Merge remote-tracking branch 'origin/mppa-work' into mppa-thread

5 files changed, 193 insertions, 52 deletions

diff --git a/test/monniaux/clock.c b/test/monniaux/clock.c
index fb636667..4ec679f6 100644
--- a/test/monniaux/clock.c
+++ b/test/monniaux/clock.c
@@ -24,9 +24,9 @@ cycle_t get_current_cycle(void) {
 }
 
 void print_total_clock(void) {
-  printf("time cycles: %lu\n", total_clock);
+  printf("time cycles: %" PRcycle "\n", total_clock);
 }
 
 void printerr_total_clock(void) {
-  fprintf(stderr, "time cycles: %lu\n", total_clock);
+  fprintf(stderr, "time cycles: %" PRcycle "\n", total_clock);
 }
diff --git a/test/monniaux/cycles.h b/test/monniaux/cycles.h
index 21541145..c7dc582b 100644
--- a/test/monniaux/cycles.h
+++ b/test/monniaux/cycles.h
@@ -1,13 +1,11 @@
+#include <stdint.h>
 #include <inttypes.h>
 #include <stdio.h>
-typedef unsigned long cycle_t;
-
-#ifdef MAX_MEASURES
-  static cycle_t _last_stop[MAX_MEASURES] = {0};
-  static cycle_t _total_cycles[MAX_MEASURES] = {0};
-#endif
 
 #ifdef __K1C__
+typedef uint64_t cycle_t;
+#define PRcycle PRId64
+
 #include <../../k1-cos/include/hal/cos_registers.h>
 
 static inline void cycle_count_config(void)
@@ -27,18 +25,57 @@ static inline cycle_t get_cycle(void)
 #else // not K1C
 static inline void cycle_count_config(void) { }
 
-#ifdef  __x86_64__
+#if defined(__i386__) || defined( __x86_64__)
+#define PRcycle PRId64
+typedef uint64_t cycle_t;
 #include <x86intrin.h>
 static inline cycle_t get_cycle(void) { return __rdtsc(); }
 
 #elif __riscv
+#ifdef __riscv32
+#define PRcycle PRId32
+typedef uint32_t cycle_t;
+#else
+#define PRcycle PRId64
+typedef uint64_t cycle_t;
+#endif
 static inline cycle_t get_cycle(void) {
   cycle_t cycles;
   asm volatile ("rdcycle %0" : "=r" (cycles));
   return cycles;
 }
 
+#elif defined (__ARM_ARCH) && (__ARM_ARCH >= 6)
+#if (__ARM_ARCH < 8)
+typedef uint32_t cycle_t;
+#define PRcycle PRId32
+
+/* need this kernel module
+https://github.com/zertyz/MTL/tree/master/cpp/time/kernel/arm */
+static inline cycle_t get_cycle(void) {
+  cycle_t cycles;
+  __asm__ volatile ("mrc p15, 0, %0, c9, c13, 0":"=r" (cycles));
+  return cycles;
+}
 #else
+#define PRcycle PRId64
+typedef uint64_t cycle_t;
+/* need this kernel module:
+https://github.com/jerinjacobk/armv8_pmu_cycle_counter_el0
+
+on 5+ kernels, remove first argument of access_ok macro */
+
+static inline cycle_t get_cycle(void)
+{
+  uint64_t val;
+  __asm__ volatile("mrs %0, pmccntr_el0" : "=r"(val));
+  return val;
+}
+#endif
+
+#else
+#define PRcycle PRId32
+typedef uint32_t cycle_t;
 static inline cycle_t get_cycle(void) { return 0; }
 #endif
 #endif
@@ -48,3 +85,9 @@ static inline cycle_t get_cycle(void) { return 0; }
   #define TIMESTOP(i) {cycle_t cur = get_cycle(); _total_cycles[i] += cur - _last_stop[i]; _last_stop[i] = cur;}
   #define TIMEPRINT(n) { for (int i = 0; i <= n; i++) printf("%d cycles: %" PRIu64 "\n", i, _total_cycles[i]); }
 #endif
+
+
+#ifdef MAX_MEASURES
+  static cycle_t _last_stop[MAX_MEASURES] = {0};
+  static cycle_t _total_cycles[MAX_MEASURES] = {0};
+#endif
diff --git a/test/monniaux/quicksort/quicksort_run.c b/test/monniaux/quicksort/quicksort_run.c
index c35d0752..3c640b24 100644
--- a/test/monniaux/quicksort/quicksort_run.c
+++ b/test/monniaux/quicksort/quicksort_run.c
@@ -13,7 +13,7 @@ int main (void) {
   quicksort(vec, len);
   quicksort_time = get_cycle() - quicksort_time;
   printf("sorted=%s\n"
-	 "time cycles:%" PRIu64 "\n",
+	 "time cycles:%" PRcycle "\n",
 	 data_vec_is_sorted(vec, len)?"true":"false",
 	 quicksort_time);
   free(vec);
diff --git a/test/monniaux/yarpgen/Makefile b/test/monniaux/yarpgen/Makefile
index 9da82deb..28bd5ae0 100644
--- a/test/monniaux/yarpgen/Makefile
+++ b/test/monniaux/yarpgen/Makefile
@@ -1,52 +1,98 @@
-YARPGEN=yarpgen
-MAX=300
+TARGET_CCOMP=../../../ccomp
+TARGET_CC=gcc
+
+all:
+
+.SECONDARY:
+
+ifndef YARPGEN
+YARPGEN=./yarpgen
+GENERATOR=yarpgen
+endif
+
+ifdef BITS
+YARPGEN+=-m $(BITS)
+CFLAGS+=-m$(BITS)
+endif
+
+MAX=129
 PREFIX=ran%06.f
-include ../rules.mk
-
-K1C_CCOMPFLAGS += -funprototyped -fbitfields
-CCOMPFLAGS += -funprototyped -fbitfields
-
-TARGETS_C=$(shell seq --format $(PREFIX)/func.c 0 $(MAX)) \
-	$(shell seq --format $(PREFIX)/driver.c 0 $(MAX)) \
-	$(shell seq --format $(PREFIX)/init.h 0 $(MAX))
-TARGETS_CCOMP_K1C_S=$(shell seq --format $(PREFIX)/func.ccomp.k1c.s 0 $(MAX)) \
-	$(shell seq --format $(PREFIX)/driver.ccomp.k1c.s 0 $(MAX))
-TARGETS_GCC_K1C_S=$(shell seq --format $(PREFIX)/func.gcc.k1c.s 0 $(MAX)) \
-	$(shell seq --format $(PREFIX)/driver.gcc.k1c.s 0 $(MAX))
-TARGETS_CCOMP_HOST_S=$(shell seq --format $(PREFIX)/func.ccomp.host.s 0 $(MAX)) \
-	$(shell seq --format $(PREFIX)/driver.ccomp.host.s 0 $(MAX))
-TARGETS_GCC_HOST_S=$(shell seq --format $(PREFIX)/func.gcc.host.s 0 $(MAX)) \
-	$(shell seq --format $(PREFIX)/driver.gcc.host.s 0 $(MAX))
-TARGETS_CCOMP_K1C_OUT=$(shell seq --format $(PREFIX)/example.ccomp.k1c.out 0 $(MAX))
-TARGETS_GCC_K1C_OUT=$(shell seq --format $(PREFIX)/example.gcc.k1c.out 0 $(MAX))
-TARGETS_GCC_HOST_OUT=$(shell seq --format $(PREFIX)/example.gcc.host.out 0 $(MAX))
-TARGETS_CCOMP_HOST_OUT=$(shell seq --format $(PREFIX)/example.ccomp.host.out 0 $(MAX))
-TARGETS_CMP=$(shell seq --format $(PREFIX)/example.k1c.cmp 0 $(MAX))
-
-all:	$(TARGETS_CCOMP_K1C_OUT) $(TARGETS_GCC_K1C_OUT) $(TARGETS_GCC_HOST_OUT) $(TARGETS_CCOMP_HOST_OUT) $(TARGETS_CCOMP_K1C_S) $(TARGETS_GCC_K1C_S) $(TARGETS_GCC_HOST_S) $(TARGETS_CCOMP_HOST_S) $(TARGETS_CMP) $(TARGETS_C)
-
-ran%/func.ccomp.k1c.s ran%/func.gcc.k1c.s ran%/func.ccomp.host.s ran%/func.gcc.host.s : ran%/init.h
-
-ran%/example.ccomp.k1c: ran%/func.ccomp.k1c.o ran%/driver.ccomp.k1c.o 
-	$(K1C_CCOMP) $(K1C_CCOMPFLAGS) $+ -o $@
-
-ran%/example.gcc.k1c: ran%/func.gcc.k1c.o ran%/driver.gcc.k1c.o 
-	$(K1C_CC) $(K1C_CFLAGS) $+ -o $@
-
-ran%/example.gcc.host: ran%/func.gcc.host.o ran%/driver.gcc.host.o 
-	$(CC) $(CFLAGS) $+ -o $@
 
-ran%/example.ccomp.host: ran%/func.ccomp.host.o ran%/driver.ccomp.host.o 
-	$(CCOMP) $(CCOMPFLAGS) $+ -o $@
+CCOMPOPTS=-static
+CCOMPFLAGS+=-funprototyped -fbitfields -fno-cse2 -stdlib ../../../runtime # FIXME
+
+TESTS_C=$(shell seq --format $(PREFIX)/func.c 1 $(MAX)) \
+	$(shell seq --format $(PREFIX)/driver.c 1 $(MAX)) \
+	$(shell seq --format $(PREFIX)/init.c 1 $(MAX)) \
+	$(shell seq --format $(PREFIX)/hash.c 1 $(MAX)) \
+	$(shell seq --format $(PREFIX)/check.c 1 $(MAX)) \
+	$(shell seq --format $(PREFIX)/init.h 1 $(MAX))
+
+$(TESTS_C): $(GENERATOR)
+
+TESTS_CCOMP_TARGET_S=$(TEST_C:.c=.ccomp.target.s)
+TESTS_GCC_TARGET_S=$(TEST_C:.c=.gcc.target.s)
+TESTS_GCC_HOST_S=$(TEST_C:.c=.gcc.host.s)
+TESTS_CCOMP_TARGET_OUT=$(shell seq --format $(PREFIX)/example.ccomp.target.out 1 $(MAX))
+TESTS_GCC_TARGET_OUT=$(shell seq --format $(PREFIX)/example.gcc.target.out 1 $(MAX))
+TESTS_GCC_HOST_OUT=$(shell seq --format $(PREFIX)/example.gcc.host.out 1 $(MAX))
+TESTS_CMP=$(shell seq --format $(PREFIX)/example.target.cmp 1 $(MAX)) # $(shell seq --format $(PREFIX)/example.host_target.cmp 1 $(MAX))
+
+all:	$(TESTS_CCOMP_TARGET_OUT) $(TESTS_GCC_TARGET_OUT) $(TESTS_CCOMP_TARGET_S) $(TESTS_GCC_TARGET_S) $(TESTS_CMP) $(TESTS_C)
+
+tests_c: $(TESTS_C)
+
+tests_s: $(TESTS_CCOMP_TARGET_S)
+
+%.ccomp.target.s : %.c
+	$(TARGET_CCOMP) $(CCOMPOPTS) $(CCOMPFLAGS) -S -o $@ $<
+
+%.gcc.target.s : %.c
+	$(TARGET_CC) $(CCOMPOPTS) -S -o $@ $<
 
-ran%/driver.c ran%/func.c ran%/init.h:
-	-mkdir ran$*
+%.gcc.host.s : %.c
+	$(CC) $(CFLAGS) -S -o $@ $<
+
+%.target.o : %.target.s
+	$(TARGET_CC) -c -o $@ $<
+
+%.host.o : %.host.s
+	$(CC) $(CFLAGS) -c -o $@ $<
+
+%.target.out : %.target
+	$(EXECUTE) $< | tee $@
+
+%.host.out : %.host
+	./$< | tee $@
+
+ran%/func.ccomp.target.s ran%/func.gcc.target.s ran%/func.ccomp.host.s ran%/func.gcc.host.s ran%/init.gcc.host.s : ran%/init.h
+
+ran%/example.ccomp.target: ran%/func.ccomp.target.o ran%/driver.ccomp.target.o  ran%/init.ccomp.target.o ran%/check.ccomp.target.o ran%/hash.ccomp.target.o
+	$(TARGET_CCOMP) $(CCOMPOPTS) $(CCOMPFLAGS) $+ -o $@
+
+ran%/example.gcc.target: ran%/func.gcc.target.o ran%/driver.gcc.target.o ran%/init.gcc.target.o ran%/check.gcc.target.o ran%/hash.gcc.target.o
+	$(TARGET_CC) $(TARGET_CFLAGS) $+ -o $@
+
+ran%/example.gcc.host: ran%/func.gcc.host.o ran%/driver.gcc.host.o ran%/init.gcc.host.o ran%/check.gcc.host.o ran%/hash.gcc.host.o
+	$(CC) $(CFLAGS) $+ -o $@
+
+ran%/driver.c ran%/func.c ran%/init.c ran%/check.c ran%/hash.c ran%/init.h:
+	mkdir -p ran$*
 	$(YARPGEN) --seed=$* --out-dir=ran$*/ --std=c99
 
-ran%/example.k1c.cmp : ran%/example.gcc.k1c.out ran%/example.ccomp.k1c.out
+ran%/example.target.cmp : ran%/example.gcc.target.out ran%/example.ccomp.target.out
 	cmp $+ > $@
 
-.PHONY: all clean
+ran%/example.host_target.cmp : ran%/example.gcc.host.out ran%/example.ccomp.target.out
+	cmp $+ > $@
+
+yarpgen:
+	curl -L -o yarpgen_v1.1.tar.gz https://github.com/intel/yarpgen/archive/v1.1.tar.gz
+	tar xfz yarpgen_v1.1.tar.gz
+	$(MAKE) CXX=g++ -C yarpgen-1.1
+	cp yarpgen-1.1/yarpgen $@
+
+.PHONY: all clean tests_c tests_c
 
 clean:
 	-rm -rf ran*
diff --git a/test/monniaux/yarpgen/Makefile.old b/test/monniaux/yarpgen/Makefile.old
new file mode 100644
index 00000000..9da82deb
--- /dev/null
+++ b/test/monniaux/yarpgen/Makefile.old
@@ -0,0 +1,52 @@
+YARPGEN=yarpgen
+MAX=300
+PREFIX=ran%06.f
+include ../rules.mk
+
+K1C_CCOMPFLAGS += -funprototyped -fbitfields
+CCOMPFLAGS += -funprototyped -fbitfields
+
+TARGETS_C=$(shell seq --format $(PREFIX)/func.c 0 $(MAX)) \
+	$(shell seq --format $(PREFIX)/driver.c 0 $(MAX)) \
+	$(shell seq --format $(PREFIX)/init.h 0 $(MAX))
+TARGETS_CCOMP_K1C_S=$(shell seq --format $(PREFIX)/func.ccomp.k1c.s 0 $(MAX)) \
+	$(shell seq --format $(PREFIX)/driver.ccomp.k1c.s 0 $(MAX))
+TARGETS_GCC_K1C_S=$(shell seq --format $(PREFIX)/func.gcc.k1c.s 0 $(MAX)) \
+	$(shell seq --format $(PREFIX)/driver.gcc.k1c.s 0 $(MAX))
+TARGETS_CCOMP_HOST_S=$(shell seq --format $(PREFIX)/func.ccomp.host.s 0 $(MAX)) \
+	$(shell seq --format $(PREFIX)/driver.ccomp.host.s 0 $(MAX))
+TARGETS_GCC_HOST_S=$(shell seq --format $(PREFIX)/func.gcc.host.s 0 $(MAX)) \
+	$(shell seq --format $(PREFIX)/driver.gcc.host.s 0 $(MAX))
+TARGETS_CCOMP_K1C_OUT=$(shell seq --format $(PREFIX)/example.ccomp.k1c.out 0 $(MAX))
+TARGETS_GCC_K1C_OUT=$(shell seq --format $(PREFIX)/example.gcc.k1c.out 0 $(MAX))
+TARGETS_GCC_HOST_OUT=$(shell seq --format $(PREFIX)/example.gcc.host.out 0 $(MAX))
+TARGETS_CCOMP_HOST_OUT=$(shell seq --format $(PREFIX)/example.ccomp.host.out 0 $(MAX))
+TARGETS_CMP=$(shell seq --format $(PREFIX)/example.k1c.cmp 0 $(MAX))
+
+all:	$(TARGETS_CCOMP_K1C_OUT) $(TARGETS_GCC_K1C_OUT) $(TARGETS_GCC_HOST_OUT) $(TARGETS_CCOMP_HOST_OUT) $(TARGETS_CCOMP_K1C_S) $(TARGETS_GCC_K1C_S) $(TARGETS_GCC_HOST_S) $(TARGETS_CCOMP_HOST_S) $(TARGETS_CMP) $(TARGETS_C)
+
+ran%/func.ccomp.k1c.s ran%/func.gcc.k1c.s ran%/func.ccomp.host.s ran%/func.gcc.host.s : ran%/init.h
+
+ran%/example.ccomp.k1c: ran%/func.ccomp.k1c.o ran%/driver.ccomp.k1c.o 
+	$(K1C_CCOMP) $(K1C_CCOMPFLAGS) $+ -o $@
+
+ran%/example.gcc.k1c: ran%/func.gcc.k1c.o ran%/driver.gcc.k1c.o 
+	$(K1C_CC) $(K1C_CFLAGS) $+ -o $@
+
+ran%/example.gcc.host: ran%/func.gcc.host.o ran%/driver.gcc.host.o 
+	$(CC) $(CFLAGS) $+ -o $@
+
+ran%/example.ccomp.host: ran%/func.ccomp.host.o ran%/driver.ccomp.host.o 
+	$(CCOMP) $(CCOMPFLAGS) $+ -o $@
+
+ran%/driver.c ran%/func.c ran%/init.h:
+	-mkdir ran$*
+	$(YARPGEN) --seed=$* --out-dir=ran$*/ --std=c99
+
+ran%/example.k1c.cmp : ran%/example.gcc.k1c.out ran%/example.ccomp.k1c.out
+	cmp $+ > $@
+
+.PHONY: all clean
+
+clean:
+	-rm -rf ran*