UserspaceEmulator: Add a fast path for forward REP STOSB

This is used by memset() so we get a lot of mileage out of optimizing
this instruction.

Note that we currently audit every individual byte accessed separately.
This could be greatly improved by adding a range auditing mechanism to
MallocTracer.

DevTools/UserspaceEmulator/SoftCPU.cpp

@@ -3048,6 +3048,22 @@ void SoftCPU::STI(const X86::Instruction&) { TODO_INSN(); }
 
 void SoftCPU::STOSB(const X86::Instruction& insn)
 {
+    if (insn.has_rep_prefix() && !df()) {
+        // Fast path for 8-bit forward memory fill.
+        if (m_emulator.mmu().fast_fill_memory8({ es(), destination_index(insn.a32()).value() }, ecx().value(), al())) {
+            if (insn.a32()) {
+                // FIXME: Should an uninitialized ECX taint EDI here?
+                set_edi({ (u32)(edi().value() + ecx().value()), 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()), di().shadow() });
+                set_cx(shadow_wrap_as_initialized<u16>(0));
+            }
+            return;
+        }
+    }
+
     do_once_or_repeat<false>(insn, [&] {
         write_memory8({ es(), destination_index(insn.a32()).value() }, al());
         step_destination_index(insn.a32(), 1);

DevTools/UserspaceEmulator/SoftMMU.cpp

@@ -26,9 +26,11 @@
 
 #include "SoftMMU.h"
 #include "Emulator.h"
+#include "MmapRegion.h"
 #include "Report.h"
 #include "SharedBufferRegion.h"
 #include <AK/ByteBuffer.h>
+#include <AK/Memory.h>
 
 namespace UserspaceEmulator {
 
@@ -244,4 +246,29 @@ SharedBufferRegion* SoftMMU::shbuf_region(int shbuf_id)
     return (SharedBufferRegion*)m_shbuf_regions.get(shbuf_id).value_or(nullptr);
 }
 
+bool SoftMMU::fast_fill_memory8(X86::LogicalAddress address, size_t size, ValueWithShadow<u8> value)
+{
+    if (!size)
+        return true;
+    auto* region = find_region(address);
+    if (!region)
+        return false;
+    if (!region->contains(address.offset() + size - 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 < size; i += sizeof(u8)) {
+                tracer->audit_write(address.offset() + (i * sizeof(u8)), sizeof(u8));
+            }
+        }
+    }
+
+    size_t offset_in_region = address.offset() - region->base();
+    memset(region->data() + offset_in_region, value.value(), size);
+    memset(region->shadow_data() + offset_in_region, value.shadow(), size);
+    return true;
+}
+
 }

DevTools/UserspaceEmulator/SoftMMU.h

@@ -108,6 +108,8 @@ public:
 
     void set_tls_region(NonnullOwnPtr<Region>);
 
+    bool fast_fill_memory8(X86::LogicalAddress, size_t size, ValueWithShadow<u8>);
+
     void copy_to_vm(FlatPtr destination, const void* source, size_t);
     void copy_from_vm(void* destination, const FlatPtr source, size_t);
     ByteBuffer copy_buffer_from_vm(const FlatPtr source, size_t);