[Qemu-devel] [PATCH 10/10] cutils: Rewrite x86 buffer zero checking

[Paolo Bonzini <pbonzini@redhat.com>]

From: Richard Henderson <rth@twiddle.net>

Handle alignment of buffers, so that the vector paths can be
used more often.  Add versions for AVX1 and SSE4.1, both of
which have incremental improvements over SSE2.

Signed-off-by: Richard Henderson <rth@twiddle.net>
Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 util/bufferiszero.c | 139 ++++++++++++++++++++++++++++++++++++----------------
 1 file changed, 96 insertions(+), 43 deletions(-)

diff --git a/util/bufferiszero.c b/util/bufferiszero.c
index 4d8a8c8..c23589c 100644
--- a/util/bufferiszero.c
+++ b/util/bufferiszero.c
@@ -26,38 +26,6 @@
 #include "qemu/cutils.h"
 #include "qemu/bswap.h"
 
-
-/* vector definitions */
-
-extern void link_error(void);
-
-#define ACCEL_BUFFER_ZERO(NAME, SIZE, VECTYPE, NONZERO)         \
-static bool NAME(const void *buf, size_t len)                   \
-{                                                               \
-    const void *end = buf + len;                                \
-    do {                                                        \
-        const VECTYPE *p = buf;                                 \
-        VECTYPE t;                                              \
-        __builtin_prefetch(buf + SIZE);                         \
-        barrier();                                              \
-        if (SIZE == sizeof(VECTYPE) * 4) {                      \
-            t = (p[0] | p[1]) | (p[2] | p[3]);                  \
-        } else if (SIZE == sizeof(VECTYPE) * 8) {               \
-            t  = p[0] | p[1];                                   \
-            t |= p[2] | p[3];                                   \
-            t |= p[4] | p[5];                                   \
-            t |= p[6] | p[7];                                   \
-        } else {                                                \
-            link_error();                                       \
-        }                                                       \
-        if (unlikely(NONZERO(t))) {                             \
-            return false;                                       \
-        }                                                       \
-        buf += SIZE;                                            \
-    } while (buf < end);                                        \
-    return true;                                                \
-}
-
 static bool
 buffer_zero_int(const void *buf, size_t len)
 {
@@ -102,24 +70,110 @@ buffer_zero_int(const void *buf, size_t len)
 #pragma GCC push_options
 #pragma GCC target("sse2")
 #include <emmintrin.h>
-#define SSE2_NONZERO(X) \
-    (_mm_movemask_epi8(_mm_cmpeq_epi8((X), _mm_setzero_si128())) != 0xFFFF)
-ACCEL_BUFFER_ZERO(buffer_zero_sse2, 64, __m128i, SSE2_NONZERO)
+
+static bool
+buffer_zero_sse2(const void *buf, size_t len)
+{
+    __m128i t = _mm_loadu_si128(buf);
+    __m128i *p = (__m128i *)(((uintptr_t)buf + 5 * 16) & -16);
+    __m128i *e = (__m128i *)(((uintptr_t)buf + len) & -16);
+    __m128i zero = _mm_setzero_si128();
+
+    /* Loop over 16-byte aligned blocks of 64.  */
+    while (likely(p <= e)) {
+        __builtin_prefetch(p);
+        t = _mm_cmpeq_epi8(t, zero);
+        if (unlikely(_mm_movemask_epi8(t) != 0xFFFF)) {
+            return false;
+        }
+        t = p[-4] | p[-3] | p[-2] | p[-1];
+        p += 4;
+    }
+
+    /* Finish the aligned tail.  */
+    t |= e[-3];
+    t |= e[-2];
+    t |= e[-1];
+
+    /* Finish the unaligned tail.  */
+    t |= _mm_loadu_si128(buf + len - 16);
+
+    return _mm_movemask_epi8(_mm_cmpeq_epi8(t, zero)) == 0xFFFF;
+}
 #pragma GCC pop_options
 
 #ifdef CONFIG_AVX2_OPT
 #pragma GCC push_options
 #pragma GCC target("sse4")
 #include <smmintrin.h>
-#define SSE4_NONZERO(X)  !_mm_testz_si128((X), (X))
-ACCEL_BUFFER_ZERO(buffer_zero_sse4, 64, __m128i, SSE4_NONZERO)
+
+static bool
+buffer_zero_sse4(const void *buf, size_t len)
+{
+    __m128i t = _mm_loadu_si128(buf);
+    __m128i *p = (__m128i *)(((uintptr_t)buf + 5 * 16) & -16);
+    __m128i *e = (__m128i *)(((uintptr_t)buf + len) & -16);
+
+    /* Loop over 16-byte aligned blocks of 64.  */
+    while (likely(p <= e)) {
+        __builtin_prefetch(p);
+        if (unlikely(!_mm_testz_si128(t, t))) {
+            return false;
+        }
+        t = p[-4] | p[-3] | p[-2] | p[-1];
+        p += 4;
+    }
+
+    /* Finish the aligned tail.  */
+    t |= e[-3];
+    t |= e[-2];
+    t |= e[-1];
+
+    /* Finish the unaligned tail.  */
+    t |= _mm_loadu_si128(buf + len - 16);
+
+    return _mm_testz_si128(t, t);
+}
 #pragma GCC pop_options
 
 #pragma GCC push_options
 #pragma GCC target("avx2")
 #include <immintrin.h>
-#define AVX2_NONZERO(X)  !_mm256_testz_si256((X), (X))
-ACCEL_BUFFER_ZERO(buffer_zero_avx2, 128, __m256i, AVX2_NONZERO)
+
+static bool
+buffer_zero_avx2(const void *buf, size_t len)
+{
+    /* Begin with an unaligned head of 32 bytes.  */
+    __m256i t = _mm256_loadu_si256(buf);
+    __m256i *p = (__m256i *)(((uintptr_t)buf + 5 * 32) & -32);
+    __m256i *e = (__m256i *)(((uintptr_t)buf + len) & -32);
+
+    if (likely(p <= e)) {
+        /* Loop over 32-byte aligned blocks of 128.  */
+        do {
+            __builtin_prefetch(p);
+            if (unlikely(!_mm256_testz_si256(t, t))) {
+                return false;
+            }
+            t = p[-4] | p[-3] | p[-2] | p[-1];
+            p += 4;
+        } while (p <= e);
+    } else {
+        t |= _mm256_loadu_si256(buf + 32);
+        if (len <= 128) {
+            goto last2;
+        }
+    }
+
+    /* Finish the last block of 128 unaligned.  */
+    t |= _mm256_loadu_si256(buf + len - 4 * 32);
+    t |= _mm256_loadu_si256(buf + len - 3 * 32);
+ last2:
+    t |= _mm256_loadu_si256(buf + len - 2 * 32);
+    t |= _mm256_loadu_si256(buf + len - 1 * 32);
+
+    return _mm256_testz_si256(t, t);
+}
 #pragma GCC pop_options
 #endif
 
@@ -177,16 +231,15 @@ bool test_buffer_is_zero_next_accel(void)
 
 static bool select_accel_fn(const void *buf, size_t len)
 {
-    uintptr_t ibuf = (uintptr_t)buf;
 #ifdef CONFIG_AVX2_OPT
-    if (len % 128 == 0 && ibuf % 32 == 0 && (cpuid_cache & CACHE_AVX2)) {
+    if (len >= 128 && (cpuid_cache & CACHE_AVX2)) {
         return buffer_zero_avx2(buf, len);
     }
-    if (len % 64 == 0 && ibuf % 16 == 0 && (cpuid_cache & CACHE_SSE4)) {
+    if (len >= 64 && (cpuid_cache & CACHE_SSE4)) {
         return buffer_zero_sse4(buf, len);
     }
 #endif
-    if (len % 64 == 0 && ibuf % 16 == 0 && (cpuid_cache & CACHE_SSE2)) {
+    if (len >= 64 && (cpuid_cache & CACHE_SSE2)) {
         return buffer_zero_sse2(buf, len);
     }
     return buffer_zero_int(buf, len);
-- 
1.8.3.1