rsx: Write the ibo tail in the correct order

Fixes MSG4’s graveyard scene.
RPCS3 · Oct 29, 2019 · b4327a8 · b4327a8
1 parent 066286c
commit b4327a8
Showing 1 changed file with 21 additions and 28 deletions.
diff --git a/rpcs3/Emu/RSX/Common/BufferUtils.cpp b/rpcs3/Emu/RSX/Common/BufferUtils.cpp
@@ -735,7 +735,7 @@ namespace
 	{
 		AVX2_FUNC
 		static
-		std::tuple<u16, u16, u32> upload_u16_swapped_avx2(const void *src, void *dst, u32 count, u16 restart_index)
+		std::tuple<u16, u16> upload_u16_swapped_avx2(const void *src, void *dst, u32 iterations, u16 restart_index)
 		{
 			const __m256i shuffle_mask = _mm256_set_epi8(
 				0xE, 0xF, 0xC, 0xD,
@@ -754,7 +754,6 @@ namespace
 			__m256i min = _mm256_set1_epi16(0xffff);
 			__m256i max = _mm256_set1_epi16(0);
 
-			const auto iterations = count / 16;
 			for (unsigned n = 0; n < iterations; ++n)
 			{
 				const __m256i raw = _mm256_loadu_si256(src_stream++);
@@ -802,12 +801,12 @@ namespace
 			const u16 min_index = u16(_mm_cvtsi128_si32(min2) & 0xFFFF);
 			const u16 max_index = u16(_mm_cvtsi128_si32(max2) & 0xFFFF);
 
-			return std::make_tuple(min_index, max_index, count);
+			return std::make_tuple(min_index, max_index);
 		}
 
 		SSE4_1_FUNC
 		static
-		std::tuple<u16, u16, u32> upload_u16_swapped_sse4_1(const void *src, void *dst, u32 count, u16 restart_index)
+		std::tuple<u16, u16> upload_u16_swapped_sse4_1(const void *src, void *dst, u32 iterations, u16 restart_index)
 		{
 			const __m128i shuffle_mask = _mm_set_epi8(
 				0xE, 0xF, 0xC, 0xD,
@@ -822,7 +821,6 @@ namespace
 			__m128i min = _mm_set1_epi16(0xffff);
 			__m128i max = _mm_set1_epi16(0);
 
-			const auto iterations = count / 8;
 			for (unsigned n = 0; n < iterations; ++n)
 			{
 				const __m128i raw = _mm_loadu_si128(src_stream++);
@@ -864,12 +862,12 @@ namespace
 			const u16 min_index = u16(_mm_cvtsi128_si32(min) & 0xFFFF);
 			const u16 max_index = u16(_mm_cvtsi128_si32(max) & 0xFFFF);
 
-			return std::make_tuple(min_index, max_index, count);
+			return std::make_tuple(min_index, max_index);
 		}
 
 		SSE4_1_FUNC
 		static
-		std::tuple<u32, u32, u32> upload_u32_swapped_sse4_1(const void *src, void *dst, u32 count, u32 restart_index)
+		std::tuple<u32, u32> upload_u32_swapped_sse4_1(const void *src, void *dst, u32 iterations, u32 restart_index)
 		{
 			const __m128i shuffle_mask = _mm_set_epi8(
 				0xC, 0xD, 0xE, 0xF,
@@ -884,7 +882,6 @@ namespace
 			__m128i min = _mm_set1_epi32(0xffffffff);
 			__m128i max = _mm_set1_epi32(0);
 
-			const auto iterations = count / 4;
 			for (unsigned n = 0; n < iterations; ++n)
 			{
 				const __m128i raw = _mm_loadu_si128(src_stream++);
@@ -918,57 +915,53 @@ namespace
 			const u32 min_index = u32(_mm_cvtsi128_si32(min));
 			const u32 max_index = u32(_mm_cvtsi128_si32(max));
 
-			return std::make_tuple(min_index, max_index, count);
+			return std::make_tuple(min_index, max_index);
 		}
 
 		template<typename T>
 		static
 		std::tuple<T, T, u32> upload_untouched(gsl::span<to_be_t<const T>> src, gsl::span<T> dst, T restart_index, bool skip_restart)
 		{
-			T min_index, max_index;
+			T min_index = index_limit<T>();
+			T max_index = 0;
 			u32 written = 0;
-			u32 remaining = src.size();
+			u32 length = src.size();
 
-			if (remaining >= 32 && !skip_restart)
+			if (length >= 32 && !skip_restart)
 			{
 				if constexpr (std::is_same<T, u16>::value)
 				{
 					if (s_use_avx2)
 					{
-						const auto count = (remaining & ~0xf);
-						std::tie(min_index, max_index, written) = upload_u16_swapped_avx2(src.data(), dst.data(), count, restart_index);
+						u32 iterations = length >> 4;
+						written = length & 0xf;
+						std::tie(min_index, max_index) = upload_u16_swapped_avx2(src.data(), dst.data(), iterations, restart_index);
 					}
 					else if (s_use_sse4_1)
 					{
-						const auto count = (remaining & ~0x7);
-						std::tie(min_index, max_index, written) = upload_u16_swapped_sse4_1(src.data(), dst.data(), count, restart_index);
+						u32 iterations = length >> 3;
+						written = length & 0x7;
+						std::tie(min_index, max_index) = upload_u16_swapped_sse4_1(src.data(), dst.data(), iterations, restart_index);
 					}
 				}
 				else if constexpr (std::is_same<T, u32>::value)
 				{
 					if (s_use_sse4_1)
 					{
-						const auto count = (remaining & ~0x3);
-						std::tie(min_index, max_index, written) = upload_u32_swapped_sse4_1(src.data(), dst.data(), count, restart_index);
+						u32 iterations = length >> 2;
+						written = length & 0x3;
+						std::tie(min_index, max_index) = upload_u32_swapped_sse4_1(src.data(), dst.data(), iterations, restart_index);
 					}
 				}
 				else
 				{
 					fmt::throw_exception("Unreachable" HERE);
 				}
-
-				remaining -= written;
-			}
-			else
-			{
-				min_index = index_limit<T>();
-				max_index = 0;
-				written = 0;
 			}
 
-			while (remaining--)
+			for (u32 i = written; i < length; ++i)
 			{
-				T index = src[written];
+				T index = src[i];
 				if (index == restart_index)
 				{
 					if (!skip_restart)