feat: optimize GCC interpret flags

src/Makefile

@@ -218,17 +218,21 @@ ifneq (,$(filter gcc,$(CC)))
 # 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 -fpeel-loops
-# GCC manual says that we should disable gcse when using computed gotos
-INTERPRET_CXXFLAGS+=-fno-gcse
+INTERPRET_CXXFLAGS+=-fgcse-after-reload -fpeel-loops
 # The interpreter dispatch loop performs better as a big inlined function
 INTERPRET_CXXFLAGS+=-finline-limit=1024
-# The following optimization improves register allocation in the interpret hot loop
-INTERPRET_CXXFLAGS+=-funroll-loops
-INTERPRET_CXXFLAGS+=$(MYINTERPRET_CXXFLAGS)
+# 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 ordered by hand, we don't want the compiler to shuffle it
+INTERPRET_CXXFLAGS+=-freorder-blocks-algorithm=simple
+# The following is known to save instructions in the hop loop
+INTERPRET_CXXFLAGS+=-fgcse-sm
 endif
 endif
 
+# Make testing new optimization options easier
+INTERPRET_CXXFLAGS+=$(MYINTERPRET_CXXFLAGS)
+
 # Link time optimizations
 ifeq ($(lto),yes)
 OPTFLAGS+=-flto=auto

src/interpret.cpp

@@ -2011,7 +2011,7 @@ static NO_INLINE execute_status write_csr_satp(STATE_ACCESS &a, uint64_t val) {
 }
 
 template <typename STATE_ACCESS>
-static execute_status write_csr_mstatus(STATE_ACCESS &a, uint64_t val) {
+static NO_INLINE execute_status write_csr_mstatus(STATE_ACCESS &a, uint64_t val) {
     const uint64_t old_mstatus = a.read_mstatus() & MSTATUS_R_MASK;
 
     // M-mode software can determine whether a privilege mode is implemented