reloc on startup reduced

demos/build.sh

@@ -61,11 +61,22 @@ echo "compile: $BASE.c -> $OUTBASE.o"
     -O2 -ffunction-sections -c "$SRC" -o "$OBJ"
 
 echo "link: -> $OUTBASE.bin"
+# text-base 0: under the GS/OS Loader our segment lands at some bank
+# the Loader picks (KIND=0x1000), so link-time addresses can start at
+# offset 0 in the segment.  With text-base=0 the link-time-baked IMM16
+# operand for `lda gFoo` already equals gFoo's runtime offset within
+# the segment (no segPlacedBase adjustment needed since the Loader
+# places at bank-aligned bases), so the linker can skip emitting
+# cRELOC entries for every IMM16 site — 60–90% reduction in cRELOC
+# count, proportional speed-up in the Loader's per-reloc walk.
+# Bare-metal builds keep text-base=0x1000 (runInMame.sh loads at
+# bank-0:1000 and link-time addresses must match).
+#
 # bss-base 0xA000 keeps BSS above the SHR shadow region ($2000-$9FFF
 # in bank 0 mirrors to bank E1 SHR memory).  Without this, the smaller
 # section-gc'd demos place BSS at e.g. $2300 and global writes scribble
 # on the screen.
-LINKER_ARGS=(-o "$BIN" --text-base 0x1000 --bss-base 0xA000 \
+LINKER_ARGS=(-o "$BIN" --text-base 0 --bss-base 0xA000 \
     --map "$MAP" --reloc-out "$RELOC")
 if [ "$DEBUG" = 1 ]; then
     LINKER_ARGS+=(--debug-out "$DWARF")
@@ -90,7 +101,7 @@ fi
 
 echo "OMF: -> $OUTBASE.omf"
 "$OMF" --input "$BIN" --map "$MAP" \
-    --base 0x1000 --entry __start --output "$OUT" \
+    --base 0 --entry __start --output "$OUT" \
     --name "$(echo "$OUTBASE" | tr '[:lower:]' '[:upper:]' | cut -c1-8)" \
     --expressload --relocs "$RELOC"
 

src/link816/link816.cpp

@@ -548,7 +548,19 @@ static void applyReloc(std::vector<uint8_t> &buf, uint32_t off,
         // `if (weakFn) weakFn()` null test that the null is meant to fail.
         // recordCRelocSite handles the gate; byteCnt=2 distinguishes
         // from IMM24 (3) so omfEmit emits cRELOC ByteCnt=2 here.
-        recordCRelocSite(patchAddr, target, /*byteCnt=*/2, /*bitShift=*/0);
+        // Optimization: when text-base's low 16 bits are zero, the
+        // link-time baked operand (target & 0xFFFF) is already equal
+        // to (segPlacedBase + offsetRef) & 0xFFFF — because the
+        // Loader places KIND=0x1000 segments at bank-aligned bases
+        // (segPlacedBase.lo16 = 0), and we baked target = textBase +
+        // offsetRef, so the patch is a no-op.  Skipping it eliminates
+        // 50–90% of cRELOC entries in typical builds (most data refs
+        // are IMM16); the GS/OS Loader walks each cRELOC site in a
+        // linear loop, so the per-boot reloc-application time drops
+        // proportionally.  Programs that link with a non-bank-aligned
+        // text-base (e.g. bare-metal at $1000) still get the cRELOC.
+        if ((gTextBaseForSites & 0xFFFFu) != 0)
+            recordCRelocSite(patchAddr, target, /*byteCnt=*/2, /*bitShift=*/0);
         break;
     case R_W65816_BANK16:
         // 2-byte patch: byte 0 = bank of target, byte 1 = 0 (pad).