Bug 1921862 - Refresh and reappply the patch on top of libaom. r=media-playback-reviewers,aosmond

Differential Revision: https://phabricator.services.mozilla.com/D230507

media/libaom/mmloadusi64.patch

@@ -1,19 +1,19 @@
 diff --git a/aom_dsp/x86/synonyms.h b/aom_dsp/x86/synonyms.h
 --- a/aom_dsp/x86/synonyms.h
 +++ b/aom_dsp/x86/synonyms.h
-@@ -41,23 +41,34 @@ static INLINE __m128i xx_loadl_64(const void *a) {
- static INLINE __m128i xx_load_128(const void *a) {
+@@ -41,23 +41,34 @@ static inline __m128i xx_loadl_64(const 
+ static inline __m128i xx_load_128(const void *a) {
    return _mm_load_si128((const __m128i *)a);
  }
-
- static INLINE __m128i xx_loadu_128(const void *a) {
+ 
+ static inline __m128i xx_loadu_128(const void *a) {
    return _mm_loadu_si128((const __m128i *)a);
  }
-
+ 
 +// _mm_loadu_si64 has been introduced in GCC 9, reimplement the function
 +// manually on older compilers.
 +#if !defined(__clang__) && __GNUC_MAJOR__ < 9
-+static INLINE __m128i xx_loadu_2x64(const void *hi, const void *lo) {
++static inline __m128i xx_loadu_2x64(const void *hi, const void *lo) {
 +  __m64 hi_, lo_;
 +  memcpy(&hi_, hi, sizeof(hi_));
 +  memcpy(&lo_, lo, sizeof(lo_));
@@ -23,44 +23,44 @@ diff --git a/aom_dsp/x86/synonyms.h b/aom_dsp/x86/synonyms.h
  // Load 64 bits from each of hi and low, and pack into an SSE register
  // Since directly loading as `int64_t`s and using _mm_set_epi64 may violate
  // the strict aliasing rule, this takes a different approach
- static INLINE __m128i xx_loadu_2x64(const void *hi, const void *lo) {
+ static inline __m128i xx_loadu_2x64(const void *hi, const void *lo) {
    return _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)lo),
                              _mm_loadl_epi64((const __m128i *)hi));
  }
 +#endif
-
- static INLINE void xx_storel_32(void *const a, const __m128i v) {
+ 
+ static inline void xx_storel_32(void *const a, const __m128i v) {
    const int val = _mm_cvtsi128_si32(v);
    memcpy(a, &val, sizeof(val));
  }
-
- static INLINE void xx_storel_64(void *const a, const __m128i v) {
+ 
+ static inline void xx_storel_64(void *const a, const __m128i v) {
    _mm_storel_epi64((__m128i *)a, v);
 diff --git a/aom_dsp/x86/synonyms_avx2.h b/aom_dsp/x86/synonyms_avx2.h
 --- a/aom_dsp/x86/synonyms_avx2.h
 +++ b/aom_dsp/x86/synonyms_avx2.h
-@@ -55,21 +55,36 @@ static INLINE __m256i yy_set2_epi16(int16_t a, int16_t b) {
-
- // Some compilers don't have _mm256_set_m128i defined in immintrin.h. We
- // therefore define an equivalent function using a different intrinsic.
- // ([ hi ], [ lo ]) -> [ hi ][ lo ]
- static INLINE __m256i yy_set_m128i(__m128i hi, __m128i lo) {
-   return _mm256_insertf128_si256(_mm256_castsi128_si256(lo), hi, 1);
+@@ -71,21 +71,36 @@ static inline __m256i yy_loadu_4x64(cons
+   __m128d v23 = _mm_loadh_pd(v2, (const double *)e3);
+   // Note this can be replaced with
+   // `_mm256_castpd_si256(_mm256_set_m128d(v23, v01))` if immintrin.h contains
+   // _mm256_set_m128d() with all supported compilers. This version is used to
+   // match the behavior with yy_set_m128i().
+   return yy_set_m128i(_mm_castpd_si128(v23), _mm_castpd_si128(v01));
  }
-
+ 
 +#define GCC_VERSION (__GNUC__ * 10000 \
 +                     + __GNUC_MINOR__ * 100 \
 +                     + __GNUC_PATCHLEVEL__)
 +
 +// _mm256_loadu2_m128i has been introduced in GCC 10.1
 +#if !defined(__clang__) && GCC_VERSION < 101000
-+static INLINE __m256i yy_loadu2_128(const void *hi, const void *lo) {
++static inline __m256i yy_loadu2_128(const void *hi, const void *lo) {
 +  __m128i mhi = _mm_loadu_si128((const __m128i *)(hi));
 +  __m128i mlo = _mm_loadu_si128((const __m128i *)(lo));
 +  return _mm256_set_m128i(mhi, mlo);
 +}
 +#else
- static INLINE __m256i yy_loadu2_128(const void *hi, const void *lo) {
+ static inline __m256i yy_loadu2_128(const void *hi, const void *lo) {
    __m128i mhi = _mm_loadu_si128((const __m128i *)(hi));
    __m128i mlo = _mm_loadu_si128((const __m128i *)(lo));
    return yy_set_m128i(mhi, mlo);
@@ -68,11 +68,11 @@ diff --git a/aom_dsp/x86/synonyms_avx2.h b/aom_dsp/x86/synonyms_avx2.h
 +#endif
 +
 +#undef GCC_VERSION
-
- static INLINE void yy_storeu2_128(void *hi, void *lo, const __m256i a) {
+ 
+ static inline void yy_storeu2_128(void *hi, void *lo, const __m256i a) {
    _mm_storeu_si128((__m128i *)hi, _mm256_extracti128_si256(a, 1));
    _mm_storeu_si128((__m128i *)lo, _mm256_castsi256_si128(a));
  }
-
- static INLINE __m256i yy_roundn_epu16(__m256i v_val_w, int bits) {
+ 
+ static inline __m256i yy_roundn_epu16(__m256i v_val_w, int bits) {
    const __m256i v_s_w = _mm256_srli_epi16(v_val_w, bits - 1);

third_party/aom/aom_dsp/x86/synonyms.h

@@ -46,6 +46,16 @@ static inline __m128i xx_loadu_128(const void *a) {
   return _mm_loadu_si128((const __m128i *)a);
 }
 
+// _mm_loadu_si64 has been introduced in GCC 9, reimplement the function
+// manually on older compilers.
+#if !defined(__clang__) && __GNUC_MAJOR__ < 9
+static inline __m128i xx_loadu_2x64(const void *hi, const void *lo) {
+  __m64 hi_, lo_;
+  memcpy(&hi_, hi, sizeof(hi_));
+  memcpy(&lo_, lo, sizeof(lo_));
+  return _mm_set_epi64(hi_, lo_);
+}
+#else
 // Load 64 bits from each of hi and low, and pack into an SSE register
 // Since directly loading as `int64_t`s and using _mm_set_epi64 may violate
 // the strict aliasing rule, this takes a different approach
@@ -53,6 +63,7 @@ static inline __m128i xx_loadu_2x64(const void *hi, const void *lo) {
   return _mm_unpacklo_epi64(_mm_loadl_epi64((const __m128i *)lo),
                             _mm_loadl_epi64((const __m128i *)hi));
 }
+#endif
 
 static inline void xx_storel_32(void *const a, const __m128i v) {
   const int val = _mm_cvtsi128_si32(v);

third_party/aom/aom_dsp/x86/synonyms_avx2.h

@@ -76,11 +76,26 @@ static inline __m256i yy_loadu_4x64(const void *e3, const void *e2,
   return yy_set_m128i(_mm_castpd_si128(v23), _mm_castpd_si128(v01));
 }
 
+#define GCC_VERSION (__GNUC__ * 10000 \
+                     + __GNUC_MINOR__ * 100 \
+                     + __GNUC_PATCHLEVEL__)
+
+// _mm256_loadu2_m128i has been introduced in GCC 10.1
+#if !defined(__clang__) && GCC_VERSION < 101000
+static inline __m256i yy_loadu2_128(const void *hi, const void *lo) {
+  __m128i mhi = _mm_loadu_si128((const __m128i *)(hi));
+  __m128i mlo = _mm_loadu_si128((const __m128i *)(lo));
+  return _mm256_set_m128i(mhi, mlo);
+}
+#else
 static inline __m256i yy_loadu2_128(const void *hi, const void *lo) {
   __m128i mhi = _mm_loadu_si128((const __m128i *)(hi));
   __m128i mlo = _mm_loadu_si128((const __m128i *)(lo));
   return yy_set_m128i(mhi, mlo);
 }
+#endif
+
+#undef GCC_VERSION
 
 static inline void yy_storeu2_128(void *hi, void *lo, const __m256i a) {
   _mm_storeu_si128((__m128i *)hi, _mm256_extracti128_si256(a, 1));