util/bufferiszero: Reorganize for early test for acceleration

Test for length >= 256 inline, where is is often a constant.
Before calling into the accelerated routine, sample three bytes
from the buffer, which handles most non-zero buffers.

Signed-off-by: Alexander Monakov <amonakov@ispras.ru>
Signed-off-by: Mikhail Romanov <mmromanov@ispras.ru>
Message-Id: <20240206204809.9859-3-amonakov@ispras.ru>
[rth: Use __builtin_constant_p; move the indirect call out of line.]
Signed-off-by: Richard Henderson <richard.henderson@linaro.org>

include/qemu/cutils.h

@@ -187,9 +187,39 @@ char *freq_to_str(uint64_t freq_hz);
 /* used to print char* safely */
 #define STR_OR_NULL(str) ((str) ? (str) : "null")
 
-bool buffer_is_zero(const void *buf, size_t len);
+/*
+ * Check if a buffer is all zeroes.
+ */
+
+bool buffer_is_zero_ool(const void *vbuf, size_t len);
+bool buffer_is_zero_ge256(const void *vbuf, size_t len);
 bool test_buffer_is_zero_next_accel(void);
 
+static inline bool buffer_is_zero_sample3(const char *buf, size_t len)
+{
+    /*
+     * For any reasonably sized buffer, these three samples come from
+     * three different cachelines.  In qemu-img usage, we find that
+     * each byte eliminates more than half of all buffer testing.
+     * It is therefore critical to performance that the byte tests
+     * short-circuit, so that we do not pull in additional cache lines.
+     * Do not "optimize" this to !(a | b | c).
+     */
+    return !buf[0] && !buf[len - 1] && !buf[len / 2];
+}
+
+#ifdef __OPTIMIZE__
+static inline bool buffer_is_zero(const void *buf, size_t len)
+{
+    return (__builtin_constant_p(len) && len >= 256
+            ? buffer_is_zero_sample3(buf, len) &&
+              buffer_is_zero_ge256(buf, len)
+            : buffer_is_zero_ool(buf, len));
+}
+#else
+#define buffer_is_zero  buffer_is_zero_ool
+#endif
+
 /*
  * Implementation of ULEB128 (http://en.wikipedia.org/wiki/LEB128)
  * Input is limited to 14-bit numbers

util/bufferiszero.c

@@ -26,8 +26,9 @@
 #include "qemu/bswap.h"
 #include "host/cpuinfo.h"
 
-static bool
-buffer_zero_int(const void *buf, size_t len)
+static bool (*buffer_is_zero_accel)(const void *, size_t);
+
+static bool buffer_is_zero_integer(const void *buf, size_t len)
 {
     if (unlikely(len < 8)) {
         /* For a very small buffer, simply accumulate all the bytes.  */
@@ -128,60 +129,38 @@ buffer_zero_avx2(const void *buf, size_t len)
 }
 #endif /* CONFIG_AVX2_OPT */
 
-/*
- * Make sure that these variables are appropriately initialized when
- * SSE2 is enabled on the compiler command-line, but the compiler is
- * too old to support CONFIG_AVX2_OPT.
- */
-#if defined(CONFIG_AVX2_OPT)
-# define INIT_USED     0
-# define INIT_LENGTH   0
-# define INIT_ACCEL    buffer_zero_int
-#else
-# ifndef __SSE2__
-#  error "ISA selection confusion"
-# endif
-# define INIT_USED     CPUINFO_SSE2
-# define INIT_LENGTH   64
-# define INIT_ACCEL    buffer_zero_sse2
-#endif
-
-static unsigned used_accel = INIT_USED;
-static unsigned length_to_accel = INIT_LENGTH;
-static bool (*buffer_accel)(const void *, size_t) = INIT_ACCEL;
-
 static unsigned __attribute__((noinline))
 select_accel_cpuinfo(unsigned info)
 {
     /* Array is sorted in order of algorithm preference. */
     static const struct {
         unsigned bit;
-        unsigned len;
         bool (*fn)(const void *, size_t);
     } all[] = {
 #ifdef CONFIG_AVX2_OPT
-        { CPUINFO_AVX2,    128, buffer_zero_avx2 },
+        { CPUINFO_AVX2,    buffer_zero_avx2 },
 #endif
-        { CPUINFO_SSE2,     64, buffer_zero_sse2 },
-        { CPUINFO_ALWAYS,    0, buffer_zero_int },
+        { CPUINFO_SSE2,    buffer_zero_sse2 },
+        { CPUINFO_ALWAYS,  buffer_is_zero_integer },
     };
 
     for (unsigned i = 0; i < ARRAY_SIZE(all); ++i) {
         if (info & all[i].bit) {
-            length_to_accel = all[i].len;
-            buffer_accel = all[i].fn;
+            buffer_is_zero_accel = all[i].fn;
             return all[i].bit;
         }
     }
     return 0;
 }
 
-#if defined(CONFIG_AVX2_OPT)
+static unsigned used_accel;
+
 static void __attribute__((constructor)) init_accel(void)
 {
     used_accel = select_accel_cpuinfo(cpuinfo_init());
 }
-#endif /* CONFIG_AVX2_OPT */
+
+#define INIT_ACCEL NULL
 
 bool test_buffer_is_zero_next_accel(void)
 {
@@ -194,36 +173,37 @@ bool test_buffer_is_zero_next_accel(void)
     used_accel |= used;
     return used;
 }
-
-static bool select_accel_fn(const void *buf, size_t len)
-{
-    if (likely(len >= length_to_accel)) {
-        return buffer_accel(buf, len);
-    }
-    return buffer_zero_int(buf, len);
-}
-
 #else
-#define select_accel_fn  buffer_zero_int
 bool test_buffer_is_zero_next_accel(void)
 {
     return false;
 }
+
+#define INIT_ACCEL buffer_is_zero_integer
 #endif
 
-/*
- * Checks if a buffer is all zeroes
- */
-bool buffer_is_zero(const void *buf, size_t len)
+static bool (*buffer_is_zero_accel)(const void *, size_t) = INIT_ACCEL;
+
+bool buffer_is_zero_ool(const void *buf, size_t len)
 {
     if (unlikely(len == 0)) {
         return true;
     }
+    if (!buffer_is_zero_sample3(buf, len)) {
+        return false;
+    }
+    /* All bytes are covered for any len <= 3.  */
+    if (unlikely(len <= 3)) {
+        return true;
+    }
 
-    /* Fetch the beginning of the buffer while we select the accelerator.  */
-    __builtin_prefetch(buf);
-
-    /* Use an optimized zero check if possible.  Note that this also
-       includes a check for an unrolled loop over 64-bit integers.  */
-    return select_accel_fn(buf, len);
+    if (likely(len >= 256)) {
+        return buffer_is_zero_accel(buf, len);
+    }
+    return buffer_is_zero_integer(buf, len);
+}
+
+bool buffer_is_zero_ge256(const void *buf, size_t len)
+{
+    return buffer_is_zero_accel(buf, len);
 }