git: 0be13a45f7cb - main - xregs_sig: Add a ability to handle multiple register banks
- Go to: [ bottom of page ] [ top of archives ] [ this month ]
Date: Sun, 04 Jun 2023 09:08:36 UTC
The branch main has been updated by dchagin:
URL: https://cgit.FreeBSD.org/src/commit/?id=0be13a45f7cb18f8b86c800a28457bf918e130f2
commit 0be13a45f7cb18f8b86c800a28457bf918e130f2
Author: Dmitry Chagin <dchagin@FreeBSD.org>
AuthorDate: 2023-06-04 09:06:25 +0000
Commit: Dmitry Chagin <dchagin@FreeBSD.org>
CommitDate: 2023-06-04 09:06:25 +0000
xregs_sig: Add a ability to handle multiple register banks
While here replace the .global directive by traditionally used .globl.
Reviewed by: kib
Differential Revision: https://reviews.freebsd.org/D40410
---
tools/test/xregs_sig/c2x2c_aarch64.S | 18 ++---
tools/test/xregs_sig/c2x2c_amd64.S | 18 ++---
tools/test/xregs_sig/xregs_sig.c | 128 ++++++++++++++++++++++-------------
3 files changed, 100 insertions(+), 64 deletions(-)
diff --git a/tools/test/xregs_sig/c2x2c_aarch64.S b/tools/test/xregs_sig/c2x2c_aarch64.S
index 6f1637f5bcdf..f53550b867fd 100644
--- a/tools/test/xregs_sig/c2x2c_aarch64.S
+++ b/tools/test/xregs_sig/c2x2c_aarch64.S
@@ -5,9 +5,11 @@
* $FreeBSD$
*/
- .global cpu_to_xregs
- .type cpu_to_xregs, @function
-cpu_to_xregs:
+ .text
+
+ .globl cpu_to_vfp
+ .type cpu_to_vfp, @function
+cpu_to_vfp:
stp q0, q1, [x0, #( 0 * 16)]
stp q2, q3, [x0, #( 2 * 16)]
stp q4, q5, [x0, #( 4 * 16)]
@@ -26,12 +28,12 @@ cpu_to_xregs:
stp q30, q31, [x0, #(30 * 16)]
ret
- .size cpu_to_xregs, . - cpu_to_xregs
+ .size cpu_to_vfp, . - cpu_to_vfp
- .global xregs_to_cpu
- .type xregs_to_cpu, @function
-xregs_to_cpu:
+ .globl vfp_to_cpu
+ .type vfp_to_cpu, @function
+vfp_to_cpu:
ldp q0, q1, [x0, #( 0 * 16)]
ldp q2, q3, [x0, #( 2 * 16)]
ldp q4, q5, [x0, #( 4 * 16)]
@@ -50,6 +52,6 @@ xregs_to_cpu:
ldp q30, q31, [x0, #(30 * 16)]
ret
- .size xregs_to_cpu, . - xregs_to_cpu
+ .size vfp_to_cpu, . - vfp_to_cpu
.section .note.GNU-stack,"",@progbits
diff --git a/tools/test/xregs_sig/c2x2c_amd64.S b/tools/test/xregs_sig/c2x2c_amd64.S
index e5ae5af99e97..7348e4871d53 100644
--- a/tools/test/xregs_sig/c2x2c_amd64.S
+++ b/tools/test/xregs_sig/c2x2c_amd64.S
@@ -5,9 +5,11 @@
* $FreeBSD$
*/
- .global cpu_to_xregs
- .type cpu_to_xregs, @function
-cpu_to_xregs:
+ .text
+
+ .globl cpu_to_xmm
+ .type cpu_to_xmm, @function
+cpu_to_xmm:
movdqu %xmm0, (%rdi)
movdqu %xmm1, 1 * 16(%rdi)
movdqu %xmm2, 2 * 16(%rdi)
@@ -26,12 +28,12 @@ cpu_to_xregs:
movdqu %xmm15, 15 * 16(%rdi)
retq
- .size cpu_to_xregs, . - cpu_to_xregs
+ .size cpu_to_xmm, . - cpu_to_xmm
- .global xregs_to_cpu
- .type xregs_to_cpu, @function
-xregs_to_cpu:
+ .globl xmm_to_cpu
+ .type xmm_to_cpu, @function
+xmm_to_cpu:
movdqu (%rdi), %xmm0
movdqu 1 * 16(%rdi), %xmm1
movdqu 2 * 16(%rdi), %xmm2
@@ -50,6 +52,6 @@ xregs_to_cpu:
movdqu 15 * 16(%rdi), %xmm15
retq
- .size xregs_to_cpu, . - xregs_to_cpu
+ .size xmm_to_cpu, . - xmm_to_cpu
.section .note.GNU-stack,"",@progbits
diff --git a/tools/test/xregs_sig/xregs_sig.c b/tools/test/xregs_sig/xregs_sig.c
index a67379a0080e..4e29165d39e2 100644
--- a/tools/test/xregs_sig/xregs_sig.c
+++ b/tools/test/xregs_sig/xregs_sig.c
@@ -44,30 +44,49 @@
#ifndef __unused
#define __unused __attribute__((__unused__))
#endif
-#ifndef nitems
-#define nitems(x) (sizeof((x)) / sizeof((x)[0]))
-#endif
-#if defined(__amd64__)
-#define XREGSRNAM "xmm"
-#define NREGS 16
-#elif defined(__aarch64__)
-#define XREGSRNAM "q"
-#define NREGS 32
-#endif
-
-struct xregsreg {
- uint8_t xregs_bytes[16];
+struct xregs_bank {
+ const char *b_name;
+ const char *r_name;
+ uint32_t regs;
+ uint32_t bytes;
+ void (*x2c)(uint8_t *);
+ void (*c2x)(uint8_t *);
};
-struct xregs {
- struct xregsreg xregsreg[NREGS];
+#if defined(__amd64__)
+void cpu_to_xmm(uint8_t *);
+void xmm_to_cpu(uint8_t *);
+
+static const struct xregs_bank xregs_banks[] = {
+ {
+ .b_name = "SSE",
+ .r_name = "xmm",
+ .regs = 16,
+ .bytes = 16,
+ .x2c = xmm_to_cpu,
+ .c2x = cpu_to_xmm,
+ },
};
-
-void cpu_to_xregs(struct xregs *xregs);
-void xregs_to_cpu(struct xregs *xregs);
+#elif defined(__aarch64__)
+void cpu_to_vfp(uint8_t *);
+void vfp_to_cpu(uint8_t *);
+
+static const struct xregs_bank xregs_banks[] = {
+ {
+ .b_name = "VFP",
+ .r_name = "q",
+ .regs = 32,
+ .bytes = 16,
+ .x2c = vfp_to_cpu,
+ .c2x = cpu_to_vfp,
+ },
+};
+#endif
static atomic_uint sigs;
+static int max_bank_idx;
+
static void
sigusr1_handler(int sig __unused, siginfo_t *si __unused, void *m __unused)
@@ -87,23 +106,22 @@ sigalrm_handler(int sig __unused)
alarm(TIMO);
}
-static struct xregs zero_xregs = {};
static void
-fill_xregs(struct xregs *xregs)
+fill_xregs(uint8_t *xregs, int bank)
{
- arc4random_buf(xregs, sizeof(*xregs));
+ arc4random_buf(xregs, xregs_banks[bank].regs * xregs_banks[bank].bytes);
}
static void
-dump_xregs(const struct xregsreg *r)
+dump_xregs(const uint8_t *r, int bank)
{
unsigned k;
- for (k = 0; k < nitems(r->xregs_bytes); k++) {
+ for (k = 0; k < xregs_banks[bank].bytes; k++) {
if (k != 0)
printf(" ");
- printf("%02x", r->xregs_bytes[k]);
+ printf("%02x", r[k]);
}
printf("\n");
}
@@ -111,9 +129,9 @@ dump_xregs(const struct xregsreg *r)
static pthread_mutex_t show_lock;
static void
-show_diff(const struct xregs *xregs1, const struct xregs *xregs2)
+show_diff(const uint8_t *xregs1, const uint8_t *xregs2, int bank)
{
- const struct xregsreg *r1, *r2;
+ const uint8_t *r1, *r2;
unsigned i, j;
#if defined(__FreeBSD__)
@@ -121,14 +139,14 @@ show_diff(const struct xregs *xregs1, const struct xregs *xregs2)
#elif defined(__linux__)
printf("thr %ld\n", syscall(SYS_gettid));
#endif
- for (i = 0; i < nitems(xregs1->xregsreg); i++) {
- r1 = &xregs1->xregsreg[i];
- r2 = &xregs2->xregsreg[i];
- for (j = 0; j < nitems(r1->xregs_bytes); j++) {
- if (r1->xregs_bytes[j] != r2->xregs_bytes[j]) {
- printf("%%%s%u\n", XREGSRNAM, i);
- dump_xregs(r1);
- dump_xregs(r2);
+ for (i = 0; i < xregs_banks[bank].regs; i++) {
+ r1 = xregs1 + i * xregs_banks[bank].bytes;
+ r2 = xregs2 + i * xregs_banks[bank].bytes;
+ for (j = 0; j < xregs_banks[bank].bytes; j++) {
+ if (r1[j] != r2[j]) {
+ printf("%%%s%u\n", xregs_banks[bank].r_name, i);
+ dump_xregs(r1, bank);
+ dump_xregs(r2, bank);
break;
}
}
@@ -142,28 +160,32 @@ my_pause(void)
}
static void *
-worker_thread(void *arg __unused)
+worker_thread(void *arg)
{
- struct xregs xregs, xregs_cpu;
+ int bank = (uintptr_t)arg;
+ int sz = xregs_banks[bank].regs * xregs_banks[bank].bytes;
+ uint8_t xregs[sz], xregs_cpu[sz], zero_xregs[sz];
+
+ memset(zero_xregs, 0, sz);
- fill_xregs(&xregs);
+ fill_xregs(xregs, bank);
for (;;) {
- xregs_to_cpu(&xregs);
+ xregs_banks[bank].x2c(xregs);
my_pause();
- cpu_to_xregs(&xregs_cpu);
- if (memcmp(&xregs, &xregs_cpu, sizeof(struct xregs)) != 0) {
+ xregs_banks[bank].c2x(xregs_cpu);
+ if (memcmp(xregs, xregs_cpu, sz) != 0) {
pthread_mutex_lock(&show_lock);
- show_diff(&xregs, &xregs_cpu);
+ show_diff(xregs, xregs_cpu, bank);
abort();
pthread_mutex_unlock(&show_lock);
}
- xregs_to_cpu(&zero_xregs);
+ xregs_banks[bank].x2c(zero_xregs);
my_pause();
- cpu_to_xregs(&xregs_cpu);
- if (memcmp(&zero_xregs, &xregs_cpu, sizeof(struct xregs)) != 0) {
+ xregs_banks[bank].c2x(xregs_cpu);
+ if (memcmp(zero_xregs, xregs_cpu, sz) != 0) {
pthread_mutex_lock(&show_lock);
- show_diff(&zero_xregs, &xregs_cpu);
+ show_diff(zero_xregs, xregs_cpu, bank);
abort();
pthread_mutex_unlock(&show_lock);
}
@@ -175,7 +197,9 @@ int
main(void)
{
struct sigaction sa;
- int error, i, ncpu;
+ int error, i, ncpu, bank;
+
+ max_bank_idx = 0;
bzero(&sa, sizeof(sa));
sa.sa_handler = sigalrm_handler;
@@ -199,14 +223,22 @@ main(void)
}
ncpu = sysconf(_SC_NPROCESSORS_ONLN);
- ncpu *= 2;
+ if (max_bank_idx == 0)
+ ncpu *= 2;
+ bank = 0;
pthread_t wt[ncpu];
+nextbank:
+ printf("Starting %d threads for registers bank %s sized [%d][%d]\n", ncpu,
+ xregs_banks[bank].b_name, xregs_banks[bank].regs, xregs_banks[bank].bytes);
for (i = 0; i < ncpu; i++) {
- error = pthread_create(&wt[i], NULL, worker_thread, NULL);
+ error = pthread_create(&wt[i], NULL, worker_thread,
+ (void *)(uintptr_t)bank);
if (error != 0) {
fprintf(stderr, "pthread_create %s\n", strerror(error));
}
}
+ if (++bank <= max_bank_idx)
+ goto nextbank;
alarm(TIMO);
for (;;) {