From 577a6e24597ab56c7a364bcc1742d8816ca8f164 Mon Sep 17 00:00:00 2001
From: Elad <18193363+elad335@users.noreply.github.com>
Date: Mon, 29 Jun 2026 17:42:30 +0300
Subject: [PATCH] SPU Precompilation: Add support for unaligned segments

---
 rpcs3/Emu/Cell/SPUCommonRecompiler.cpp | 35 ++++++++++++++++++++++----
 1 file changed, 30 insertions(+), 5 deletions(-)

diff --git a/rpcs3/Emu/Cell/SPUCommonRecompiler.cpp b/rpcs3/Emu/Cell/SPUCommonRecompiler.cpp
index e8f2a31fa8..dbb7e048f8 100644
--- a/rpcs3/Emu/Cell/SPUCommonRecompiler.cpp
+++ b/rpcs3/Emu/Cell/SPUCommonRecompiler.cpp
@@ -646,18 +646,32 @@ spu_cache::~spu_cache()
 
 extern void utilize_spu_data_segment(u32 vaddr, const void* ls_data_vaddr, u32 size)
 {
-	if (vaddr % 4)
+	if (u32 rem = vaddr % 4)
 	{
-		return;
+		// The remainder that it needs to be aligned up to the next DWORD
+		rem = 4 - rem;
+
+		if (size < rem)
+		{
+			return;
+		}
+
+		vaddr += rem;
+		ls_data_vaddr = reinterpret_cast<const char*>(ls_data_vaddr) + rem;
+		size -= rem;
 	}
 
 	size &= -4;
 
-	if (!size || vaddr + size > SPU_LS_SIZE)
+	if (!size || vaddr >= SPU_LS_SIZE)
 	{
 		return;
 	}
 
+	// Let SPU block search mistakes pass through
+	// SPU code mining is too important
+	size = std::min<u32>(SPU_LS_SIZE - vaddr, size);
+
 	if (!g_cfg.core.llvm_precompilation)
 	{
 		return;
@@ -2364,11 +2378,22 @@ std::vector<u32> spu_thread::discover_functions(u32 base_addr, std::span<const u
 	// TODO: Does not detect jumptables or fixed-addr indirect calls
 	const v128 brasl_mask = is_known_addr ? v128::from32p(0x62u << 23) : v128::from32p(umax);
 
-	for (u32 i = utils::align<u32>(base_addr, 0x10); i < std::min<u32>(base_addr + ::size32(ls), 0x3FFF0); i += 0x10)
+	for (u32 i = base_addr, end_ls = std::min<u32>(base_addr + ::size32(ls), SPU_LS_SIZE); i < end_ls; i = utils::align<u32>(i + 1, 0x10))
 	{
 		// Search for BRSL LR and BRASL LR or BR
 		// TODO: BISL
-		const v128 inst = read_from_ptr<be_t<v128>>(ls, i - base_addr);
+		be_t<v128> inst_be{};
+
+		if (end_ls - i < 16)
+		{
+			std::memcpy(&inst_be, ls.data() + (i - base_addr), end_ls - i);
+		}
+		else
+		{
+			inst_be = read_from_ptr<be_t<v128>>(ls, i - base_addr);
+		}
+
+		const v128 inst = inst_be;
 		const v128 cleared_i16 = gv_and32(inst, v128::from32p(std::rotl<u32>(~0xffff, 7)));
 		const v128 eq_brsl = gv_eq32(cleared_i16, v128::from32p(0x66u << 23));
 		const v128 eq_brasl = gv_eq32(cleared_i16, brasl_mask);