From 17ebb38ebd91d4594e578232c5478add1e4fa5e3 Mon Sep 17 00:00:00 2001
From: Ryan Houdek <Sonicadvance1@gmail.com>
Date: Tue, 15 Sep 2026 14:21:08 -0700
Subject: [PATCH] FEXCore/Utils: Make crc32 match between architectures

I forgot that we actually need this to match arm64 behaviour on x86.
Implement the pseudocodefor the instruction.
---
 FEXCore/Source/Utils/crc32.h         | 51 +++++++++++++++++-----------
 FEXCore/unittests/APITests/crc32.cpp |  5 ---
 2 files changed, 32 insertions(+), 24 deletions(-)

diff --git a/FEXCore/Source/Utils/crc32.h b/FEXCore/Source/Utils/crc32.h
index 5af0e7cc4..f9fc7b114 100644
--- a/FEXCore/Source/Utils/crc32.h
+++ b/FEXCore/Source/Utils/crc32.h
@@ -4,8 +4,6 @@
 #include <cstdint>
 #if defined(ARCHITECTURE_arm64)
 #include <arm_acle.h>
-#elif defined(ARCHITECTURE_x86_64)
-#include <nmmintrin.h>
 #endif
 
 namespace FEXCore::Utils {
@@ -23,24 +21,39 @@ static inline uint32_t crc32(const T* Ptr, size_t Size) {
   do_crc(uint32_t, w);
   do_crc(uint16_t, h);
   do_crc(uint8_t, b);
-
-#elif defined(ARCHITECTURE_x86_64)
-#define do_crc(type, intrinsic)                                      \
-  while (Size >= sizeof(type)) {                                     \
-    Result = intrinsic(Result, *reinterpret_cast<const type*>(Ptr)); \
-    Ptr += sizeof(type);                                             \
-    Size -= sizeof(type);                                            \
-  }
-
-  do_crc(uint64_t, _mm_crc32_u64);
-  do_crc(uint32_t, _mm_crc32_u32);
-  do_crc(uint16_t, _mm_crc32_u16);
-  do_crc(uint8_t, _mm_crc32_u8);
-#else
-  // Unsupported on non-arm.
-  return 0;
-#endif
 #undef do_crc
+#else
+  // Emulate arm64 crc32.
+  // This is basically just the pseudo-code for crc32b.
+  // Doesn't need to be fast, just needs to match.
+  auto reverse_bits = [](auto bits) {
+    decltype(bits) Result {};
+    for (size_t i = 0; i < (sizeof(decltype(bits)) * 8); ++i) {
+      Result = (Result << 1) | ((bits >> i) & 1);
+    }
+    return Result;
+  };
+
+  auto Poly32Mod2 = [](uint64_t data) -> uint32_t {
+    constexpr static size_t bits = 40;
+    constexpr static uint64_t poly = 0x04C11DB7U;
+    for (size_t i = (bits - 1); i >= 32; --i) {
+      if (((data >> i) & 1) != 0) {
+        const uint64_t poly_shift = poly << (i - 32);
+        const uint64_t data_mask = (1ULL << i) - 1;
+        data = (data & data_mask) ^ poly_shift;
+      }
+    }
+
+    return data;
+  };
+
+  for (size_t i = 0; i < Size; ++i) {
+    uint64_t TempAcc = static_cast<uint64_t>(reverse_bits(Result)) << 8;
+    uint64_t TempVal = static_cast<uint64_t>(reverse_bits(reinterpret_cast<const uint8_t*>(Ptr)[i])) << 32;
+    Result = reverse_bits(Poly32Mod2(TempAcc ^ TempVal));
+  }
+#endif
   return Result;
 }
 } // namespace FEXCore::Utils
diff --git a/FEXCore/unittests/APITests/crc32.cpp b/FEXCore/unittests/APITests/crc32.cpp
index eadb6adc0..7ec76b83d 100644
--- a/FEXCore/unittests/APITests/crc32.cpp
+++ b/FEXCore/unittests/APITests/crc32.cpp
@@ -5,10 +5,5 @@
 
 TEST_CASE("Simple") {
   uint32_t data = 0x41424344U;
-  // Arm and x86 use different polynomials by default.
-#if defined(ARCHITECTURE_arm64)
   CHECK(FEXCore::Utils::crc32(&data, sizeof(data)) == 0xa53ea072);
-#elif defined(ARCHITECTURE_x86_64)
-  CHECK(FEXCore::Utils::crc32(&data, sizeof(data)) == 0x968d46eb);
-#endif
 }