PPU/x64: Fix gv_cvtu32_tofs

rpcs3/Emu/Cell/PPUTranslator.cpp

@@ -1115,7 +1115,24 @@ void PPUTranslator::VCFSX(ppu_opcode_t op)
 void PPUTranslator::VCFUX(ppu_opcode_t op)
 {
 	const auto b = get_vr<u32[4]>(op.vb);
-	set_vr(op.vd, fpcast<f32[4]>(b) * fsplat<f32[4]>(std::pow(2, -static_cast<int>(op.vuimm))));
+
+#ifdef ARCH_ARM64
+	return set_vr(op.vd, fpcast<f32[4]>(b) * fsplat<f32[4]>(std::pow(2, -static_cast<int>(op.vuimm))));
+#else
+	if (m_use_avx512)
+	{
+		return set_vr(op.vd, fpcast<f32[4]>(b) * fsplat<f32[4]>(std::pow(2, -static_cast<int>(op.vuimm))));
+	}
+
+	constexpr int bit_shift = 9;
+	const auto shifted = (b >> bit_shift);
+	const auto cleared = shifted << bit_shift;
+	const auto low_bits = b - cleared;
+	const auto high_part = fpcast<f32[4]>(noncast<s32[4]>(shifted)) * fsplat<f32[4]>(static_cast<f32>(1u << bit_shift));
+	const auto low_part = fpcast<f32[4]>(noncast<s32[4]>(low_bits));
+	const auto temp = high_part + low_part;
+	set_vr(op.vd, temp * fsplat<f32[4]>(std::pow(2, -static_cast<int>(op.vuimm))));
+#endif
 }
 
 void PPUTranslator::VCMPBFP(ppu_opcode_t op)

rpcs3/util/simd.hpp

@@ -2213,8 +2213,11 @@ inline v128 gv_cvtu32_tofs(const v128& src)
 #if defined(__AVX512VL__)
 	return _mm_cvtepu32_ps(src);
 #elif defined(ARCH_X64)
-	const auto fix = _mm_and_ps(_mm_castsi128_ps(_mm_srai_epi32(src, 31)), _mm_set1_ps(0x80000000));
-	return _mm_add_ps(_mm_cvtepi32_ps(_mm_and_si128(src, _mm_set1_epi32(0x7fffffff))), fix);
+	constexpr u64 bit_shift = 9;
+	const auto shifted = _mm_srli_epi32(src, bit_shift);
+	const auto cleared = _mm_slli_epi32(shifted, bit_shift);
+	const auto low_bits = _mm_sub_epi32(src, cleared);
+	return _mm_add_ps(_mm_cvtepi32_ps(low_bits), _mm_mul_ps(_mm_cvtepi32_ps(shifted), _mm_set_ps1(1u << bit_shift)));
 #elif defined(ARCH_ARM64)
 	return vcvtq_f32_u32(src);
 #endif