
-		"pmcstat -s OTHER_ASSISTS.SSE_TO_AVX -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	otherassistsse},
+/*1*/	{ "eff1", "thresh < .75", 
+	  "pmcstat -s UOPS_RETIRED.RETIRE_SLOTS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  efficiency1, 2 },
+/*2*/	{ "eff2", "thresh > 1.0", 
+	  "pmcstat -s INST_RETIRED.ANY_P -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  efficiency2, 2 },
+/*3*/	{ "itlbmiss", "thresh > .05", 
+	  "pmcstat -s ITLB_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  itlb_miss, 2 },
+/*4*/	{ "icachemiss", "thresh > .05", 
+	  "pmcstat -s ICACHE.IFETCH_STALL -s ITLB_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  icache_miss, 3 },
+/*5*/	{ "lcpstall", "thresh > .05", 
+	  "pmcstat -s ILD_STALL.LCP -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  lcp_stall, 2 },
+/*6*/	{ "cache1", "thresh >= .2", 
+	  "pmcstat -s MEM_LOAD_UOPS_LLC_MISS_RETIRED.LOCAL_DRAM -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  cache1ib, 2 },
+/*7*/	{ "cache2", "thresh >= .2", 
+	  "pmcstat -s MEM_LOAD_UOPS_RETIRED.LLC_HIT -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  cache2ib, 2 },
+/*8*/	{ "contested", "thresh >= .05", 
+	  "pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM  -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  contested, 2 },
+/*9*/	{ "datashare", "thresh >= .05",
+	  "pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  datasharing, 2 },
+/*10*/	{ "blockstorefwd", "thresh >= .05", 
+	  "pmcstat -s LD_BLOCKS_STORE_FORWARD -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  blockstoreforward, 2 },
+/*11*/	{ "splitload", "thresh >= .1", 
+	  "pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s L1D_PEND_MISS.PENDING -s MEM_LOAD_UOPS_RETIRED.L1_MISS -s LD_BLOCKS.NO_SR -w 1",
+	  splitloadib, 4 },
+/*12*/	{ "splitstore", "thresh >= .01", 
+	  "pmcstat -s MEM_UOPS_RETIRED.SPLIT_STORES -s MEM_UOPS_RETIRED.ALL_STORES -w 1",
+	  splitstore, 2 },
+/*13*/	{ "aliasing_4k", "thresh >= .1", 
+	  "pmcstat -s LD_BLOCKS_PARTIAL.ADDRESS_ALIAS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  aliasing, 2 },
+/*14*/	{ "dtlbmissload", "thresh >= .1", 
+	  "pmcstat -s DTLB_LOAD_MISSES.STLB_HIT -s DTLB_LOAD_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  dtlb_missload , 3},
+/*15*/	{ "dtlbmissstore", "thresh >= .05", 
+	  "pmcstat -s DTLB_STORE_MISSES.STLB_HIT -s DTLB_STORE_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  dtlb_missstore, 3 },
+/*16*/	{ "br_miss", "thresh >= .2", 
+	  "pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s BR_MISP_RETIRED.ALL_BRANCHES -s MACHINE_CLEARS.MEMORY_ORDERING -s MACHINE_CLEARS.SMC -s MACHINE_CLEARS.MASKMOV -s UOPS_ISSUED.ANY -s UOPS_RETIRED.RETIRE_SLOTS -s INT_MISC.RECOVERY_CYCLES -w 1",
+	  br_mispredictib, 8 },
+/*17*/	{ "clears", "thresh >= .02", 
+	  "pmcstat -s MACHINE_CLEARS.MEMORY_ORDERING -s MACHINE_CLEARS.SMC -s MACHINE_CLEARS.MASKMOV -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  clears, 4 },
+/*18*/	{ "microassist", "thresh >= .05", 
+	  "pmcstat -s IDQ.MS_UOPS,cmask=1 -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  microassist, 2 },
+/*19*/	{ "fpassist", "look for a excessive value", 
+	  "pmcstat -s FP_ASSIST.ANY -s INST_RETIRED.ANY_P -w 1",
+	  fpassists, 2 },
+/*20*/	{ "otherassistavx", "look for a excessive value", 
+	  "pmcstat -s OTHER_ASSISTS.AVX_TO_SSE -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  otherassistavx , 2},
+/*21*/	{ "otherassistsse", "look for a excessive value", 
+	  "pmcstat -s OTHER_ASSISTS.SSE_TO_AVX -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  otherassistsse, 2 },
 };
 
 #define HASWELL_COUNT 20
 static struct cpu_entry haswell[HASWELL_COUNT] = {
-	 /*1*/ {"eff1", "thresh < .75",
-		"pmcstat -s UOPS_RETIRED.RETIRE_SLOTS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	efficiency1},
-	 /*2*/ {"eff2", "thresh > 1.0",
-		"pmcstat -s INST_RETIRED.ANY_P -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	efficiency2},
-	 /*3*/ {"itlbmiss", "thresh > .05",
-		"pmcstat -s ITLB_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	itlb_miss},
-	 /*4*/ {"icachemiss", "thresh > .05",
-		"pmcstat -s ICACHE.MISSES --s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	icache_miss_has},
-	 /*5*/ {"lcpstall", "thresh > .05",
-		"pmcstat -s ILD_STALL.LCP -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	lcp_stall},
-	 /*6*/ {"cache1", "thresh >= .2",
-		"pmcstat -s MEM_LOAD_UOPS_LLC_MISS_RETIRED.LOCAL_DRAM -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	cache1ib},
-	 /*7*/ {"cache2", "thresh >= .2",
-		"pmcstat -s MEM_LOAD_UOPS_RETIRED.LLC_HIT  -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	cache2has},
-	 /*8*/ {"contested", "thresh >= .05",
-		"pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM  -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	contested_has},
-	 /*9*/ {"datashare", "thresh >= .05",
-		"pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	datasharing_has},
-	 /*10*/ {"blockstorefwd", "thresh >= .05",
-		"pmcstat -s LD_BLOCKS_STORE_FORWARD -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	blockstoreforward},
-	 /*11*/ {"splitload", "thresh >= .1",
-		"pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s MEM_UOP_RETIRED.SPLIT_LOADS -w 1",
-	splitload},
-	 /*12*/ {"splitstore", "thresh >= .01",
-		"pmcstat -s MEM_UOP_RETIRED.SPLIT_STORES -s MEM_UOP_RETIRED.ALL_STORES -w 1",
-	splitstore},
-	 /*13*/ {"aliasing_4k", "thresh >= .1",
-		"pmcstat -s LD_BLOCKS_PARTIAL.ADDRESS_ALIAS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	aliasing},
-	 /*14*/ {"dtlbmissload", "thresh >= .1",
-		"pmcstat -s DTLB_LOAD_MISSES.STLB_HIT -s DTLB_LOAD_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	dtlb_missload},
-	 /*15*/ {"br_miss", "thresh >= .2",
-		"pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s BR_MISP_RETIRED.ALL_BRANCHES -w 1",
-	br_mispredict},
-	 /*16*/ {"clears", "thresh >= .02",
-		"pmcstat -s MACHINE_CLEARS.MEMORY_ORDERING -s MACHINE_CLEARS.SMC -s MACHINE_CLEARS.MASKMOV -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	clears},
-	 /*17*/ {"microassist", "thresh >= .05",
-		"pmcstat -s IDQ.MS_UOPS,cmask=1 -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	microassist},
-	 /*18*/ {"fpassist", "look for a excessive value",
-		"pmcstat -s FP_ASSIST.ANY -s INST_RETIRED.ANY_P -w 1",
-	fpassists},
-	 /*19*/ {"otherassistavx", "look for a excessive value",
-		"pmcstat -s OTHER_ASSISTS.AVX_TO_SSE -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	otherassistavx},
-	 /*20*/ {"otherassistsse", "look for a excessive value",
-		"pmcstat -s OTHER_ASSISTS.SSE_TO_AVX -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	otherassistsse},
+/*1*/	{ "eff1", "thresh < .75", 
+	  "pmcstat -s UOPS_RETIRED.RETIRE_SLOTS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  efficiency1, 2 },
+/*2*/	{ "eff2", "thresh > 1.0", 
+	  "pmcstat -s INST_RETIRED.ANY_P -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  efficiency2, 2 },
+/*3*/	{ "itlbmiss", "thresh > .05", 
+	  "pmcstat -s ITLB_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  itlb_miss, 2 },
+/*4*/	{ "icachemiss", "thresh > .05", 
+	  "pmcstat -s ICACHE.MISSES -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  icache_miss_has, 2 },
+/*5*/	{ "lcpstall", "thresh > .05", 
+	  "pmcstat -s ILD_STALL.LCP -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  lcp_stall, 2 },
+/*6*/	{ "cache1", "thresh >= .2", 
+	  "pmcstat -s MEM_LOAD_UOPS_LLC_MISS_RETIRED.LOCAL_DRAM -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  cache1ib, 2 },
+/*7*/	{ "cache2", "thresh >= .2", 
+	  "pmcstat -s MEM_LOAD_UOPS_RETIRED.LLC_HIT  -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  cache2has, 4 },
+/*8*/	{ "contested", "thresh >= .05", 
+	  "pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM  -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  contested_has, 2 },
+/*9*/	{ "datashare", "thresh >= .05",
+	  "pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  datasharing_has, 2 },
+/*10*/	{ "blockstorefwd", "thresh >= .05", 
+	  "pmcstat -s LD_BLOCKS_STORE_FORWARD -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  blockstoreforward, 2 },
+/*11*/	{ "splitload", "thresh >= .1", 
+	  "pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s MEM_UOPS_RETIRED.SPLIT_LOADS -w 1",
+	  splitload , 2},
+/*12*/	{ "splitstore", "thresh >= .01", 
+	  "pmcstat -s MEM_UOPS_RETIRED.SPLIT_STORES -s MEM_UOPS_RETIRED.ALL_STORES -w 1",
+	  splitstore, 2 },
+/*13*/	{ "aliasing_4k", "thresh >= .1", 
+	  "pmcstat -s LD_BLOCKS_PARTIAL.ADDRESS_ALIAS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  aliasing, 2 },
+/*14*/	{ "dtlbmissload", "thresh >= .1", 
+	  "pmcstat -s DTLB_LOAD_MISSES.STLB_HIT -s DTLB_LOAD_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  dtlb_missload, 3 },
+/*15*/	{ "br_miss", "thresh >= .2", 
+	  "pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s BR_MISP_RETIRED.ALL_BRANCHES -w 1",
+	  br_mispredict, 2 },
+/*16*/	{ "clears", "thresh >= .02", 
+	  "pmcstat -s MACHINE_CLEARS.MEMORY_ORDERING -s MACHINE_CLEARS.SMC -s MACHINE_CLEARS.MASKMOV -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  clears, 4 },
+/*17*/	{ "microassist", "thresh >= .05", 
+	  "pmcstat -s IDQ.MS_UOPS,cmask=1 -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  microassist, 2 },
+/*18*/	{ "fpassist", "look for a excessive value", 
+	  "pmcstat -s FP_ASSIST.ANY -s INST_RETIRED.ANY_P -w 1",
+	  fpassists, 2 },
+/*19*/	{ "otherassistavx", "look for a excessive value", 
+	  "pmcstat -s OTHER_ASSISTS.AVX_TO_SSE -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  otherassistavx, 2 },
+/*20*/	{ "otherassistsse", "look for a excessive value", 
+	  "pmcstat -s OTHER_ASSISTS.SSE_TO_AVX -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  otherassistsse, 2 },
 };
 
 
@@ -1851,7 +1838,6 @@ static void
 explain_name_broad(const char *name)
 {
 	const char *mythresh;
-
 	if (strcmp(name, "eff1") == 0) {
 		printf("Examine (UOPS_RETIRED.RETIRE_SLOTS)/(4 *CPU_CLK_UNHALTED.THREAD_P)\n");
 		mythresh = "thresh < .75";
@@ -1860,9 +1846,9 @@ explain_name_broad(const char *name)
 		mythresh = "thresh > 1.0";
 	} else if (strcmp(name, "itlbmiss") == 0) {
 		printf("Examine (7 * ITLB_MISSES_STLB_HIT_4K + ITLB_MISSES.WALK_DURATION)/ CPU_CLK_UNHALTED.THREAD_P\n");
-		mythresh = "thresh > .05";
+		mythresh = "thresh > .05"; 
 	} else if (strcmp(name, "icachemiss") == 0) {
-		printf("Examine ( 36.0 * CACHE.MISSES)/ CPU_CLK_UNHALTED.THREAD_P ??? may not be right \n");
+		printf("Examine ( 36.0 * ICACHE.MISSES)/ CPU_CLK_UNHALTED.THREAD_P ??? may not be right \n");
 		mythresh = "thresh > .05";
 	} else if (strcmp(name, "lcpstall") == 0) {
 		printf("Examine ILD_STALL.LCP/CPU_CLK_UNHALTED.THREAD_P\n");
@@ -1913,66 +1899,67 @@ explain_name_broad(const char *name)
 	} else {
 		printf("Unknown name:%s\n", name);
 		mythresh = "unknown entry";
-	}
+        }
 	printf("If the value printed is %s we may have the ability to improve performance\n", mythresh);
 }
 
 
 #define BROADWELL_COUNT 17
 static struct cpu_entry broadwell[BROADWELL_COUNT] = {
-	 /*1*/ {"eff1", "thresh < .75",
-		"pmcstat -s UOPS_RETIRED.RETIRE_SLOTS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	efficiency1},
-	 /*2*/ {"eff2", "thresh > 1.0",
-		"pmcstat -s INST_RETIRED.ANY_P -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	efficiency2},
-	 /*3*/ {"itlbmiss", "thresh > .05",
-		"pmcstat -s ITLB_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -s ITLB_MISSES.STLB_HIT_4K -w 1",
-	itlb_miss_broad},
-	 /*4*/ {"icachemiss", "thresh > .05",
-		"pmcstat -s ICACHE.MISSES --s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	icache_miss_has},
-	 /*5*/ {"lcpstall", "thresh > .05",
-		"pmcstat -s ILD_STALL.LCP -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	lcp_stall},
-	 /*6*/ {"cache1", "thresh >= .1",
-		"pmcstat -s MEM_LOAD_UOPS_RETIRED.L3_MISS  -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	cache1broad},
-	 /*7*/ {"cache2", "thresh >= .2",
-		"pmcstat -s MEM_LOAD_UOPS_RETIRED.L3_HIT  -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	cache2broad},
-	 /*8*/ {"contested", "thresh >= .05",
-		"pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM  -s CPU_CLK_UNHALTED.THREAD_P  -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_MISS -w 1",
-	contestedbroad},
-	 /*9*/ {"datashare", "thresh >= .05",
-		"pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	datasharing_has},
-	 /*10*/ {"blockstorefwd", "thresh >= .05",
-		"pmcstat -s LD_BLOCKS_STORE_FORWARD -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	blockstoreforward},
-	 /*11*/ {"aliasing_4k", "thresh >= .1",
-		"pmcstat -s LD_BLOCKS_PARTIAL.ADDRESS_ALIAS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	aliasing_broad},
-	 /*12*/ {"dtlbmissload", "thresh >= .1",
-		"pmcstat -s DTLB_LOAD_MISSES.STLB_HIT -s DTLB_LOAD_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	dtlb_missload},
-	 /*13*/ {"br_miss", "thresh >= .2",
-		"pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s BR_MISP_RETIRED.ALL_BRANCHES -s MACHINE_CLEARS.CYCLES -s UOPS_ISSUED.ANY -s UOPS_RETIRED.RETIRE_SLOTS -s INT_MISC.RECOVERY_CYCLES -w 1",
-	br_mispredict_broad},
-	 /*14*/ {"clears", "thresh >= .02",
-		"pmcstat -s MACHINE_CLEARS.CYCLES -s MACHINE_CLEARS.MEMORY_ORDERING -s MACHINE_CLEARS.SMC -s MACHINE_CLEARS.MASKMOV -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	clears_broad},
-	 /*15*/ {"fpassist", "look for a excessive value",
-		"pmcstat -s FP_ASSIST.ANY -s INST_RETIRED.ANY_P -w 1",
-	fpassists},
-	 /*16*/ {"otherassistavx", "look for a excessive value",
-		"pmcstat -s OTHER_ASSISTS.AVX_TO_SSE -s CPU_CLK_UNHALTED.THREAD_P -w 1",
-	otherassistavx},
-	 /*17*/ {"microassist", "thresh >= .2",
-		"pmcstat -s IDQ.MS_UOPS,cmask=1 -s CPU_CLK_UNHALTED.THREAD_P -s UOPS_ISSUED.ANY -s UOPS_RETIRED.RETIRE_SLOTS  -w 1",
-	microassist_broad},
+/*1*/	{ "eff1", "thresh < .75", 
+	  "pmcstat -s UOPS_RETIRED.RETIRE_SLOTS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  efficiency1, 2 }, 
+/*2*/	{ "eff2", "thresh > 1.0", 
+	  "pmcstat -s INST_RETIRED.ANY_P -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  efficiency2, 2 },
+/*3*/	{ "itlbmiss", "thresh > .05", 
+	  "pmcstat -s ITLB_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -s ITLB_MISSES.STLB_HIT_4K -w 1",
+	  itlb_miss_broad, 3 },
+/*4*/	{ "icachemiss", "thresh > .05", 
+	  "pmcstat -s ICACHE.MISSES -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  icache_miss_has, 2 },
+/*5*/	{ "lcpstall", "thresh > .05", 
+	  "pmcstat -s ILD_STALL.LCP -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  lcp_stall, 2 },
+/*6*/	{ "cache1", "thresh >= .1", 
+	  "pmcstat -s MEM_LOAD_UOPS_RETIRED.L3_MISS  -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  cache1broad, 2 },
+/*7*/	{ "cache2", "thresh >= .2", 
+	  "pmcstat -s MEM_LOAD_UOPS_RETIRED.L3_HIT  -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  cache2broad, 2 },
+/*8*/	{ "contested", "thresh >= .05", 
+	  "pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HITM  -s CPU_CLK_UNHALTED.THREAD_P  -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_MISS -w 1",
+	  contestedbroad, 2 },
+/*9*/	{ "datashare", "thresh >= .05",
+	  "pmcstat -s MEM_LOAD_UOPS_LLC_HIT_RETIRED.XSNP_HIT -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  datasharing_has, 2 },
+/*10*/	{ "blockstorefwd", "thresh >= .05", 
+	  "pmcstat -s LD_BLOCKS_STORE_FORWARD -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  blockstoreforward, 2 },
+/*11*/	{ "aliasing_4k", "thresh >= .1", 
+	  "pmcstat -s LD_BLOCKS_PARTIAL.ADDRESS_ALIAS -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  aliasing_broad, 2 }, 
+/*12*/	{ "dtlbmissload", "thresh >= .1", 
+	  "pmcstat -s DTLB_LOAD_MISSES.STLB_HIT_4K -s DTLB_LOAD_MISSES.WALK_DURATION -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  dtlb_missload, 3 },
+/*13*/	{ "br_miss", "thresh >= .2", 
+	  "pmcstat -s CPU_CLK_UNHALTED.THREAD_P -s BR_MISP_RETIRED.ALL_BRANCHES -s MACHINE_CLEARS.CYCLES -s UOPS_ISSUED.ANY -s UOPS_RETIRED.RETIRE_SLOTS -s INT_MISC.RECOVERY_CYCLES -w 1",
+	  br_mispredict_broad, 7 },
+/*14*/	{ "clears", "thresh >= .02", 
+	  "pmcstat -s MACHINE_CLEARS.CYCLES -s MACHINE_CLEARS.MEMORY_ORDERING -s MACHINE_CLEARS.SMC -s MACHINE_CLEARS.MASKMOV -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  clears_broad, 5 },
+/*15*/	{ "fpassist", "look for a excessive value", 
+	  "pmcstat -s FP_ASSIST.ANY -s INST_RETIRED.ANY_P -w 1",
+	  fpassists, 2 },
+/*16*/	{ "otherassistavx", "look for a excessive value", 
+	  "pmcstat -s OTHER_ASSISTS.AVX_TO_SSE -s CPU_CLK_UNHALTED.THREAD_P -w 1",
+	  otherassistavx, 2 },
+/*17*/	{ "microassist", "thresh >= .2", 
+	  "pmcstat -s IDQ.MS_UOPS,cmask=1 -s CPU_CLK_UNHALTED.THREAD_P -s UOPS_ISSUED.ANY -s UOPS_RETIRED.RETIRE_SLOTS  -w 1",
+	  microassist_broad, 4 },
 };
 
+
 static void
 set_sandybridge(void)
 {
@@ -2001,6 +1988,7 @@ set_haswell(void)
 	the_cpu.explain = explain_name_has;
 }
 
+
 static void
 set_broadwell(void)
 {
@@ -2010,41 +1998,57 @@ set_broadwell(void)
 	the_cpu.explain = explain_name_broad;
 }
 
-static void
-set_expression(char *name)
+
+static int
+set_expression(const char *name)
 {
 	int found = 0, i;
-
-	for (i = 0; i < the_cpu.number; i++) {
+	for(i=0 ; i< the_cpu.number; i++) {
 		if (strcmp(name, the_cpu.ents[i].name) == 0) {
 			found = 1;
 			expression = the_cpu.ents[i].func;
 			command = the_cpu.ents[i].command;
 			threshold = the_cpu.ents[i].thresh;
+			if  (the_cpu.ents[i].counters_required > max_pmc_counters) {
+				printf("Test %s requires that the CPU have %d counters and this CPU has only %d\n",
+				       the_cpu.ents[i].name,
+				       the_cpu.ents[i].counters_required, max_pmc_counters);
+				printf("Sorry this test can not be run\n");
+				if (run_all == 0) {
+					exit(-1);
+				} else {
+					return(-1);
+				}
+			}
 			break;
 		}
 	}
 	if (!found) {
 		printf("For CPU type %s we have no expression:%s\n",
-		    the_cpu.cputype, name);
+		       the_cpu.cputype, name);
 		exit(-1);
 	}
+	return(0);
 }
 
+
+
+
+
 static int
-validate_expression(char *name)
+validate_expression(char *name) 
 {
 	int i, found;
 
 	found = 0;
-	for (i = 0; i < the_cpu.number; i++) {
+	for(i=0 ; i< the_cpu.number; i++) {
 		if (strcmp(name, the_cpu.ents[i].name) == 0) {
 			found = 1;
 			break;
 		}
 	}
 	if (!found) {
-		return (-1);
+		return(-1);
 	}
 	return (0);
 }
@@ -2052,9 +2056,9 @@ validate_expression(char *name)
 static void
 do_expression(struct counters *cpu, int pos)
 {
-	if (expression == NULL)
+	if (expression == NULL) 
 		return;
-	(*expression) (cpu, pos);
+	(*expression)(cpu, pos);
 }
 
 static void
@@ -2062,57 +2066,60 @@ process_header(int idx, char *p)
 {
 	struct counters *up;
 	int i, len, nlen;
-
-	/*
-	 * Given header element idx, at p in form 's/NN/nameof' process the
-	 * entry to pull out the name and the CPU number.
+	/* 
+	 * Given header element idx, at p in
+	 * form 's/NN/nameof'
+	 * process the entry to pull out the name and
+	 * the CPU number.
 	 */
 	if (strncmp(p, "s/", 2)) {
 		printf("Check -- invalid header no s/ in %s\n",
-		    p);
+		       p);
 		return;
 	}
 	up = &cnts[idx];
 	up->cpu = strtol(&p[2], NULL, 10);
 	len = strlen(p);
-	for (i = 2; i < len; i++) {
+	for (i=2; i<len; i++) {
 		if (p[i] == '/') {
-			nlen = strlen(&p[(i + 1)]);
-			if (nlen < (MAX_NLEN - 1)) {
-				strcpy(up->counter_name, &p[(i + 1)]);
+			nlen = strlen(&p[(i+1)]);
+			if (nlen < (MAX_NLEN-1)) {
+				strcpy(up->counter_name, &p[(i+1)]);
 			} else {
-				strncpy(up->counter_name, &p[(i + 1)], (MAX_NLEN - 1));
+				strncpy(up->counter_name, &p[(i+1)], (MAX_NLEN-1));
 			}
 		}
 	}
 }
 
 static void
-build_counters_from_header(FILE * io)
+build_counters_from_header(FILE *io)
 {
 	char buffer[8192], *p;
 	int i, len, cnt;
 	size_t mlen;
 
-	/*
-	 * We have a new start, lets setup our headers and cpus.
+	/* We have a new start, lets 
+	 * setup our headers and cpus.
 	 */
 	if (fgets(buffer, sizeof(buffer), io) == NULL) {
 		printf("First line can't be read from file err:%d\n", errno);
 		return;
 	}
 	/*
-	 * Ok output is an array of counters. Once we start to read the
-	 * values in we must put them in there slot to match there CPU and
-	 * counter being updated. We create a mass array of the counters,
-	 * filling in the CPU and counter name.
+	 * Ok output is an array of counters. Once
+	 * we start to read the values in we must
+	 * put them in there slot to match there CPU and 
+	 * counter being updated. We create a mass array
+	 * of the counters, filling in the CPU and 
+	 * counter name. 
 	 */
 	/* How many do we get? */
 	len = strlen(buffer);
-	for (i = 0, cnt = 0; i < len; i++) {
+	for (i=0, cnt=0; i<len; i++) {
 		if (strncmp(&buffer[i], "s/", 2) == 0) {
 			cnt++;
-			for (; i < len; i++) {
+			for(;i<len;i++) {
 				if (buffer[i] == ' ')
 					break;
 			}
@@ -2126,10 +2133,10 @@ build_counters_from_header(FILE * io)
 		return;
 	}
 	memset(cnts, 0, mlen);
-	for (i = 0, cnt = 0; i < len; i++) {
+	for (i=0, cnt=0; i<len; i++) {
 		if (strncmp(&buffer[i], "s/", 2) == 0) {
 			p = &buffer[i];
-			for (; i < len; i++) {
+			for(;i<len;i++) {
 				if (buffer[i] == ' ') {
 					buffer[i] = 0;
 					break;
@@ -2140,22 +2147,22 @@ build_counters_from_header(FILE * io)
 		}
 	}
 	if (verbose)
-		printf("We have %d entries\n", cnt);
+		printf("We have %d entries\n", cnt);	
 }
 extern int max_to_collect;
 int max_to_collect = MAX_COUNTER_SLOTS;
 
 static int
-read_a_line(FILE * io)
+read_a_line(FILE *io) 
 {
-	char buffer[8192], *p, *stop;
+	char buffer[8192], *p, *stop;	
 	int pos, i;
 
 	if (fgets(buffer, sizeof(buffer), io) == NULL) {
-		return (0);
+		return(0);
 	}
 	p = buffer;
-	for (i = 0; i < ncnts; i++) {
+	for (i=0; i<ncnts; i++) {
 		pos = cnts[i].pos;
 		cnts[i].vals[pos] = strtol(p, &stop, 0);
 		cnts[i].pos++;
@@ -2166,7 +2173,7 @@ read_a_line(FILE * io)
 }
 
 extern int cpu_count_out;
-int cpu_count_out = 0;
+int cpu_count_out=0;
 
 static void
 print_header(void)
@@ -2174,13 +2181,13 @@ print_header(void)
 	int i, cnt, printed_cnt;
 
 	printf("*********************************\n");
-	for (i = 0, cnt = 0; i < MAX_CPU; i++) {
+	for(i=0, cnt=0; i<MAX_CPU; i++) {
 		if (glob_cpu[i]) {
 			cnt++;
 		}
-	}
+	}	
 	cpu_count_out = cnt;
-	for (i = 0, printed_cnt = 0; i < MAX_CPU; i++) {
+	for(i=0, printed_cnt=0; i<MAX_CPU; i++) {
 		if (glob_cpu[i]) {
 			printf("CPU%d", i);
 			printed_cnt++;
@@ -2200,7 +2207,7 @@ lace_cpus_together(void)
 	int i, j, lace_cpu;
 	struct counters *cpat, *at;
 
-	for (i = 0; i < ncnts; i++) {
+	for(i=0; i<ncnts; i++) {
 		cpat = &cnts[i];
 		if (cpat->next_cpu) {
 			/* Already laced in */
@@ -2218,7 +2225,7 @@ lace_cpus_together(void)
 			continue;
 		}
 		/* Ok look forward for cpu->cpu and link in */
-		for (j = (i + 1); j < ncnts; j++) {
+		for(j=(i+1); j<ncnts; j++) {
 			at = &cnts[j];
 			if (at->next_cpu) {
 				continue;
@@ -2239,19 +2246,15 @@ process_file(char *filename)
 	FILE *io;
 	int i;
 	int line_at, not_done;
-	pid_t pid_of_command = 0;
+	pid_t pid_of_command=0;
 
-	if (filename == NULL) {
+	if (filename ==  NULL) {
 		io = my_popen(command, "r", &pid_of_command);
-		if (io == NULL) {
-			printf("Can't popen the command %s\n", command);
-			return;
-		}
 	} else {
 		io = fopen(filename, "r");
 		if (io == NULL) {
 			printf("Can't process file %s err:%d\n",
-			    filename, errno);
+			       filename, errno);
 			return;
 		}
 	}
@@ -2259,25 +2262,23 @@ process_file(char *filename)
 	if (cnts == NULL) {
 		/* Nothing we can do */
 		printf("Nothing to do -- no counters built\n");
-		if (filename) {
+		if (io) {
 			fclose(io);
-		} else {
-			my_pclose(io, pid_of_command);
 		}
 		return;
 	}
 	lace_cpus_together();
 	print_header();
 	if (verbose) {
-		for (i = 0; i < ncnts; i++) {
+		for (i=0; i<ncnts; i++) {
 			printf("Counter:%s cpu:%d index:%d\n",
-			    cnts[i].counter_name,
-			    cnts[i].cpu, i);
+			       cnts[i].counter_name,
+			       cnts[i].cpu, i);
 		}
 	}
 	line_at = 0;
 	not_done = 1;
-	while (not_done) {
+	while(not_done) {
 		if (read_a_line(io)) {
 			line_at++;
 		} else {
@@ -2288,10 +2289,9 @@ process_file(char *filename)
 		}
 		if (filename == NULL) {
 			int cnt;
-
 			/* For the ones we dynamically open we print now */
-			for (i = 0, cnt = 0; i < MAX_CPU; i++) {
-				do_expression(glob_cpu[i], (line_at - 1));
+			for(i=0, cnt=0; i<MAX_CPU; i++) {
+				do_expression(glob_cpu[i], (line_at-1));
 				cnt++;
 				if (cnt == cpu_count_out) {
 					printf("\n");
@@ -2308,12 +2308,25 @@ process_file(char *filename)
 		my_pclose(io, pid_of_command);
 	}
 }
-
 #if defined(__amd64__)
 #define cpuid(in,a,b,c,d)\
   asm("cpuid": "=a" (a), "=b" (b), "=c" (c), "=d" (d) : "a" (in));
+
+static __inline void
+do_cpuid(u_int ax, u_int cx, u_int *p)
+{
+	__asm __volatile("cpuid"
+			 : "=a" (p[0]), "=b" (p[1]), "=c" (p[2]), "=d" (p[3])
+			 :  "0" (ax), "c" (cx) );
+}
+
 #else
-#define cpuid(in, a, b, c, d)
+#define cpuid(in, a, b, c, d) 
+static __inline void
+do_cpuid(u_int ax, u_int cx, u_int *p)
+{
+}
+
 #endif
 
 static void
@@ -2321,10 +2334,11 @@ get_cpuid_set(void)
 {
 	unsigned long eax, ebx, ecx, edx;
 	int model;
-	pid_t pid_of_command = 0;
+	pid_t pid_of_command=0;
 	size_t sz, len;
 	FILE *io;
 	char linebuf[1024], *str;
+	u_int reg[4];
 
 	eax = ebx = ecx = edx = 0;
 
@@ -2357,25 +2371,20 @@ get_cpuid_set(void)
 			printf("Intel Pentium P6\n");
 			goto not_supported;
 			break;
-		case 0x3:
+		case 0x3: 
 		case 0x5:
 			printf("Intel PII\n");
 			goto not_supported;
 			break;
-		case 0x6:
-		case 0x16:
+		case 0x6: case 0x16:
 			printf("Intel CL\n");
 			goto not_supported;
 			break;
-		case 0x7:
-		case 0x8:
-		case 0xA:
-		case 0xB:
+		case 0x7: case 0x8: case 0xA: case 0xB:
 			printf("Intel PIII\n");
 			goto not_supported;
 			break;
-		case 0x9:
-		case 0xD:
+		case 0x9: case 0xD:
 			printf("Intel PM\n");
 			goto not_supported;
 			break;
@@ -2396,10 +2405,14 @@ get_cpuid_set(void)
 			goto not_supported;
 			break;
 		case 0x1A:
-		case 0x1E:	/* Per Intel document 253669-032 9/2009, pages
-				 * A-2 and A-57 */
-		case 0x1F:	/* Per Intel document 253669-032 9/2009, pages
-				 * A-2 and A-57 */
+		case 0x1E:	/*
+				 * Per Intel document 253669-032 9/2009,
+				 * pages A-2 and A-57
+				 */
+		case 0x1F:	/*
+				 * Per Intel document 253669-032 9/2009,
+				 * pages A-2 and A-57
+				 */
 			printf("Intel COREI7\n");
 			goto not_supported;
 			break;
@@ -2455,7 +2468,7 @@ get_cpuid_set(void)
 			break;
 		case 0x4f:
 		case 0x56:
-			printf("Intel BROADWEL (Xeon)L\n");
+			printf("Intel BROADWEL (Xeon)\n");
 			set_broadwell();
 			break;
 
@@ -2466,7 +2479,7 @@ get_cpuid_set(void)
 			break;
 		default:
 			printf("Intel model 0x%x is not known -- sorry\n",
-			    model);
+			       model);
 			goto not_supported;
 			break;
 		}
@@ -2476,6 +2489,9 @@ get_cpuid_set(void)
 		goto not_supported;
 		break;
 	}
+	do_cpuid(0xa, 0, reg);
+	max_pmc_counters = (reg[3] & 0x0000000f) + 1;
+	printf("We have %d PMC counters to work with\n", max_pmc_counters);
 	/* Ok lets load the list of all known PMC's */
 	io = my_popen("/usr/sbin/pmccontrol -L", "r", &pid_of_command);
 	if (valid_pmcs == NULL) {
@@ -2484,26 +2500,27 @@ get_cpuid_set(void)
 		sz = sizeof(char *) * pmc_allocated_cnt;
 		valid_pmcs = malloc(sz);
 		if (valid_pmcs == NULL) {
-			printf("No memory allocation fails at startup?\n");
+			printf("No memory allocation fails at startup?\n");	
 			exit(-1);
 		}
 		memset(valid_pmcs, 0, sz);
 	}
+	
 	while (fgets(linebuf, sizeof(linebuf), io) != NULL) {
 		if (linebuf[0] != '\t') {
 			/* sometimes headers ;-) */
 			continue;
 		}
 		len = strlen(linebuf);
-		if (linebuf[(len - 1)] == '\n') {
+		if (linebuf[(len-1)] == '\n') {
 			/* Likely */
-			linebuf[(len - 1)] = 0;
+			linebuf[(len-1)] = 0;
 		}
 		str = &linebuf[1];
 		len = strlen(str) + 1;
 		valid_pmcs[valid_pmc_cnt] = malloc(len);
 		if (valid_pmcs[valid_pmc_cnt] == NULL) {
-			printf("No memory2 allocation fails at startup?\n");
+			printf("No memory2 allocation fails at startup?\n");	
 			exit(-1);
 		}
 		memset(valid_pmcs[valid_pmc_cnt], 0, len);
@@ -2516,7 +2533,7 @@ get_cpuid_set(void)
 			sz = sizeof(char *) * (pmc_allocated_cnt * 2);
 			more = malloc(sz);
 			if (more == NULL) {
-				printf("No memory3 allocation fails at startup?\n");
+				printf("No memory3 allocation fails at startup?\n");	
 				exit(-1);
 			}
 			memset(more, 0, sz);
@@ -2526,10 +2543,10 @@ get_cpuid_set(void)
 			valid_pmcs = more;
 		}
 	}
-	my_pclose(io, pid_of_command);
+	my_pclose(io, pid_of_command);	
 	return;
 not_supported:
-	printf("Not supported\n");
+	printf("Not supported\n");	
 	exit(-1);
 }
 
@@ -2537,12 +2554,11 @@ static void
 explain_all(void)
 {
 	int i;
-
-	printf("For CPU's of type %s the following expressions are available:\n", the_cpu.cputype);
+	printf("For CPU's of type %s the following expressions are available:\n",the_cpu.cputype);
 	printf("-------------------------------------------------------------\n");
-	for (i = 0; i < the_cpu.number; i++) {
+	for(i=0; i<the_cpu.number; i++){
 		printf("For -e %s ", the_cpu.ents[i].name);
-		(*the_cpu.explain) (the_cpu.ents[i].name);
+		(*the_cpu.explain)(the_cpu.ents[i].name);
 		printf("----------------------------\n");
 	}
 }
@@ -2551,7 +2567,7 @@ static void
 test_for_a_pmc(const char *pmc, int out_so_far)
 {
 	FILE *io;
-	pid_t pid_of_command = 0;
+	pid_t pid_of_command=0;	
 	char my_command[1024];
 	char line[1024];
 	char resp[1024];
@@ -2559,12 +2575,12 @@ test_for_a_pmc(const char *pmc, int out_so_far)
 
 	if (out_so_far < 50) {
 		len = 50 - out_so_far;
-		for (i = 0; i < len; i++) {
+		for(i=0; i<len; i++) {
 			printf(" ");
 		}
 	}
 	sprintf(my_command, "/usr/sbin/pmcstat -w .25 -c 0 -s %s", pmc);
-	io = my_popen(my_command, "r", &pid_of_command);
+	io = my_popen(my_command, "r", &pid_of_command);	
 	if (io == NULL) {
 		printf("Failed -- popen fails\n");
 		return;
@@ -2576,11 +2592,11 @@ test_for_a_pmc(const char *pmc, int out_so_far)
 		goto out;
 	}
 	llen = strlen(line);
-	if (line[(llen - 1)] == '\n') {
-		line[(llen - 1)] = 0;
+	if (line[(llen-1)] == '\n') {
+		line[(llen-1)] = 0;
 		llen--;
 	}
-	for (i = 2; i < (llen - len); i++) {
+	for(i=2; i<(llen-len); i++) {
 		if (strncmp(&line[i], "ERROR", 5) == 0) {
 			printf("Failed %s\n", line);
 			goto out;
@@ -2592,9 +2608,9 @@ test_for_a_pmc(const char *pmc, int out_so_far)
 				goto out;
 			}
 			len = strlen(line);
-			for (j = 0; j < len; j++) {
+			for (j=0; j<len; j++) {
 				if (line[j] == ' ') {
-					j++;
+					j++; 
 				} else {
 					break;
 				}
@@ -2602,7 +2618,7 @@ test_for_a_pmc(const char *pmc, int out_so_far)
 			printf("Pass");
 			len = strlen(&line[j]);
 			if (len < 20) {
-				for (k = 0; k < (20 - len); k++) {
+				for(k=0; k<(20-len); k++) {
 					printf(" ");
 				}
 			}
@@ -2616,8 +2632,8 @@ test_for_a_pmc(const char *pmc, int out_so_far)
 	}
 	printf("Failed -- '%s' not '%s'\n", line, resp);
 out:
-	my_pclose(io, pid_of_command);
-
+	my_pclose(io, pid_of_command);		
+	
 }
 
 static int
@@ -2625,16 +2641,15 @@ add_it_to(char **vars, int cur_cnt, char *name)
 {
 	int i;
 	size_t len;
-
-	for (i = 0; i < cur_cnt; i++) {
+	for(i=0; i<cur_cnt; i++) {
 		if (strcmp(vars[i], name) == 0) {
 			/* Already have */
-			return (0);
+			return(0);
 		}
 	}
 	if (vars[cur_cnt] != NULL) {
-		printf("Cur_cnt:%d filled with %s??\n",
-		    cur_cnt, vars[cur_cnt]);
+		printf("Cur_cnt:%d filled with %s??\n", 
+		       cur_cnt, vars[cur_cnt]);
 		exit(-1);
 	}
 	/* Ok its new */
@@ -2646,16 +2661,18 @@ add_it_to(char **vars, int cur_cnt, char *name)
 	}
 	memset(vars[cur_cnt], 0, len);
 	strcpy(vars[cur_cnt], name);
-	return (1);
+	return(1);
 }
 
 static char *
 build_command_for_exp(struct expression *exp)
 {
 	/*
-	 * Build the pmcstat command to handle the passed in expression.
-	 * /usr/sbin/pmcstat -w 1 -s NNN -s QQQ where NNN and QQQ represent
-	 * the PMC's in the expression uniquely..
+	 * Build the pmcstat command to handle
+	 * the passed in expression.
+	 * /usr/sbin/pmcstat -w 1 -s NNN -s QQQ
+	 * where NNN and QQQ represent the PMC's in the expression
+	 * uniquely..
 	 */
 	char forming[1024];
 	int cnt_pmc, alloced_pmcs, i;
@@ -2686,39 +2703,39 @@ build_command_for_exp(struct expression *exp)
 	at = exp;
 	while (at) {
 		if (at->type == TYPE_VALUE_PMC) {
-			if (add_it_to(vars, alloced_pmcs, at->name)) {
+			if(add_it_to(vars, alloced_pmcs, at->name)) {
 				alloced_pmcs++;
 			}
 		}
 		at = at->next;
 	}
 	/* Now we have a unique list in vars so create our command */
-	mal = 23;		/* "/usr/sbin/pmcstat -w 1"  + \0 */
-	for (i = 0; i < alloced_pmcs; i++) {
+	mal = 23; /*	"/usr/sbin/pmcstat -w 1"  + \0 */
+	for(i=0; i<alloced_pmcs; i++) {
 		mal += strlen(vars[i]) + 4;	/* var + " -s " */
 	}
-	cmd = malloc((mal + 2));
+	cmd = malloc((mal+2));
 	if (cmd == NULL) {
 		printf("%s out of mem\n", __FUNCTION__);
 		exit(-1);
 	}
-	memset(cmd, 0, (mal + 2));
+	memset(cmd, 0, (mal+2));
 	strcpy(cmd, "/usr/sbin/pmcstat -w 1");
 	at = exp;
-	for (i = 0; i < alloced_pmcs; i++) {
+	for(i=0; i<alloced_pmcs; i++) {
 		sprintf(forming, " -s %s", vars[i]);
 		strcat(cmd, forming);
 		free(vars[i]);
 		vars[i] = NULL;
 	}
 	free(vars);
-	return (cmd);
+	return(cmd);
 }
 
 static int
 user_expr(struct counters *cpu, int pos)
 {
-	int ret;
+	int ret;	
 	double res;
 	struct counters *var;
 	struct expression *at;
@@ -2741,7 +2758,7 @@ user_expr(struct counters *cpu, int pos)
 	}
 	res = run_expr(master_exp, 1, NULL);
 	ret = printf("%1.3f", res);
-	return (ret);
+	return(ret);
 }
 
 
@@ -2757,10 +2774,9 @@ static void
 run_tests(void)
 {
 	int i, lenout;
-
 	printf("Running tests on %d PMC's this may take some time\n", valid_pmc_cnt);
 	printf("------------------------------------------------------------------------\n");
-	for (i = 0; i < valid_pmc_cnt; i++) {
+	for(i=0; i<valid_pmc_cnt; i++) {
 		lenout = printf("%s", valid_pmcs[i]);
 		fflush(stdout);
 		test_for_a_pmc(valid_pmcs[i], lenout);
@@ -2770,12 +2786,11 @@ static void
 list_all(void)
 {
 	int i, cnt, j;
-
 	printf("PMC                                               Abbreviation\n");
 	printf("--------------------------------------------------------------\n");
-	for (i = 0; i < valid_pmc_cnt; i++) {
+	for(i=0; i<valid_pmc_cnt; i++) {
 		cnt = printf("%s", valid_pmcs[i]);
-		for (j = cnt; j < 52; j++) {
+		for(j=cnt; j<52; j++) {
 			printf(" ");
 		}
 		printf("%%%d\n", i);
@@ -2787,23 +2802,27 @@ int
 main(int argc, char **argv)
 {
 	int i, j, cnt;
-	char *filename = NULL;
-	char *name = NULL;
+	char *filename=NULL;
+	const char *name=NULL;
 	int help_only = 0;
 	int test_mode = 0;
+	int test_at = 0;
 
 	get_cpuid_set();
 	memset(glob_cpu, 0, sizeof(glob_cpu));
-	while ((i = getopt(argc, argv, "LHhvm:i:?e:TE:")) != -1) {
+	while ((i = getopt(argc, argv, "ALHhvm:i:?e:TE:")) != -1) {
 		switch (i) {
+		case 'A':
+			run_all = 1;
+			break;
 		case 'L':
 			list_all();
-			return (0);
+			return(0);
 		case 'H':
 			printf("**********************************\n");
 			explain_all();
 			printf("**********************************\n");
-			return (0);
+			return(0);
 			break;
 		case 'T':
 			test_mode = 1;
@@ -2817,7 +2836,7 @@ main(int argc, char **argv)
 		case 'e':
 			if (validate_expression(optarg)) {
 				printf("Unknown expression %s\n", optarg);
-				return (0);
+				return(0);
 			}
 			name = optarg;
 			set_expression(optarg);
@@ -2840,9 +2859,9 @@ main(int argc, char **argv)
 			break;
 		case '?':
 		default:
-	use:
+		use:
 			printf("Use %s [ -i inputfile -v -m max_to_collect -e expr -E -h -? -H]\n",
-			    argv[0]);
+			       argv[0]);
 			printf("-i inputfile -- use source as inputfile not stdin (if stdin collect)\n");
 			printf("-v -- verbose dump debug type things -- you don't want this\n");
 			printf("-m N -- maximum to collect is N measurments\n");
@@ -2851,56 +2870,85 @@ main(int argc, char **argv)
 			printf("-h -- Don't do the expression I put in -e xxx just explain what it does and exit\n");
 			printf("-H -- Don't run anything, just explain all canned expressions\n");
 			printf("-T -- Test all PMC's defined by this processor\n");
-			return (0);
+			printf("-A -- Run all canned tests\n");
+			return(0);
 			break;
 		};
 	}
-	if ((name == NULL) && (filename == NULL) && (test_mode == 0) && (master_exp == NULL)) {
+	if ((run_all == 0) && (name == NULL) && (filename == NULL) &&
+	    (test_mode == 0) && (master_exp == NULL)) {
 		printf("Without setting an expression we cannot dynamically gather information\n");
 		printf("you must supply a filename (and you probably want verbosity)\n");
 		goto use;
 	}
+	if (run_all && max_to_collect > 10) {
+		max_to_collect = 3;
+	}
 	if (test_mode) {
 		run_tests();
-		return (0);
+		return(0);
 	}
 	printf("*********************************\n");
-	if (master_exp == NULL) {
-		(*the_cpu.explain) (name);
-	} else {
+	if ((master_exp == NULL) && name) {
+		(*the_cpu.explain)(name);
+	} else if (master_exp) {
 		printf("Examine your expression ");
 		print_exp(master_exp);
 		printf("User defined threshold\n");
 	}
 	if (help_only) {
-		return (0);
+		return(0);
+	}
+	if (run_all) {
+	more:
+		name = the_cpu.ents[test_at].name;
+		printf("***Test %s (threshold %s)****\n", name, the_cpu.ents[test_at].thresh);
+		test_at++;
+		if (set_expression(name) == -1) {
+			if (test_at >= the_cpu.number) {
+				goto done;
+			} else
+				goto more;
+		}
+
 	}
 	process_file(filename);
 	if (verbose >= 2) {
-		for (i = 0; i < ncnts; i++) {
+		for (i=0; i<ncnts; i++) {
 			printf("Counter:%s cpu:%d index:%d\n",
-			    cnts[i].counter_name,
-			    cnts[i].cpu, i);
-			for (j = 0; j < cnts[i].pos; j++) {
+			       cnts[i].counter_name,
+			       cnts[i].cpu, i);
+			for(j=0; j<cnts[i].pos; j++) {
 				printf(" val - %ld\n", (long int)cnts[i].vals[j]);
 			}
 			printf(" sum - %ld\n", (long int)cnts[i].sum);
 		}
 	}
 	if (expression == NULL) {
-		return (0);
+		return(0);
 	}
-	for (i = 0, cnt = 0; i < MAX_CPU; i++) {
-		if (glob_cpu[i]) {
-			do_expression(glob_cpu[i], -1);
-			cnt++;
-			if (cnt == cpu_count_out) {
-				printf("\n");
-				break;
-			} else {
-				printf("\t");
+	if (max_to_collect > 1) {
+		for(i=0, cnt=0; i<MAX_CPU; i++) {
+			if (glob_cpu[i]) {
+				do_expression(glob_cpu[i], -1);
+				cnt++;
+				if (cnt == cpu_count_out) {
+					printf("\n");
+					break;
+				} else {
+					printf("\t");
+				}
 			}
 		}
 	}
-	return (0);
+	if (run_all && (test_at < the_cpu.number)) {
+		memset(glob_cpu, 0, sizeof(glob_cpu));
+		ncnts = 0;
+		printf("*********************************\n");
+		goto more;
+	} else if (run_all) {
+	done:
+		printf("*********************************\n");
+	}
+	return(0);	
 }