feat: optimize instruction fetch and decoding with big jump tables

src/Makefile

@@ -214,18 +214,27 @@ SHA3_CFLAGS=-O3
 
 # Optimization flags for the interpreter
 ifneq (,$(filter yes,$(relwithdebinfo) $(release)))
-ifneq (,$(filter gcc,$(CC)))
-# The following flag helps GCC to eliminate more redundant computations in the interpret loop,
-# saving some host instructions and improving performance.
-# This flag is usually enabled by default at -O3,
-# but we don't use -O3 because it enables some other flags that are not worth for the interpreter.
-INTERPRET_CXXFLAGS+=-fgcse-after-reload -fpredictive-commoning -fsplit-paths -ftree-partial-pre
+ifneq (,$(findstring gcc,$(CC)))
+# The following improves computed goto dispatch as stated in GCC manual
+INTERPRET_CXXFLAGS+=-fno-gcse
+# The following remove extra jumps in the computed goto dispatch
+INTERPRET_CXXFLAGS+=-fno-crossjumping
+# The following remove extra NOPs before jumping back to the interpret hot loop
+INTERPRET_CXXFLAGS+=-fno-align-loops
+# The interpreter dispatch loop performs better as a big inlined function
+INTERPRET_CXXFLAGS+=-finline-limit=1024
+# The interpreter hot loop is big and puts pressure on register allocation, this improves register use
+INTERPRET_CXXFLAGS+=-frename-registers -fweb
+# The interpreter instruction dispatch is big, the following reduces its size minimizing CPU cache pressure
+INTERPRET_CXXFLAGS+=-freorder-blocks-algorithm=simple
+# Some distributions enables stack protector by default, make sure it's disabled
+INTERPRET_CXXFLAGS+=-fno-stack-protector
 endif
-# Disable jump tables, because it degrades the instruction decoding performance in the interpret loop,
-# since it generates a memory indirection that has a high cost in opcode switches.
-INTERPRET_CXXFLAGS+=-fno-jump-tables
 endif
 
+# Make testing new optimization options easier
+INTERPRET_CXXFLAGS+=$(MYINTERPRET_CXXFLAGS)
+
 # Link time optimizations
 ifeq ($(lto),yes)
 OPTFLAGS+=-flto=auto
@@ -262,7 +271,7 @@ PGO_WORKLOAD=\
 	whetstone 25000
 
 LINTER_IGNORE_SOURCES=
-LINTER_IGNORE_HEADERS=
+LINTER_IGNORE_HEADERS=interpret-jump-table.h
 LINTER_SOURCES=$(filter-out $(LINTER_IGNORE_SOURCES),$(strip $(wildcard *.cpp) $(wildcard *.c)))
 LINTER_HEADERS=$(filter-out $(LINTER_IGNORE_HEADERS),$(strip $(wildcard *.hpp) $(wildcard *.h)))
 
@@ -273,7 +282,7 @@ CLANG_FORMAT=clang-format
 CLANG_FORMAT_UARCH_FILES:=$(wildcard ../uarch/*.cpp)
 CLANG_FORMAT_UARCH_FILES:=$(filter-out %uarch-printf%,$(strip $(CLANG_FORMAT_UARCH_FILES)))
 CLANG_FORMAT_FILES:=$(wildcard *.cpp) $(wildcard *.c) $(wildcard *.h) $(wildcard *.hpp) $(CLANG_FORMAT_UARCH_FILES)
-CLANG_FORMAT_IGNORE_FILES:=
+CLANG_FORMAT_IGNORE_FILES:=interpret-jump-table.h
 CLANG_FORMAT_FILES:=$(strip $(CLANG_FORMAT_FILES))
 CLANG_FORMAT_FILES:=$(filter-out $(CLANG_FORMAT_IGNORE_FILES),$(strip $(CLANG_FORMAT_FILES)))
 
@@ -541,12 +550,12 @@ jsonrpc-discover.cpp: jsonrpc-discover.json
 	echo '} // namespace cartesi' >> jsonrpc-discover.cpp
 
 %.clang-tidy: %.cpp machine-c-version.h
-	@$(CLANG_TIDY) --header-filter='$(CLANG_TIDY_HEADER_FILTER)' $(CLANG_TIDY_FLAGS) $< -- $(CXXFLAGS) $(LUA_INC) 2>/dev/null
+	@$(CLANG_TIDY) --header-filter='$(CLANG_TIDY_HEADER_FILTER)' $(CLANG_TIDY_FLAGS) $< -- $(CXXFLAGS) $(LUA_INC) -DCLANG_TIDY_LINT 2>/dev/null
 	@$(CXX) $(CXXFLAGS) $(LUA_INC) $< -MM -MT $@ -MF $@.d > /dev/null 2>&1
 	@touch $@
 
 %.clang-tidy: %.c
-	@$(CLANG_TIDY) --header-filter='$(CLANG_TIDY_HEADER_FILTER)' $(CLANG_TIDY_FLAGS) $< -- $(CFLAGS) 2>/dev/null
+	@$(CLANG_TIDY) --header-filter='$(CLANG_TIDY_HEADER_FILTER)' $(CLANG_TIDY_FLAGS) $< -- $(CFLAGS) -DCLANG_TIDY_LINT 2>/dev/null
 	@$(CC) $(CFLAGS) $< -MM -MT $@ -MF $@.d > /dev/null 2>&1
 	@touch $@
 

src/device-state-access.h

@@ -36,7 +36,7 @@ namespace cartesi {
 template <typename STATE_ACCESS>
 class device_state_access : public i_device_state_access {
 public:
-    explicit device_state_access(STATE_ACCESS &a, uint64_t mcycle) : m_a(a), m_mcycle(mcycle) {
+    explicit device_state_access(STATE_ACCESS a, uint64_t mcycle) : m_a(a), m_mcycle(mcycle) {
         static_assert(is_an_i_state_access<STATE_ACCESS>::value, "not an i_state_access");
     }
 
@@ -52,7 +52,7 @@ class device_state_access : public i_device_state_access {
     ~device_state_access() override = default;
 
 private:
-    STATE_ACCESS &m_a; // NOLINT(cppcoreguidelines-avoid-const-or-ref-data-members)
+    STATE_ACCESS m_a; // NOLINT(cppcoreguidelines-avoid-const-or-ref-data-members)
     uint64_t m_mcycle;
 
     void do_set_mip(uint64_t mask) override {

src/i-state-access.h

@@ -24,6 +24,7 @@
 #include <type_traits>
 #include <utility>
 
+#include "compiler-defines.h"
 #include "meta.h"
 #include "shadow-tlb.h"
 
@@ -773,15 +774,15 @@ class i_state_access { // CRTP
     }
 
     /// \brief Invalidates all TLB entries of all types.
-    void flush_all_tlb() {
+    NO_INLINE void flush_all_tlb() {
         derived().template flush_tlb_type<TLB_CODE>();
         derived().template flush_tlb_type<TLB_READ>();
         derived().template flush_tlb_type<TLB_WRITE>();
     }
 
     /// \brief Invalidates TLB entries for a specific virtual address.
     /// \param vaddr Target virtual address.
-    void flush_tlb_vaddr(uint64_t vaddr) {
+    NO_INLINE void flush_tlb_vaddr(uint64_t vaddr) {
         return derived().do_flush_tlb_vaddr(vaddr);
     }