Separate translate object control from LTO

ports/broadcom/Makefile

@@ -107,7 +107,7 @@ OBJ += $(addprefix $(BUILD)/, $(SRC_S:.s=.o))
 OBJ += $(addprefix $(BUILD)/, $(SRC_MOD:.c=.o))
 
 # BCM CLFAGS
-CFLAGS += -ffreestanding -nostartfiles -DMICROPY_HW_MCU_NAME="\"$(CHIP_VARIANT)\""
+CFLAGS += -nostartfiles -DMICROPY_HW_MCU_NAME="\"$(CHIP_VARIANT)\""
 
 
 OPTIMIZATION_FLAGS ?= -O3

ports/cxd56/mpconfigport.mk

@@ -3,6 +3,8 @@ USB_HIGHSPEED = 1
 # Number of USB endpoint pairs.
 USB_NUM_ENDPOINT_PAIRS = 6
 
+CIRCUITPY_TRANSLATE_OBJECT = 1
+
 # Longints can be implemented as mpz, as longlong, or not
 LONGINT_IMPL = MPZ
 

py/circuitpy_defns.mk

@@ -51,7 +51,8 @@ BASE_CFLAGS = \
 	-DCIRCUITPY_SOFTWARE_SAFE_MODE=0x0ADABEEF \
 	-DCIRCUITPY_CANARY_WORD=0xADAF00 \
 	-DCIRCUITPY_SAFE_RESTART_WORD=0xDEADBEEF \
-	-DCIRCUITPY_BOARD_ID="\"$(BOARD)\""
+	-DCIRCUITPY_BOARD_ID="\"$(BOARD)\"" \
+	--param max-inline-insns-single=500
 
 #        Use these flags to debug build times and header includes.
 #        -ftime-report
@@ -73,6 +74,16 @@ else
 CFLAGS += -DCIRCUITPY_LTO=0
 endif
 
+# Produce an object file for translate.c instead of including it in a header.
+# The header version can be optimized on non-LTO builds *if* inlining is allowed
+# otherwise, it blows up the binary sizes with tons of translate copies.
+ifeq ($(CIRCUITPY_LTO), 0)
+CIRCUITPY_TRANSLATE_OBJECT ?= 0
+else
+CIRCUITPY_TRANSLATE_OBJECT ?= 1
+endif
+CFLAGS += -DCIRCUITPY_TRANSLATE_OBJECT=$(CIRCUITPY_TRANSLATE_OBJECT)
+
 ###
 # Handle frozen modules.
 

supervisor/shared/translate/translate.c

@@ -135,6 +135,6 @@ char *decompress(const compressed_string_t *compressed, char *decompressed) {
     return decompressed;
 }
 
-#if CIRCUITPY_LTO == 1
+#if CIRCUITPY_TRANSLATE_OBJECT == 1
 #include "supervisor/shared/translate/translate_impl.h"
 #endif

supervisor/shared/translate/translate.h

@@ -43,7 +43,7 @@
 // that the build process will optimize this away and replace it with the
 // appropriate compressed data for each call site.
 
-#if CIRCUITPY_LTO == 0
+#if CIRCUITPY_TRANSLATE_OBJECT == 0
 // Without LTO, we need to include a copy of this function in each compilation
 // unit so that the compile stage can do the optimization. Otherwise the linker
 // will leave this as a giant function and have each call site call into it.

supervisor/shared/translate/translate_impl.h

@@ -32,7 +32,7 @@
 
 #include "supervisor/shared/translate/compressed_string.h"
 
-#if CIRCUITPY_LTO == 0
+#if CIRCUITPY_TRANSLATE_OBJECT == 0
 static
 #endif
 inline