Blockfill optimization fixes potpourri

backends/asm/optimize_ir.c

@@ -2353,6 +2353,12 @@ FindPrevSetterForReplace(IR *irorig, Operand *dst)
 #endif
     if (SrcOnlyHwReg(dst))
         return NULL;
+
+    unsigned check_flags = FlagsUsedByCond(irorig->cond);
+    // Currently flags_safe makes some functions worse
+    // Because dead sets can't be eliminated correctly in conditional sequences
+    bool flags_safe = curfunc->optimize_flags & OPT_EXPERIMENTAL;
+
     for (ir = irorig->prev; ir; ir = ir->prev) {
         if (IsDummy(ir)) {
             continue;
@@ -2368,8 +2374,12 @@ FindPrevSetterForReplace(IR *irorig, Operand *dst)
         if (IsCallThatUsesReg(ir,dst)) {
             return NULL;
         }
+        if (InstrSetsFlags(ir,check_flags)) {
+            // This needs to go first, since the set IR also setting flags is unsafe
+            flags_safe = false;
+        }
         if (ir->dst == dst && (InstrSetsDst(ir) || ir->opc == OPC_TEST || ir->opc == OPC_TESTBN)) {
-            if (ir->cond != COND_TRUE) {
+            if (flags_safe ? !CondIsSubset(ir->cond,irorig->cond) : ir->cond != COND_TRUE) {
                 // cannot be sure that we set the value here,
                 // since the set is conditional
                 return NULL;
@@ -5276,9 +5286,18 @@ OptimizeLongfill(IRList *irl) {
         IR *prevset;
         int32_t setval;
         if (IsDummy(ir)) continue;
-        if (ir->opc == OPC_CALL && !CondIsSubset(COND_C,ir->cond) && !strcmp(ir->dst->name,"builtin_longfill_")
+        if (ir->opc == OPC_CALL && CondIsSubset(ir->cond,(IRCond)((ir->cond>>2)|COND_NC)) && !strcmp(ir->dst->name,"builtin_longfill_")
                 && (prevset = FindPrevSetterForReplace(ir,GetArgReg(1)))
                 && isConstMove(prevset,&setval)) {
+
+            // The CondIsSubset(ir->cond,(ir->cond>>2)|COND_NC)
+            // bit above requires explanation:
+            // Some conditions can't be combined with COND_NC.
+            // Those are the ones where there are some condition states(of the 4 possible)
+            // that execute the orignal condition,
+            // but after clearing C do not neccessarily execute the combined condition
+            // The bitwise operation is somewhat mindboggling.
+
             int addr = ir->addr; // Some opts require addresses to be sorta-correct;
             // Since we replace a funccall, we can clobber flags and args
             IR *sub = NewIR(OPC_SUB);
@@ -5288,9 +5307,8 @@ OptimizeLongfill(IRList *irl) {
             sub->addr = addr;
             sub->cond = ir->cond;
             IR *setq = NewIR(OPC_SETQ);
-            setq->cond = COND_NC;
-            setq->dst = GetArgReg(2);
             setq->cond = (IRCond)(COND_NC | ir->cond);
+            setq->dst = GetArgReg(2);
             setq->addr = addr;
             IR *wrlong = NewIR(OPC_WRLONG);
             wrlong->cond = (IRCond)(COND_NC | ir->cond);
@@ -5352,9 +5370,7 @@ OptimizeIRLocal(IRList *irl, Function *f)
 
     // multiply divide optimization need only be performed once,
     // and should be done before other optimizations confuse things
-    OptimizeMulDiv(irl);
-    // Similarily for longfill (opt can only become available from inlining)
-    OptimizeLongfill(irl);
+    OPT_PASS(OptimizeMulDiv(irl));
 again:
     do {
         change = 0;
@@ -5396,6 +5412,7 @@ OptimizeIRLocal(IRList *irl, Function *f)
             OPT_PASS(OptimizeP2(irl));
         }
         if (gl_p2) {
+            OPT_PASS(OptimizeLongfill(irl));
             OPT_PASS(FixupLoneCORDIC(irl));
             if (flags & OPT_CONST_PROPAGATE) {
                 OPT_PASS(CORDICconstPropagate(irl));

sys/gcalloc.spin

@@ -223,11 +223,8 @@ pri _gc_doalloc(size, reserveflag) : ptr | zptr
 
   if ptr
     ' zero the returned memory
-    size := ((size << pagesizeshift) - headersize)/4
-    zptr := ptr ' skip the header
-    repeat size
-      long[zptr] := 0
-      zptr += 4
+    size := ((size << pagesizeshift) - headersize)+/4
+    longfill(ptr,0,size)
   return ptr
 
 '