[PATCH v4 0/9] FP/VEC/VSX switching optimisations

@@ -1,2 +1,4 @@
 fpu_syscall
 vmx_syscall
+fpu_preempt
+vmx_preempt

@@ -1,4 +1,4 @@
-TEST_PROGS := fpu_syscall vmx_syscall
+TEST_PROGS := fpu_syscall fpu_preempt vmx_syscall vmx_preempt
 
 all: $(TEST_PROGS)
 

@@ -6,7 +6,10 @@ $(TEST_PROGS): ../harness.c
 $(TEST_PROGS): CFLAGS += -O2 -g -pthread -m64 -maltivec
 
 fpu_syscall: fpu_asm.S
+fpu_preempt: fpu_asm.S
+
 vmx_syscall: vmx_asm.S
+vmx_preempt: vmx_asm.S
 
 include ../../lib.mk
 

@@ -159,3 +159,37 @@ FUNC_START(test_fpu)
 	POP_BASIC_STACK(256)
 	blr
 FUNC_END(test_fpu)
+
+#int preempt_fpu(double *darray, int *threads_running, int *running)
+#On starting will (atomically) decrement not_ready as a signal that the FPU
+#has been loaded with darray. Will proceed to check the validity of the FPU
+#registers while running is not zero.
+FUNC_START(preempt_fpu)
+	PUSH_BASIC_STACK(256)
+	std r3,32(sp) #double *darray
+	std r4,40(sp) #volatile int *not_ready
+	std r5,48(sp) #int *running
+	PUSH_FPU(56)
+
+	bl load_fpu
+
+	#Atomic DEC
+	ld r3,40(sp)
+1:	lwarx r4,0,r3
+	addi r4,r4,-1
+	stwcx. r4,0,r3
+	bne- 1b
+
+2:	ld r3, 32(sp)
+	bl check_fpu
+	cmpdi r3,0
+	bne 3f
+	ld r4, 48(sp)
+	ld r5, 0(r4)
+	cmpwi r5,0
+	bne 2b
+
+3:	POP_FPU(56)
+	POP_BASIC_STACK(256)
+	blr
+FUNC_END(preempt_fpu)

@@ -0,0 +1,113 @@
+/*
+ * Copyright 2015, Cyril Bur, IBM Corp.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ *
+ * This test attempts to see if the FPU registers change across preemption.
+ * Two things should be noted here a) The check_fpu function in asm only checks
+ * the non volatile registers as it is reused from the syscall test b) There is
+ * no way to be sure preemption happened so this test just uses many threads
+ * and a long wait. As such, a successful test doesn't mean much but a failure
+ * is bad.
+ */
+
+#include <stdio.h>
+#include <unistd.h>
+#include <sys/syscall.h>
+#include <sys/time.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <stdlib.h>
+#include <pthread.h>
+
+#include "utils.h"
+
+/* Time to wait for workers to get preempted (seconds) */
+#define PREEMPT_TIME 20
+/*
+ * Factor by which to multiply number of online CPUs for total number of
+ * worker threads
+ */
+#define THREAD_FACTOR 8
+
+
+__thread double darray[] = {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0,
+		     1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0,
+		     2.1};
+
+int threads_starting;
+int running;
+
+extern void preempt_fpu(double *darray, int *threads_starting, int *running);
+
+void *preempt_fpu_c(void *p)
+{
+	int i;
+	srand(pthread_self());
+	for (i = 0; i < 21; i++)
+		darray[i] = rand();
+
+	/* Test failed if it ever returns */
+	preempt_fpu(darray, &threads_starting, &running);
+
+	return p;
+}
+
+int test_preempt_fpu(void)
+{
+	int i, rc, threads;
+	pthread_t *tids;
+
+	threads = sysconf(_SC_NPROCESSORS_ONLN) * THREAD_FACTOR;
+	tids = malloc((threads) * sizeof(pthread_t));
+	FAIL_IF(!tids);
+
+	running = true;
+	threads_starting = threads;
+	for (i = 0; i < threads; i++) {
+		rc = pthread_create(&tids[i], NULL, preempt_fpu_c, NULL);
+		FAIL_IF(rc);
+	}
+
+	setbuf(stdout, NULL);
+	/* Not really necessary but nice to wait for every thread to start */
+	printf("\tWaiting for all workers to start...");
+	while(threads_starting)
+		asm volatile("": : :"memory");
+	printf("done\n");
+
+	printf("\tWaiting for %d seconds to let some workers get preempted...", PREEMPT_TIME);
+	sleep(PREEMPT_TIME);
+	printf("done\n");
+
+	printf("\tStopping workers...");
+	/*
+	 * Working are checking this value every loop. In preempt_fpu 'cmpwi r5,0; bne 2b'.
+	 * r5 will have loaded the value of running.
+	 */
+	running = 0;
+	for (i = 0; i < threads; i++) {
+		void *rc_p;
+		pthread_join(tids[i], &rc_p);
+
+		/*
+		 * Harness will say the fail was here, look at why preempt_fpu
+		 * returned
+		 */
+		if ((long) rc_p)
+			printf("oops\n");
+		FAIL_IF((long) rc_p);
+	}
+	printf("done\n");
+
+	free(tids);
+	return 0;
+}
+
+int main(int argc, char *argv[])
+{
+	return test_harness(test_preempt_fpu, "fpu_preempt");
+}

@@ -9,6 +9,7 @@
 
 #include "../basic_asm.h"
 
+#POS MUST BE 16 ALIGNED!
 #define PUSH_VMX(pos,reg) \
 	li	reg,pos; \
 	stvx	v20,reg,sp; \

@@ -35,6 +36,7 @@
 	addi	reg,reg,16; \
 	stvx	v31,reg,sp;
 
+#POS MUST BE 16 ALIGNED!
 #define POP_VMX(pos,reg) \
 	li	reg,pos; \
 	lvx	v20,reg,sp; \

@@ -93,7 +95,7 @@ FUNC_END(load_vmx)
 
 #Should be safe from C, only touches r4, r5 and v0,v1,v2
 FUNC_START(check_vmx)
-	PUSH_BASIC_STACK(16)
+	PUSH_BASIC_STACK(32)
 	mr r4,r3
 	li	r3,1 #assume a bad result
 	li	r5,0

@@ -162,7 +164,7 @@ FUNC_START(check_vmx)
 	cmpdi	r0,0xffffffff
 	bne	1f
 	li	r3,0
-1:	POP_BASIC_STACK(16)
+1:	POP_BASIC_STACK(32)
 	blr
 FUNC_END(check_vmx)
 

@@ -174,7 +176,7 @@ FUNC_START(test_vmx)
 	PUSH_BASIC_STACK(512)
 	std	r3,40(sp) #Address of varray
 	std r4,48(sp) #address of pid
-	PUSH_VMX(56, r4)
+	PUSH_VMX(64, r4)
 
 	bl load_vmx
 

@@ -187,7 +189,41 @@ FUNC_START(test_vmx)
 	ld r3,40(sp)
 	bl check_vmx
 
-	POP_VMX(56,r4)
+	POP_VMX(64,r4)
 	POP_BASIC_STACK(512)
 	blr
 FUNC_END(test_vmx)
+
+#int preempt_vmx(v4si *varray, int *threads_starting, int *running)
+#On starting will (atomically) decrement not_ready as a signal that the FPU
+#has been loaded with varray. Will proceed to check the validity of the FPU
+#registers while running is not zero.
+FUNC_START(preempt_vmx)
+	PUSH_BASIC_STACK(512)
+	std r3,32(sp) #v4si *varray
+	std r4,40(sp) #volatile int *not_ready
+	std r5,48(sp) #int *running
+	PUSH_VMX(64,r4)
+
+	bl load_vmx
+
+	#Atomic DEC
+	ld r3,40(sp)
+1:	lwarx r4,0,r3
+	addi r4,r4,-1
+	stwcx. r4,0,r3
+	bne- 1b
+
+2:	ld r3,32(sp)
+	bl check_vmx
+	cmpdi r3,0
+	bne 3f
+	ld r4,48(sp)
+	ld r5,0(r4)
+	cmpwi r5,0
+	bne 2b
+
+3:	POP_VMX(64,r4)
+	POP_BASIC_STACK(512)
+	blr
+FUNC_END(preempt_vmx)

@@ -0,0 +1,113 @@
+/*
+ * Copyright 2015, Cyril Bur, IBM Corp.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ *
+ * This test attempts to see if the VMX registers change across preemption.
+ * Two things should be noted here a) The check_vmx function in asm only checks
+ * the non volatile registers as it is reused from the syscall test b) There is
+ * no way to be sure preemption happened so this test just uses many threads
+ * and a long wait. As such, a successful test doesn't mean much but a failure
+ * is bad.
+ */
+
+#include <stdio.h>
+#include <unistd.h>
+#include <sys/syscall.h>
+#include <sys/time.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <stdlib.h>
+#include <pthread.h>
+
+#include "utils.h"
+
+/* Time to wait for workers to get preempted (seconds) */
+#define PREEMPT_TIME 20
+/*
+ * Factor by which to multiply number of online CPUs for total number of
+ * worker threads
+ */
+#define THREAD_FACTOR 8
+
+typedef int v4si __attribute__ ((vector_size (16)));
+__thread v4si varray[] = {{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10,11,12},
+	{13,14,15,16},{17,18,19,20},{21,22,23,24},
+	{25,26,27,28},{29,30,31,32},{33,34,35,36},
+	{37,38,39,40},{41,42,43,44},{45,46,47,48}};
+
+int threads_starting;
+int running;
+
+extern void preempt_vmx(v4si *varray, int *threads_starting, int *running);
+
+void *preempt_vmx_c(void *p)
+{
+	int i, j;
+	srand(pthread_self());
+	for (i = 0; i < 12; i++)
+		for (j = 0; j < 4; j++)
+			varray[i][j] = rand();
+
+	/* Test fails if it ever returns */
+	preempt_vmx(varray, &threads_starting, &running);
+	return p;
+}
+
+int test_preempt_vmx(void)
+{
+	int i, rc, threads;
+	pthread_t *tids;
+
+	threads = sysconf(_SC_NPROCESSORS_ONLN) * THREAD_FACTOR;
+	tids = malloc(threads * sizeof(pthread_t));
+	FAIL_IF(!tids);
+
+	running = true;
+	threads_starting = threads;
+	for (i = 0; i < threads; i++) {
+		rc = pthread_create(&tids[i], NULL, preempt_vmx_c, NULL);
+		FAIL_IF(rc);
+	}
+
+	setbuf(stdout, NULL);
+	/* Not really nessesary but nice to wait for every thread to start */
+	printf("\tWaiting for all workers to start...");
+	while(threads_starting)
+		asm volatile("": : :"memory");
+	printf("done\n");
+
+	printf("\tWaiting for %d seconds to let some workers get preempted...", PREEMPT_TIME);
+	sleep(PREEMPT_TIME);
+	printf("done\n");
+
+	printf("\tStopping workers...");
+	/*
+	 * Working are checking this value every loop. In preempt_vmx 'cmpwi r5,0; bne 2b'.
+	 * r5 will have loaded the value of running.
+	 */
+	running = 0;
+	for (i = 0; i < threads; i++) {
+		void *rc_p;
+		pthread_join(tids[i], &rc_p);
+
+		/*
+		 * Harness will say the fail was here, look at why preempt_vmx
+		 * returned
+		 */
+		if ((long) rc_p)
+			printf("oops\n");
+		FAIL_IF((long) rc_p);
+	}
+	printf("done\n");
+
+	return 0;
+}
+
+int main(int argc, char *argv[])
+{
+	return test_harness(test_preempt_vmx, "vmx_preempt");
+}

@@ -2,3 +2,5 @@ fpu_syscall
 vmx_syscall
 fpu_preempt
 vmx_preempt
+fpu_signal
+vmx_signal

@@ -1,4 +1,4 @@
-TEST_PROGS := fpu_syscall fpu_preempt vmx_syscall vmx_preempt
+TEST_PROGS := fpu_syscall fpu_preempt fpu_signal vmx_syscall vmx_preempt vmx_signal
 
 all: $(TEST_PROGS)
 

@@ -7,9 +7,11 @@ $(TEST_PROGS): CFLAGS += -O2 -g -pthread -m64 -maltivec
 
 fpu_syscall: fpu_asm.S
 fpu_preempt: fpu_asm.S
+fpu_signal:  fpu_asm.S
 
 vmx_syscall: vmx_asm.S
 vmx_preempt: vmx_asm.S
+vmx_signal: vmx_asm.S
 
 include ../../lib.mk
 

@@ -0,0 +1,135 @@
+/*
+ * Copyright 2015, Cyril Bur, IBM Corp.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ *
+ * This test attempts to see if the FPU registers are correctly reported in a
+ * signal context. Each worker just spins checking its FPU registers, at some
+ * point a signal will interrupt it and C code will check the signal context
+ * ensuring it is also the same.
+ */
+
+#include <stdio.h>
+#include <unistd.h>
+#include <sys/syscall.h>
+#include <sys/time.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <stdlib.h>
+#include <pthread.h>
+
+#include "utils.h"
+
+/* Number of times each thread should receive the signal */
+#define ITERATIONS 10
+/*
+ * Factor by which to multiply number of online CPUs for total number of
+ * worker threads
+ */
+#define THREAD_FACTOR 8
+
+__thread double darray[] = {0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, 1.0,
+		     1.1, 1.2, 1.3, 1.4, 1.5, 1.6, 1.7, 1.8, 1.9, 2.0,
+		     2.1};
+
+bool bad_context;
+int threads_starting;
+int running;
+
+extern long preempt_fpu(double *darray, int *threads_starting, int *running);
+
+void signal_fpu_sig(int sig, siginfo_t *info, void *context)
+{
+	int i;
+	ucontext_t *uc = context;
+	mcontext_t *mc = &uc->uc_mcontext;
+
+	/* Only the non volatiles were loaded up */
+	for (i = 14; i < 32; i++) {
+		if (mc->fp_regs[i] != darray[i - 14]) {
+			bad_context = true;
+			break;
+		}
+	}
+}
+
+void *signal_fpu_c(void *p)
+{
+	int i;
+	long rc;
+	struct sigaction act;
+	act.sa_sigaction = signal_fpu_sig;
+	act.sa_flags = SA_SIGINFO;
+	rc = sigaction(SIGUSR1, &act, NULL);
+	if (rc)
+		return p;
+
+	srand(pthread_self());
+	for (i = 0; i < 21; i++)
+		darray[i] = rand();
+
+	rc = preempt_fpu(darray, &threads_starting, &running);
+
+	return (void *) rc;
+}
+
+int test_signal_fpu(void)
+{
+	int i, j, rc, threads;
+	void *rc_p;
+	pthread_t *tids;
+
+	threads = sysconf(_SC_NPROCESSORS_ONLN) * THREAD_FACTOR;
+	tids = malloc(threads * sizeof(pthread_t));
+	FAIL_IF(!tids);
+
+	running = true;
+	threads_starting = threads;
+	for (i = 0; i < threads; i++) {
+		rc = pthread_create(&tids[i], NULL, signal_fpu_c, NULL);
+		FAIL_IF(rc);
+	}
+
+	setbuf(stdout, NULL);
+	printf("\tWaiting for all workers to start...");
+	while (threads_starting)
+		asm volatile("": : :"memory");
+	printf("done\n");
+
+	printf("\tSending signals to all threads %d times...", ITERATIONS);
+	for (i = 0; i < ITERATIONS; i++) {
+		for (j = 0; j < threads; j++) {
+			pthread_kill(tids[j], SIGUSR1);
+		}
+		sleep(1);
+	}
+	printf("done\n");
+
+	printf("\tStopping workers...");
+	running = 0;
+	for (i = 0; i < threads; i++) {
+		pthread_join(tids[i], &rc_p);
+
+		/*
+		 * Harness will say the fail was here, look at why signal_fpu
+		 * returned
+		 */
+		if ((long) rc_p || bad_context)
+			printf("oops\n");
+		if (bad_context)
+			fprintf(stderr, "\t!! bad_context is true\n");
+		FAIL_IF((long) rc_p || bad_context);
+	}
+	printf("done\n");
+
+	free(tids);
+	return 0;
+}
+
+int main(int argc, char *argv[])
+{
+	return test_harness(test_signal_fpu, "fpu_signal");
+}

@@ -0,0 +1,138 @@
+/*
+ * Copyright 2015, Cyril Bur, IBM Corp.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ *
+ * This test attempts to see if the VMX registers are correctly reported in a
+ * signal context. Each worker just spins checking its VMX registers, at some
+ * point a signal will interrupt it and C code will check the signal context
+ * ensuring it is also the same.
+ */
+
+#include <stdio.h>
+#include <unistd.h>
+#include <sys/syscall.h>
+#include <sys/time.h>
+#include <sys/types.h>
+#include <sys/wait.h>
+#include <stdlib.h>
+#include <string.h>
+#include <pthread.h>
+
+#include "utils.h"
+
+/* Number of times each thread should receive the signal */
+#define ITERATIONS 10
+/*
+ * Factor by which to multiply number of online CPUs for total number of
+ * worker threads
+ */
+#define THREAD_FACTOR 8
+
+typedef int v4si __attribute__ ((vector_size (16)));
+__thread v4si varray[] = {{1, 2, 3, 4}, {5, 6, 7, 8}, {9, 10,11,12},
+	{13,14,15,16},{17,18,19,20},{21,22,23,24},
+	{25,26,27,28},{29,30,31,32},{33,34,35,36},
+	{37,38,39,40},{41,42,43,44},{45,46,47,48}};
+
+bool bad_context;
+int running;
+int threads_starting;
+
+extern int preempt_vmx(v4si *varray, volatile int *not_ready, int *sentinal);
+
+void signal_vmx_sig(int sig, siginfo_t *info, void *context)
+{
+	int i;
+	ucontext_t *uc = context;
+	mcontext_t *mc = &uc->uc_mcontext;
+
+	/* Only the non volatiles were loaded up */
+	for (i = 20; i < 32; i++) {
+		if (memcmp(mc->v_regs->vrregs[i], &varray[i - 20], 16)) {
+			bad_context = true;
+			break;
+		}
+	}
+}
+
+void *signal_vmx_c(void *p)
+{
+	int i, j;
+	long rc;
+	struct sigaction act;
+	act.sa_sigaction = signal_vmx_sig;
+	act.sa_flags = SA_SIGINFO;
+	rc = sigaction(SIGUSR1, &act, NULL);
+	if (rc)
+		return p;
+
+	srand(pthread_self());
+	for (i = 0; i < 12; i++)
+		for (j = 0; j < 4; j++)
+			varray[i][j] = rand();
+
+	rc = preempt_vmx(varray, &not_ready, &running);
+
+	return (void *) rc;
+}
+
+int test_signal_vmx(void)
+{
+	int i, j, rc, threads;
+	void *rc_p;
+	pthread_t *tids;
+
+	threads = sysconf(_SC_NPROCESSORS_ONLN) * THREAD_FACTOR;
+	tids = malloc(threads * sizeof(pthread_t));
+	FAIL_IF(!tids);
+
+	running = true;
+	not_ready = threads;
+	for (i = 0; i < threads; i++) {
+		rc = pthread_create(&tids[i], NULL, signal_vmx_c, NULL);
+		FAIL_IF(rc);
+	}
+
+	setbuf(stdout, NULL);
+	printf("\tWaiting for all workers to start...");
+	while (not_ready);
+	printf("done\n");
+
+	printf("\tSending signals to all threads %d times...", ITERATIONS);
+	for (i = 0; i < ITERATIONS; i++) {
+		for (j = 0; j < threads; j++) {
+			pthread_kill(tids[j], SIGUSR1);
+		}
+		sleep(1);
+	}
+	printf("done\n");
+
+	printf("\tKilling workers...");
+	running = 0;
+	for (i = 0; i < threads; i++) {
+		pthread_join(tids[i], &rc_p);
+
+		/*
+		 * Harness will say the fail was here, look at why signal_vmx
+		 * returned
+		 */
+		if ((long) rc_p || bad_context)
+			printf("oops\n");
+		if (bad_context)
+			fprintf(stderr, "\t!! bad_context is true\n");
+		FAIL_IF((long) rc_p || bad_context);
+	}
+	printf("done\n");
+
+	free(tids);
+	return 0;
+}
+
+int main(int argc, char *argv[])
+{
+	return test_harness(test_signal_vmx, "vmx_signal");
+}

@@ -22,7 +22,8 @@ SUB_DIRS = benchmarks 		\
 	   switch_endian	\
 	   syscalls		\
 	   tm			\
-	   vphn
+	   vphn         \
+	   math
 
 endif
 

@@ -0,0 +1,30 @@
+#include <ppc-asm.h>
+#include <asm/unistd.h>
+
+#define LOAD_REG_IMMEDIATE(reg,expr) \
+	lis	reg,(expr)@highest;	\
+	ori	reg,reg,(expr)@higher;	\
+	rldicr	reg,reg,32,31;	\
+	oris	reg,reg,(expr)@high;	\
+	ori	reg,reg,(expr)@l;
+
+/* It is very important to note here that _extra is the extra amount of
+ * stack space needed.
+ * This space must be accessed at sp + 32!
+ */
+#define PUSH_BASIC_STACK(_extra) \
+	mflr	r0; \
+	std	r0,16(sp); \
+	stdu	sp,-(_extra + 32)(sp); \
+	mfcr	r0; \
+	stw	r0,8(sp); \
+	std	2,24(sp);
+
+#define POP_BASIC_STACK(_extra) \
+	ld	2,24(sp); \
+	lwz	r0,8(sp); \
+	mtcr	r0; \
+	addi	sp,sp,(_extra + 32); \
+	ld	r0,16(sp); \
+	mtlr	r0;
+

@@ -0,0 +1,2 @@
+fpu_syscall
+vmx_syscall

@@ -0,0 +1,14 @@
+TEST_PROGS := fpu_syscall vmx_syscall
+
+all: $(TEST_PROGS)
+
+$(TEST_PROGS): ../harness.c
+$(TEST_PROGS): CFLAGS += -O2 -g -pthread -m64 -maltivec
+
+fpu_syscall: fpu_asm.S
+vmx_syscall: vmx_asm.S
+
+include ../../lib.mk
+
+clean:
+	rm -f $(TEST_PROGS) *.o

@@ -0,0 +1,161 @@
+/*
+ * Copyright 2015, Cyril Bur, IBM Corp.
+ *
+ * This program is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU General Public License
+ * as published by the Free Software Foundation; either version
+ * 2 of the License, or (at your option) any later version.
+ */
+
+#include "../basic_asm.h"
+
+#define PUSH_FPU(pos) \
+	stfd	f14,pos(sp); \
+	stfd	f15,pos+8(sp); \
+	stfd	f16,pos+16(sp); \
+	stfd	f17,pos+24(sp); \
+	stfd	f18,pos+32(sp); \
+	stfd	f19,pos+40(sp); \
+	stfd	f20,pos+48(sp); \
+	stfd	f21,pos+56(sp); \
+	stfd	f22,pos+64(sp); \
+	stfd	f23,pos+72(sp); \
+	stfd	f24,pos+80(sp); \
+	stfd	f25,pos+88(sp); \
+	stfd	f26,pos+96(sp); \
+	stfd	f27,pos+104(sp); \
+	stfd	f28,pos+112(sp); \
+	stfd	f29,pos+120(sp); \
+	stfd	f30,pos+128(sp); \
+	stfd	f31,pos+136(sp);
+
+#define POP_FPU(pos) \
+	lfd	f14,pos(sp); \
+	lfd	f15,pos+8(sp); \
+	lfd	f16,pos+16(sp); \
+	lfd	f17,pos+24(sp); \
+	lfd	f18,pos+32(sp); \
+	lfd	f19,pos+40(sp); \
+	lfd	f20,pos+48(sp); \
+	lfd	f21,pos+56(sp); \
+	lfd	f22,pos+64(sp); \
+	lfd	f23,pos+72(sp); \
+	lfd	f24,pos+80(sp); \
+	lfd	f25,pos+88(sp); \
+	lfd	f26,pos+96(sp); \
+	lfd	f27,pos+104(sp); \
+	lfd	f28,pos+112(sp); \
+	lfd	f29,pos+120(sp); \
+	lfd	f30,pos+128(sp); \
+	lfd	f31,pos+136(sp);
+
+#Careful calling this, it will 'clobber' fpu (by design)
+#Don't call this from C
+FUNC_START(load_fpu)
+	lfd	f14,0(r3)
+	lfd	f15,8(r3)
+	lfd	f16,16(r3)
+	lfd	f17,24(r3)
+	lfd	f18,32(r3)
+	lfd	f19,40(r3)
+	lfd	f20,48(r3)
+	lfd	f21,56(r3)
+	lfd	f22,64(r3)
+	lfd	f23,72(r3)
+	lfd	f24,80(r3)
+	lfd	f25,88(r3)
+	lfd	f26,96(r3)
+	lfd	f27,104(r3)
+	lfd	f28,112(r3)
+	lfd	f29,120(r3)
+	lfd	f30,128(r3)
+	lfd	f31,136(r3)
+	blr
+FUNC_END(load_fpu)
+
+FUNC_START(check_fpu)
+	mr r4,r3
+	li	r3,1 #assume a bad result
+	lfd	f0,0(r4)
+	fcmpu	cr1,f0,f14
+	bne	cr1,1f
+	lfd	f0,8(r4)
+	fcmpu	cr1,f0,f15
+	bne	cr1,1f
+	lfd	f0,16(r4)
+	fcmpu	cr1,f0,f16
+	bne	cr1,1f
+	lfd	f0,24(r4)
+	fcmpu	cr1,f0,f17
+	bne	cr1,1f
+	lfd	f0,32(r4)
+	fcmpu	cr1,f0,f18
+	bne	cr1,1f
+	lfd	f0,40(r4)
+	fcmpu	cr1,f0,f19
+	bne	cr1,1f
+	lfd	f0,48(r4)
+	fcmpu	cr1,f0,f20
+	bne	cr1,1f
+	lfd	f0,56(r4)
+	fcmpu	cr1,f0,f21
+	bne	cr1,1f
+	lfd	f0,64(r4)
+	fcmpu	cr1,f0,f22
+	bne	cr1,1f
+	lfd	f0,72(r4)
+	fcmpu	cr1,f0,f23
+	bne	cr1,1f
+	lfd	f0,80(r4)
+	fcmpu	cr1,f0,f24
+	bne	cr1,1f
+	lfd	f0,88(r4)
+	fcmpu	cr1,f0,f25
+	bne	cr1,1f
+	lfd	f0,96(r4)
+	fcmpu	cr1,f0,f26
+	bne	cr1,1f
+	lfd	f0,104(r4)
+	fcmpu	cr1,f0,f27
+	bne	cr1,1f
+	lfd	f0,112(r4)
+	fcmpu	cr1,f0,f28
+	bne	cr1,1f
+	lfd	f0,120(r4)
+	fcmpu	cr1,f0,f29
+	bne	cr1,1f
+	lfd	f0,128(r4)
+	fcmpu	cr1,f0,f30
+	bne	cr1,1f
+	lfd	f0,136(r4)
+	fcmpu	cr1,f0,f31
+	bne	cr1,1f
+	li	r3,0 #Sucess!!!
+1:	blr
+
+FUNC_START(test_fpu)
+	#r3 holds pointer to where to put the result of fork
+	#r4 holds pointer to the pid
+	#f14-f31 are non volatiles
+	PUSH_BASIC_STACK(256)
+	std	r3,40(sp) #Address of darray
+	std r4,48(sp) #Address of pid
+	PUSH_FPU(56)
+
+	bl load_fpu
+	nop
+	li	r0,__NR_fork
+	sc
+
+	#pass the result of the fork to the caller
+	ld	r9,48(sp)
+	std	r3,0(r9)
+
+	ld r3,40(sp)
+	bl check_fpu
+	nop
+
+	POP_FPU(56)
+	POP_BASIC_STACK(256)
+	blr
+FUNC_END(test_fpu)