Merge pull request #2812 from colinleroy/Opt-ldptr1

Optimize ldax?sp/sta/stx to ldptr1?sp

libsrc/runtime/ldptr1sp.s

@@ -0,0 +1,20 @@
+;
+; Colin Leroy-Mira, 2025-07-26
+;
+; CC65 runtime: Load ptr1 from offset in stack
+;
+
+        .export         ldptr10sp, ldptr1ysp
+        .importzp       c_sp, ptr1
+
+; Beware: The optimizer knows about the value in Y after return!
+
+ldptr10sp:
+        ldy     #1
+ldptr1ysp:
+        lda     (c_sp),y        ; get high byte
+        sta     ptr1+1          ; and save it
+        dey                     ; point to lo byte
+        lda     (c_sp),y        ; load low byte
+        sta     ptr1
+        rts

src/cc65/codeinfo.c

@@ -190,6 +190,8 @@ static const FuncInfo FuncInfoTable[] = {
     { "ldeaxi",     REG_AX,                 PSTATE_ALL | REG_EAXY | REG_PTR1        },
     { "ldeaxidx",   REG_AXY,                PSTATE_ALL | REG_EAXY | REG_PTR1        },
     { "ldeaxysp",   SLV_IND | REG_Y,        PSTATE_ALL | REG_EAXY                   },
+    { "ldptr10sp",  SLV_TOP,                PSTATE_ALL | REG_AY | REG_PTR1          },
+    { "ldptr1ysp",  REG_Y | SLV_TOP,        PSTATE_ALL | REG_AY | REG_PTR1          },
     { "leaa0sp",    REG_SP | REG_A,         PSTATE_ALL | REG_AX                     },
     { "leaaxsp",    REG_SP | REG_AX,        PSTATE_ALL | REG_AX                     },
     { "leave",      REG_SP,                 PSTATE_ALL | REG_SP | REG_Y             },

src/cc65/codeopt.c

@@ -178,6 +178,7 @@ static OptFunc DOptPtrLoad17    = { OptPtrLoad17,    "OptPtrLoad17",    190, 0,
 static OptFunc DOptPtrLoad18    = { OptPtrLoad18,    "OptPtrLoad18",    100, 0, 0, 0, 0, 0 };
 static OptFunc DOptPtrLoad19    = { OptPtrLoad19,    "OptPtrLoad19",     65, 0, 0, 0, 0, 0 };
 static OptFunc DOptPtrLoad2     = { OptPtrLoad2,     "OptPtrLoad2",     100, 0, 0, 0, 0, 0 };
+static OptFunc DOptPtrLoad20    = { OptPtrLoad20,    "OptPtrLoad20",     90, 0, 0, 0, 0, 0 };
 static OptFunc DOptPtrLoad3     = { OptPtrLoad3,     "OptPtrLoad3",     100, 0, 0, 0, 0, 0 };
 static OptFunc DOptPtrLoad4     = { OptPtrLoad4,     "OptPtrLoad4",     100, 0, 0, 0, 0, 0 };
 static OptFunc DOptPtrLoad5     = { OptPtrLoad5,     "OptPtrLoad5",      50, 0, 0, 0, 0, 0 };
@@ -306,6 +307,7 @@ static OptFunc* OptFuncs[] = {
     &DOptPtrLoad18,
     &DOptPtrLoad19,
     &DOptPtrLoad2,
+    &DOptPtrLoad20,
     &DOptPtrLoad3,
     &DOptPtrLoad4,
     &DOptPtrLoad5,
@@ -904,6 +906,8 @@ static unsigned RunOptGroup7 (CodeSeg* S)
         Changes += RunOptFunc (S, &DOptTransfers3, 1);
     }
 
+    Changes += RunOptFunc (S, &DOptPtrLoad20, 1);
+
     /* Adjust branch distances */
     Changes += RunOptFunc (S, &DOptBranchDist, 3);
 

src/cc65/coptptrload.c

@@ -1671,3 +1671,64 @@ unsigned OptPtrLoad19 (CodeSeg* S)
     /* Return the number of changes made */
     return Changes;
 }
+
+
+
+unsigned OptPtrLoad20 (CodeSeg* S)
+/* Search for the sequence:
+**
+**      jsr     ldax?sp
+**      sta     ptr1
+**      stx     ptr1+1
+**
+** and replace it by:
+**
+**      jsr    ldptr1?sp
+*/
+{
+    unsigned Changes = 0;
+    unsigned I;
+
+    /* Walk over the entries */
+    I = 0;
+    while (I < CS_GetEntryCount (S)) {
+
+        CodeEntry* E[3];
+
+        /* Get the next entry */
+        E[0] = CS_GetEntry (S, I);
+
+        if ((CE_IsCallTo(E[0], "ldax0sp") ||
+             CE_IsCallTo(E[0], "ldaxysp"))      &&
+            CS_GetEntries (S, E+1, I+1, 2) != 0 &&
+            E[1]->OPC == OP65_STA               &&
+            strcmp (E[1]->Arg, "ptr1") == 0     &&
+            E[2]->OPC == OP65_STX               &&
+            strcmp (E[2]->Arg, "ptr1+1") == 0   &&
+            !CS_RangeHasLabel (S, I+1, 2)) {
+
+            if (strcmp (E[0]->Arg, "ldaxysp") == 0) {
+                CE_SetArg (E[0], "ldptr1ysp");
+            } else {
+                CE_SetArg (E[0], "ldptr10sp");
+            }
+            /* Delete the sta/stx */
+            CS_DelEntries (S, I+1, 2);
+
+            /* Regenerate register info */
+            CS_GenRegInfo (S);
+
+            /* Remember we had changes */
+            ++Changes;
+
+        } else {
+
+            /* Next entry */
+            ++I;
+        }
+
+    }
+
+    /* Return the number of changes made */
+    return Changes;
+}

src/cc65/coptptrload.h

@@ -405,6 +405,19 @@ unsigned OptPtrLoad19 (CodeSeg* S);
 */
 
 
+unsigned OptPtrLoad20 (CodeSeg* S);
+/* Search for the sequence:
+**
+**      jsr     ldax?sp
+**      sta     ptr1
+**      stx     ptr1+1
+**
+** and replace it by:
+**
+**      jsr    ldptr1?sp
+*/
+
+
 /* End of coptptrload.h */
 
 #endif