picogame: enable on RP2040

ports/raspberrypi/boards/pajenicko_picopad/mpconfigboard.mk

@@ -50,8 +50,3 @@ CFLAGS += \
 
 # Must be accompanied by a linker script change
 CFLAGS += -DCIRCUITPY_FIRMWARE_SIZE='(1536 * 1024)'
-
-# The rp2 port default is -O3; on this Cortex-M0+ (no SIMD/FPU, 16 KB XIP cache) -O2 plus
-# these five loop passes measures within +-1% of -O3 across the picogame render kernels
-# while using ~150 KB less flash (gc.o/vm.o stay -O3 via SUPEROPT regardless).
-OPTIMIZATION_FLAGS = -O2 -funswitch-loops -fpredictive-commoning -fgcse-after-reload -ftree-partial-pre -fsplit-paths

ports/raspberrypi/boards/raspberry_pi_pico/mpconfigboard.mk

@@ -9,4 +9,7 @@ CHIP_FAMILY = rp2
 EXTERNAL_FLASH_DEVICES = "W25Q16JVxQ"
 
 CIRCUITPY__EVE = 1
+
+CIRCUITPY_PICOGAME = 1
+CIRCUITPY_PICOGAME_FAST_DISPLAY = 1
 CIRCUITPY_PICODVI = 1

ports/raspberrypi/boards/raspberry_pi_pico_w/mpconfigboard.mk

@@ -12,6 +12,9 @@ CIRCUITPY_USB_HOST = 0
 
 CIRCUITPY__EVE = 1
 
+CIRCUITPY_PICOGAME = 1
+CIRCUITPY_PICOGAME_FAST_DISPLAY = 1
+
 CIRCUITPY_CYW43 = 1
 CIRCUITPY_SSL = 1
 CIRCUITPY_HASHLIB = 1

ports/raspberrypi/mpconfigport.mk

@@ -63,6 +63,33 @@ CIRCUITPY_TOUCHIO ?= 1
 
 # delay in ms before calling cyw43_arch_init_with_country
 CIRCUITPY_CYW43_INIT_DELAY ?= 1000
+
+# -O2 plus the two flags that carry -O3 here, which is 130 KB less flash than -O3 for the same
+# speed on everything measured. That headroom is what lets a board fit.
+#
+#   -funswitch-loops           Hoists a test that cannot change inside a loop out of it and keeps
+#                              one copy of the body per value. Costs 16 KB, because it duplicates
+#                              loop bodies in 94 functions. Buys 24 % on a flipped sprite blit,
+#                              9 % on a textured floor, and 1-2 % on bitmaptools.alphablend and
+#                              the ulab reductions.
+#
+#   -fvect-cost-model=dynamic  -O2 sets very-cheap, which turns nearly every loop down. dynamic
+#                              accepts them, which here means memset, memcpy and a 16-bit row
+#                              fill get unrolled with an alignment prologue. Costs 4 KB. Buys
+#                              37 % on rectangle fills and bytearray copies and 16 % on aesio.
+#                              Runs shorter than the prologue lose: a four-pixel span fill is
+#                              7 % slower.
+#
+# The four other -O3 loop passes we measured (-fpredictive-commoning, -fgcse-after-reload,
+# -ftree-partial-pre, -fsplit-paths) emit the same code as -O2 for the loops that matter here,
+# so they are left out. So is the rest of -O3: its inline parameters alone are 92 KB, and they
+# also take the 60-factorial loop 56 % slower.
+#
+# Neither flag reaches the bytecode dispatch loop: py/py.mk builds gc.o and vm.o at -O3 whatever
+# this is set to, so pure interpreter work measures the same under all of them.
+#
+# RP2350 keeps plain -O3. There an -O2 set runs 24-41 % slower on fill loops.
+OPTIMIZATION_FLAGS ?= -O2 -funswitch-loops -fvect-cost-model=dynamic
 endif
 
 ifeq ($(CHIP_VARIANT),RP2350)