[v3,19/25] aarch64: Reorg sve reginfo to save space

Message ID 20200522023440.26261-20-richard.henderson@linaro.org
State New
Headers show
Series
  • risu cleanups and improvements
Related show

Commit Message

Richard Henderson May 22, 2020, 2:34 a.m.
Mirror the signal frame by storing all of the registers
as a lump.  Use the signal macros to pull out the values.

Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

---
 risu_reginfo_aarch64.h |  16 +----
 risu_reginfo_aarch64.c | 135 +++++++++++++++++++++--------------------
 2 files changed, 73 insertions(+), 78 deletions(-)

-- 
2.20.1

Comments

Alex Bennée June 23, 2020, 4:32 p.m. | #1
Richard Henderson <richard.henderson@linaro.org> writes:

> Mirror the signal frame by storing all of the registers

> as a lump.  Use the signal macros to pull out the values.

>

> Signed-off-by: Richard Henderson <richard.henderson@linaro.org>


Reviewed-by: Alex Bennée <alex.bennee@linaro.org>


-- 
Alex Bennée

Patch

diff --git a/risu_reginfo_aarch64.h b/risu_reginfo_aarch64.h
index c33b86f..01076b4 100644
--- a/risu_reginfo_aarch64.h
+++ b/risu_reginfo_aarch64.h
@@ -17,20 +17,8 @@ 
 
 struct simd_reginfo {
     __uint128_t vregs[32];
-    char end[0];
 };
 
-#ifdef SVE_MAGIC
-struct sve_reginfo {
-    /* SVE */
-    uint16_t    vl; /* current VL */
-    __uint128_t zregs[SVE_NUM_ZREGS][SVE_VQ_MAX];
-    uint16_t    pregs[SVE_NUM_PREGS][SVE_VQ_MAX];
-    uint16_t    ffr[SVE_VQ_MAX];
-    char end[0];
-};
-#endif
-
 /* The kernel headers set this based on future arch extensions.
    The current arch maximum is 16.  Save space below.  */
 #undef SVE_VQ_MAX
@@ -47,11 +35,13 @@  struct reginfo {
     /* FP/SIMD */
     uint32_t fpsr;
     uint32_t fpcr;
+    uint32_t sve_vl;
 
     union {
         struct simd_reginfo simd;
 #ifdef SVE_MAGIC
-        struct sve_reginfo sve;
+        char sve[SVE_SIG_CONTEXT_SIZE(16) - SVE_SIG_REGS_OFFSET]
+            __attribute__((aligned(16)));
 #endif
     };
 };
diff --git a/risu_reginfo_aarch64.c b/risu_reginfo_aarch64.c
index 7044648..a1020ac 100644
--- a/risu_reginfo_aarch64.c
+++ b/risu_reginfo_aarch64.c
@@ -71,15 +71,30 @@  void process_arch_opt(int opt, const char *arg)
 
 int reginfo_size(struct reginfo *ri)
 {
-    int size = offsetof(struct reginfo, simd.end);
 #ifdef SVE_MAGIC
-    if (test_sve) {
-        size = offsetof(struct reginfo, sve.end);
+    if (ri->sve_vl) {
+        int vq = sve_vq_from_vl(ri->sve_vl);
+        return (offsetof(struct reginfo, sve) +
+                SVE_SIG_CONTEXT_SIZE(vq) - SVE_SIG_REGS_OFFSET);
     }
 #endif
-    return size;
+    return offsetof(struct reginfo, simd) + sizeof(ri->simd);
 }
 
+#ifdef SVE_MAGIC
+static uint64_t *reginfo_zreg(struct reginfo *ri, int vq, int i)
+{
+    return (uint64_t *)(ri->sve + SVE_SIG_ZREG_OFFSET(vq, i) -
+                        SVE_SIG_REGS_OFFSET);
+}
+
+static uint16_t *reginfo_preg(struct reginfo *ri, int vq, int i)
+{
+    return (uint16_t *)(ri->sve + SVE_SIG_PREG_OFFSET(vq, i) -
+                        SVE_SIG_REGS_OFFSET);
+}
+#endif
+
 /* reginfo_init: initialize with a ucontext */
 void reginfo_init(struct reginfo *ri, ucontext_t *uc)
 {
@@ -152,8 +167,6 @@  void reginfo_init(struct reginfo *ri, ucontext_t *uc)
             return;
         }
 
-        ri->sve.vl = sve->vl;
-
         if (sve->head.size < SVE_SIG_CONTEXT_SIZE(vq)) {
             if (sve->head.size == sizeof(*sve)) {
                 /* SVE state is empty -- not an error.  */
@@ -164,24 +177,9 @@  void reginfo_init(struct reginfo *ri, ucontext_t *uc)
             return;
         }
 
-        /* Copy ZREG's one at a time */
-        for (i = 0; i < SVE_NUM_ZREGS; i++) {
-            memcpy(&ri->sve.zregs[i],
-                   (void *)sve + SVE_SIG_ZREG_OFFSET(vq, i),
-                   SVE_SIG_ZREG_SIZE(vq));
-        }
-
-        /* Copy PREG's one at a time */
-        for (i = 0; i < SVE_NUM_PREGS; i++) {
-            memcpy(&ri->sve.pregs[i],
-                   (void *)sve + SVE_SIG_PREG_OFFSET(vq, i),
-                   SVE_SIG_PREG_SIZE(vq));
-        }
-
-        /* Finally the FFR */
-        memcpy(&ri->sve.ffr, (void *)sve + SVE_SIG_FFR_OFFSET(vq),
-               SVE_SIG_FFR_SIZE(vq));
-
+        ri->sve_vl = sve->vl;
+        memcpy(ri->sve, (char *)sve + SVE_SIG_REGS_OFFSET,
+               SVE_SIG_CONTEXT_SIZE(vq) - SVE_SIG_REGS_OFFSET);
         return;
     }
 #endif /* SVE_MAGIC */
@@ -225,18 +223,20 @@  static void sve_dump_preg_diff(FILE *f, int vq, const uint16_t *p1,
     fprintf(f, "\n");
 }
 
-static void sve_dump_zreg_diff(FILE *f, int vq, const __uint128_t *z1,
-                               const __uint128_t *z2)
+static void sve_dump_zreg_diff(FILE *f, int vq, const uint64_t *za,
+                               const uint64_t *zb)
 {
     const char *pad = "";
     int q;
 
     for (q = 0; q < vq; ++q) {
-        if (z1[q] != z2[q]) {
+        uint64_t za0 = za[2 * q], za1 = za[2 * q + 1];
+        uint64_t zb0 = zb[2 * q], zb1 = zb[2 * q + 1];
+
+        if (za0 != zb0 || za1 != zb1) {
             fprintf(f, "%sq%-2d: %016" PRIx64 "%016" PRIx64
-                    " vs %016" PRIx64 "%016" PRIx64"\n", pad, q,
-                    (uint64_t)(z1[q] >> 64), (uint64_t)z1[q],
-                    (uint64_t)(z2[q] >> 64), (uint64_t)z2[q]);
+                    " vs %016" PRIx64 "%016" PRIx64"\n",
+                    pad, q, za1, za0, zb1, zb0);
             pad = "      ";
         }
     }
@@ -263,28 +263,30 @@  int reginfo_dump(struct reginfo *ri, FILE * f)
     if (test_sve) {
         int q, vq = test_sve;
 
-        fprintf(f, "  vl     : %d\n", ri->sve.vl);
+        fprintf(f, "  vl     : %d\n", ri->sve_vl);
 
-        for (i = 0; i < 32; i++) {
-            fprintf(f, "  Z%-2d q%-2d: %016" PRIx64 "%016" PRIx64 "\n", i, 0,
-                    (uint64_t)(ri->sve.zregs[i][0] >> 64),
-                    (uint64_t)ri->sve.zregs[i][0]);
+        for (i = 0; i < SVE_NUM_ZREGS; i++) {
+            uint64_t *z = reginfo_zreg(ri, vq, i);
+
+            fprintf(f, "  Z%-2d q%-2d: %016" PRIx64 "%016" PRIx64 "\n",
+                    i, 0, z[1], z[0]);
             for (q = 1; q < vq; ++q) {
-                fprintf(f, "      q%-2d: %016" PRIx64 "%016" PRIx64 "\n", q,
-                        (uint64_t)(ri->sve.zregs[i][q] >> 64),
-                        (uint64_t)ri->sve.zregs[i][q]);
+                fprintf(f, "      q%-2d: %016" PRIx64 "%016" PRIx64 "\n",
+                        q, z[q * 2 + 1], z[q * 2]);
             }
         }
 
-        for (i = 0; i < 16; i++) {
-            fprintf(f, "  P%-2d    : ", i);
-            sve_dump_preg(f, vq, &ri->sve.pregs[i][0]);
+        for (i = 0; i < SVE_NUM_PREGS + 1; i++) {
+            uint16_t *p = reginfo_preg(ri, vq, i);
+
+            if (i == SVE_NUM_PREGS) {
+                fprintf(f, "  FFR    : ");
+            } else {
+                fprintf(f, "  P%-2d    : ", i);
+            }
+            sve_dump_preg(f, vq, p);
             fprintf(f, "\n");
         }
-        fprintf(f, "  FFR    : ");
-        sve_dump_preg(f, vq, &ri->sve.ffr[0]);
-        fprintf(f, "\n");
-
         return !ferror(f);
     }
 #endif
@@ -338,31 +340,34 @@  int reginfo_dump_mismatch(struct reginfo *m, struct reginfo *a, FILE * f)
 
 #ifdef SVE_MAGIC
     if (test_sve) {
-        int vq = sve_vq_from_vl(m->sve.vl);
+        int vq = sve_vq_from_vl(m->sve_vl);
 
-        if (m->sve.vl != a->sve.vl) {
-            fprintf(f, "  vl    : %d vs %d\n", m->sve.vl, a->sve.vl);
+        if (m->sve_vl != a->sve_vl) {
+            fprintf(f, "  vl    : %d vs %d\n", m->sve_vl, a->sve_vl);
         }
 
         for (i = 0; i < SVE_NUM_ZREGS; i++) {
-            if (!sve_zreg_is_eq(vq, &m->sve.zregs[i], &a->sve.zregs[i])) {
-                fprintf(f, "  Z%-2d ", i);
-                sve_dump_zreg_diff(f, vq, &m->sve.zregs[i][0],
-                                   &a->sve.zregs[i][0]);
-            }
-        }
-        for (i = 0; i < SVE_NUM_PREGS; i++) {
-            if (!sve_preg_is_eq(vq, &m->sve.pregs[i], &a->sve.pregs[i])) {
-                fprintf(f, "  P%-2d    : ", i);
-                sve_dump_preg_diff(f, vq, &m->sve.pregs[i][0],
-                                   &a->sve.pregs[i][0]);
-            }
-        }
-        if (!sve_preg_is_eq(vq, &m->sve.ffr, &a->sve.ffr)) {
-            fprintf(f, "  FFR   : ");
-            sve_dump_preg_diff(f, vq, &m->sve.pregs[i][0], &a->sve.pregs[i][0]);
-        }
+            uint64_t *zm = reginfo_zreg(m, vq, i);
+            uint64_t *za = reginfo_zreg(a, vq, i);
 
+            if (!sve_zreg_is_eq(vq, zm, za)) {
+                fprintf(f, "  Z%-2d ", i);
+                sve_dump_zreg_diff(f, vq, zm, za);
+            }
+        }
+        for (i = 0; i < SVE_NUM_PREGS + 1; i++) {
+            uint16_t *pm = reginfo_preg(m, vq, i);
+            uint16_t *pa = reginfo_preg(a, vq, i);
+
+            if (!sve_preg_is_eq(vq, pm, pa)) {
+                if (i == SVE_NUM_PREGS) {
+                    fprintf(f, "  FFR   : ");
+                } else {
+                    fprintf(f, "  P%-2d    : ", i);
+                }
+                sve_dump_preg_diff(f, vq, pm, pa);
+            }
+        }
         return !ferror(f);
     }
 #endif