UserspaceEmulator: Add a fast path for forward REP STOSD

Same as REP STOSB, except for 32-bit fills instead.

DevTools/UserspaceEmulator/SoftCPU.cpp

@@ -3072,6 +3072,22 @@ void SoftCPU::STOSB(const X86::Instruction& insn)
 
 void SoftCPU::STOSD(const X86::Instruction& insn)
 {
+    if (insn.has_rep_prefix() && !df()) {
+        // Fast path for 32-bit forward memory fill.
+        if (m_emulator.mmu().fast_fill_memory32({ es(), destination_index(insn.a32()).value() }, ecx().value(), eax())) {
+            if (insn.a32()) {
+                // FIXME: Should an uninitialized ECX taint EDI here?
+                set_edi({ (u32)(edi().value() + (ecx().value() * sizeof(u32))), edi().shadow() });
+                set_ecx(shadow_wrap_as_initialized<u32>(0));
+            } else {
+                // FIXME: Should an uninitialized CX taint DI here?
+                set_di({ (u16)(di().value() + (cx().value() * sizeof(u32))), di().shadow() });
+                set_cx(shadow_wrap_as_initialized<u16>(0));
+            }
+            return;
+        }
+    }
+
     do_once_or_repeat<false>(insn, [&] {
         write_memory32({ es(), destination_index(insn.a32()).value() }, eax());
         step_destination_index(insn.a32(), 4);

DevTools/UserspaceEmulator/SoftMMU.cpp

@@ -259,7 +259,7 @@ bool SoftMMU::fast_fill_memory8(X86::LogicalAddress address, size_t size, ValueW
     if (region->is_mmap() && static_cast<const MmapRegion&>(*region).is_malloc_block()) {
         if (auto* tracer = Emulator::the().malloc_tracer()) {
             // FIXME: Add a way to audit an entire range of memory instead of looping here!
-            for (size_t i = 0; i < size; i += sizeof(u8)) {
+            for (size_t i = 0; i < size; ++i) {
                 tracer->audit_write(address.offset() + (i * sizeof(u8)), sizeof(u8));
             }
         }
@@ -271,4 +271,29 @@ bool SoftMMU::fast_fill_memory8(X86::LogicalAddress address, size_t size, ValueW
     return true;
 }
 
+bool SoftMMU::fast_fill_memory32(X86::LogicalAddress address, size_t count, ValueWithShadow<u32> value)
+{
+    if (!count)
+        return true;
+    auto* region = find_region(address);
+    if (!region)
+        return false;
+    if (!region->contains(address.offset() + (count * sizeof(u32)) - 1))
+        return false;
+
+    if (region->is_mmap() && static_cast<const MmapRegion&>(*region).is_malloc_block()) {
+        if (auto* tracer = Emulator::the().malloc_tracer()) {
+            // FIXME: Add a way to audit an entire range of memory instead of looping here!
+            for (size_t i = 0; i < count; ++i) {
+                tracer->audit_write(address.offset() + (i * sizeof(u32)), sizeof(u32));
+            }
+        }
+    }
+
+    size_t offset_in_region = address.offset() - region->base();
+    fast_u32_fill((u32*)(region->data() + offset_in_region), value.value(), count);
+    fast_u32_fill((u32*)(region->shadow_data() + offset_in_region), value.shadow(), count);
+    return true;
+}
+
 }

DevTools/UserspaceEmulator/SoftMMU.h

@@ -109,6 +109,7 @@ public:
     void set_tls_region(NonnullOwnPtr<Region>);
 
     bool fast_fill_memory8(X86::LogicalAddress, size_t size, ValueWithShadow<u8>);
+    bool fast_fill_memory32(X86::LogicalAddress, size_t size, ValueWithShadow<u32>);
 
     void copy_to_vm(FlatPtr destination, const void* source, size_t);
     void copy_from_vm(void* destination, const FlatPtr source, size_t);