Added peek-args optimization flag

Author: totalspectrum

Changelog.txt

@@ -1,4 +1,5 @@
 Version 7.0.4
+- Added `-Opeek-args` optimization flag (the optimization was always there, this provides a way to disable it)
 - Fixed an issue with casting to different size in nucode
 - Fixed incorrect line info in some DEBUG statements
 - Fixed incorrect re-use of parameter registers near functions that have an ABORT or THROW

backends/asm/optimize_ir.c

@@ -974,11 +974,13 @@ static bool isMulDivFunc(Operand *func) {
 // Set "actually" if you only care about the call actually using the current value
 static bool FuncUsesArgEx(Operand *func, Operand *arg, bool actually)
 {
+    bool peek_into_func = curfunc ? (curfunc->optimize_flags & OPT_PEEK_ARGS)
+        : (gl_optimize_flags & OPT_PEEK_ARGS);
     if (isMulDivFunc(func)) {
         return (arg == muldiva || arg == muldivb);
     } else if (arg == muldiva || arg == muldivb) {
         return !actually;
-    } else if (func && func->val) {
+    } else if (func && func->val && peek_into_func) {
         Function *funcObj = (Function *)func->val;
         if ((/*func->kind == IMM_COG_LABEL ||*/ func->kind == IMM_HUB_LABEL) && (actually || funcObj->is_leaf || FuncData(funcObj)->effectivelyLeaf) && !funcObj->has_throw) {
             if (arg->kind != REG_ARG) return true; // subreg or smth

cmdline.c

@@ -542,6 +542,7 @@ static FlagTable optflag[] = {
     { "cold-code", OPT_COLD_CODE},
     { "spin-relax-memory", OPT_SPIN_RELAXMEM},
     { "fast-inline-asm", OPT_FASTASM },
+    { "peek-args", OPT_PEEK_ARGS },
     { "experimental", OPT_EXPERIMENTAL },
     { "all", OPT_FLAGS_ALL },
 };

doc/general.md

@@ -439,6 +439,13 @@ If a method is called only once in a whole program, it is expanded inline at the
 
 If two functions have identical code, output only one copy of the function. Currently only checked in the assembly backend, bytecode is still to come.
 
+### Peek into function calls to see which registers are used (-O2, -Opeek-args)
+
+Looks into function definitions to determine whether an argument register is
+used there. By default it is assumed that all argument registers may be
+changed within a function, but with this optimization we check for some special
+cases to save having to copy arguments to other registers.
+
 ### Common Subexpression Elimination (-O2, -Ocse)
 
 Code like:

frontends/common.h

@@ -195,6 +195,7 @@ extern int gl_optimize_flags; /* flags for optimization */
 #define OPT_MERGE_DUPLICATES    0x00800000  /* merge duplicate functions */
 #define OPT_SPIN_RELAXMEM       0x01000000  /* relax strict memory semantics for Spin */
 #define OPT_FASTASM             0x02000000  /* optimize inline assembly invocation */
+#define OPT_PEEK_ARGS           0x04000000  /* peek into functions to see if arg registers can be reused */
 #define OPT_EXPERIMENTAL        0x80000000  /* gate new or experimental optimizations */
 #define OPT_FLAGS_ALL           0xffffffff
 
@@ -207,7 +208,7 @@ extern int gl_optimize_flags; /* flags for optimization */
 // default optimization (-O1) for ASM output
 #define DEFAULT_ASM_OPTS        (OPT_ASM_BASIC|OPT_DEADCODE|OPT_REMOVE_UNUSED_FUNCS|OPT_INLINE_SMALLFUNCS|OPT_AUTO_FCACHE|OPT_LOOP_BASIC|OPT_TAIL_CALLS|OPT_SPECIAL_FUNCS|OPT_CORDIC_REORDER|OPT_LOCAL_REUSE|OPT_LOOP_BASIC)
 // extras added with -O2
-#define EXTRA_ASM_OPTS          (OPT_INLINE_SINGLEUSE|OPT_PERFORM_CSE|OPT_PERFORM_LOOPREDUCE|OPT_REMOVE_HUB_BSS|OPT_EXPERIMENTAL|OPT_AGGRESSIVE_MEM|OPT_MERGE_DUPLICATES)
+#define EXTRA_ASM_OPTS          (OPT_INLINE_SINGLEUSE|OPT_PERFORM_CSE|OPT_PERFORM_LOOPREDUCE|OPT_REMOVE_HUB_BSS|OPT_EXPERIMENTAL|OPT_AGGRESSIVE_MEM|OPT_MERGE_DUPLICATES|OPT_PEEK_ARGS)
 
 // default optimization (-O1) for bytecode output; defaults to much less optimization than asm
 #define DEFAULT_BYTECODE_OPTS   (OPT_REMOVE_UNUSED_FUNCS|OPT_REMOVE_FEATURES|OPT_DEADCODE|OPT_MAKE_MACROS|OPT_SPECIAL_FUNCS|OPT_PEEPHOLE|OPT_LOOP_BASIC)