Fix embedded images built with the clang compiler

.github/workflows/test-build.yml

@@ -77,6 +77,17 @@ jobs:
         run: |
           sudo apt-get install -y gcc-arm-none-eabi gcc-aarch64-linux-gnu gcc-powerpc-linux-gnu gnu-efi
 
+      - name: Install Clang
+        if: |
+          inputs.config-file == './config/examples/stm32h7.config' ||
+          inputs.config-file == './config/examples/stm32h7-octospi.config' ||
+          inputs.config-file == './config/examples/stm32u5.config' ||
+          inputs.config-file == './config/examples/stm32u5-wolfcrypt-tz.config' ||
+          inputs.config-file == './config/examples/stm32u5-nonsecure-dualbank.config' ||
+          inputs.config-file == 'config/examples/stm32n567.config'
+        run: |
+          sudo apt-get install -y clang
+
       - name: make clean
         run: |
           make distclean
@@ -92,3 +103,17 @@ jobs:
       - name: Build wolfboot
         run: |
           make ${{inputs.make-args}}
+
+      - name: Rebuild wolfboot with Clang
+        if: |
+          inputs.config-file == './config/examples/stm32h7.config' ||
+          inputs.config-file == './config/examples/stm32h7-octospi.config' ||
+          inputs.config-file == './config/examples/stm32u5.config' ||
+          inputs.config-file == './config/examples/stm32u5-wolfcrypt-tz.config' ||
+          inputs.config-file == './config/examples/stm32u5-nonsecure-dualbank.config' ||
+          inputs.config-file == 'config/examples/stm32n567.config'
+        run: |
+          make distclean
+          cp ${{inputs.config-file}} .config
+          make -C tools/keytools && make -C tools/bin-assemble
+          make USE_CLANG=1 USE_GCC=0 ${{inputs.make-args}}

Makefile

@@ -24,6 +24,7 @@ DEBUG_UART?=0
 LIBS=
 SIGN_ALG=
 OBJCOPY_FLAGS=
+OBJCOPY_BIN_FLAGS=
 BIG_ENDIAN?=0
 USE_CLANG?=0
 ifeq ($(USE_CLANG),1)
@@ -196,6 +197,14 @@ ifeq ($(USE_GCC_HEADLESS),1)
   LSCRIPT_FLAGS+=-T $(LSCRIPT)
   OBJCOPY_FLAGS+=--gap-fill $(FILL_BYTE)
 endif
+
+ifeq ($(ARCH),ARM)
+  ifeq ($(USE_CLANG),1)
+    ifneq ($(TZEN),1)
+      OBJCOPY_BIN_FLAGS+=$(CLANG_ARM_OBJCOPY_FLASH_FLAGS_BOOT)
+    endif
+  endif
+endif
 ifeq ($(TARGET),ti_hercules)
   LSCRIPT_FLAGS+=--run_linker $(LSCRIPT)
 endif
@@ -322,7 +331,21 @@ wolfboot.efi: wolfboot.elf
 
 wolfboot.bin: wolfboot.elf
 	@echo "\t[BIN] $@"
-	$(Q)$(OBJCOPY) $(OBJCOPY_FLAGS) -O binary $^ $@
+ifeq ($(USE_CLANG),1)
+ifeq ($(TZEN),1)
+	$(Q)last_load="$$($(CROSS_COMPILE)readelf -Wl $< | awk '/ LOAD / { line = $$0 } END { print line }')"; \
+		set -- $$last_load; \
+		last_phys=$$(printf '%d' $$4); \
+		last_filesz=$$(printf '%d' $$5); \
+		padded_filesz=$$((($$last_filesz + 0xff) & ~0xff)); \
+		pad_to=$$((last_phys + padded_filesz)); \
+		$(OBJCOPY) $(OBJCOPY_FLAGS) $(OBJCOPY_BIN_FLAGS) --pad-to=$$(printf '0x%x' $$pad_to) -O binary $< $@
+else
+	$(Q)$(OBJCOPY) $(OBJCOPY_FLAGS) $(OBJCOPY_BIN_FLAGS) -O binary $^ $@
+endif
+else
+	$(Q)$(OBJCOPY) $(OBJCOPY_FLAGS) $(OBJCOPY_BIN_FLAGS) -O binary $^ $@
+endif
 	@echo
 	@echo "\t[SIZE]"
 	$(Q)$(SIZE) wolfboot.elf

arch.mk

@@ -1282,15 +1282,25 @@ ifeq ($(USE_CLANG),1)
   CLANG_NEWLIB_INCLUDE?=$(abspath $(dir $(CLANG_LIBC_A))/../include)
 
   CC=$(CLANG_DRIVER)
-  LD=$(CLANG_DRIVER)
+  # Keep clang for compilation, but use the GNU linker driver so linker-script
+  # LMAs for RAM sections are preserved and objcopy does not emit sparse images.
+  LD=$(CLANG_GCC_NAME)
   AS=$(CLANG_DRIVER)
   AR=$(CROSS_COMPILE)ar
   OBJCOPY?=$(CROSS_COMPILE)objcopy
   SIZE=$(CROSS_COMPILE)size
 
   CFLAGS+=-isystem $(CLANG_NEWLIB_INCLUDE)
   CFLAGS+=-DWOLFSSL_NO_ATOMIC -DWOLFSSL_NO_ATOMICS
-  LDFLAGS+=-nostdlib
+  CFLAGS+=-Wno-unknown-attributes -Wno-error=unknown-attributes
+  CFLAGS+=-fno-unwind-tables -fno-asynchronous-unwind-tables
+  LSCRIPT_FLAGS+=-T $(abspath $(WOLFBOOT_ROOT)/hal/clang-discard.ld)
+  # Clang-built ARM raw images may otherwise expand ELF loadable gaps into
+  # oversized binaries on some targets. Keep the section selection scoped to
+  # the Clang workaround paths only.
+  CLANG_ARM_OBJCOPY_FLASH_FLAGS_BASE:=-j .text -j .edidx
+  CLANG_ARM_OBJCOPY_FLASH_FLAGS_BOOT:=$(CLANG_ARM_OBJCOPY_FLASH_FLAGS_BASE) -j .ramcode -j .keystore -j .gnu.sgstubs
+  CLANG_ARM_OBJCOPY_FLASH_FLAGS_APP:=$(CLANG_ARM_OBJCOPY_FLASH_FLAGS_BASE)
 endif
 
 ifeq ($(USE_GCC),1)

hal/clang-discard.ld

@@ -0,0 +1,14 @@
+SECTIONS
+{
+    /DISCARD/ :
+    {
+        *(.ARM.extab* .gnu.linkonce.armextab.*)
+        *(.eh_frame* .gcc_except_table*)
+        *(.preinit_array*)
+        *(.init_array*)
+        *(.fini_array*)
+        *(.ctors*)
+        *(.dtors*)
+        *(.jcr*)
+    }
+}

hal/stm32l4.ld

@@ -15,6 +15,7 @@ SECTIONS
         . = ALIGN(4);
         _end_text = .;
     } > FLASH
+
     .edidx :
     {
         . = ALIGN(4);

options.mk

@@ -12,6 +12,9 @@ ifeq ($(USE_CLANG),1)
   ifeq ($(USE_GCC),1)
     $(error USE_CLANG=1 is incompatible with USE_GCC=1; set USE_GCC=0)
   endif
+  ifeq ($(ARMORED),1)
+    $(error USE_CLANG=1 requires ARMORED=0)
+  endif
 endif
 
 # Support for Built-in ROT into OTP flash memory

test-app/Makefile

@@ -88,7 +88,7 @@ endif
 include ../arch.mk
 
 ifeq ($(USE_CLANG),1)
-  APP_OBJS+=../src/string.o
+  APP_NEEDS_STRING:=1
 endif
 
 # Optional alias for clearer TZ PSA selection in app builds.
@@ -102,9 +102,23 @@ LDFLAGS+=-T $(LSCRIPT) -Wl,-gc-sections -Wl,-Map=image.map -nostartfiles
 
 # Setup default objcopy flags
 OBJCOPY_FLAGS+=--gap-fill $(FILL_BYTE)
+OBJCOPY_IMAGE_FLAGS:=
+
+ifeq ($(USE_CLANG),1)
+  # Clang-built ARM ELFs can keep RAM sections as loadable segments, and raw
+  # objcopy output then expands the flash-to-RAM gap into a huge sparse image.
+  # The app image only needs the flash-backed output sections.
+  ifneq ($(TZEN),1)
+    OBJCOPY_IMAGE_FLAGS+=$(CLANG_ARM_OBJCOPY_FLASH_FLAGS_APP)
+  endif
+endif
 
 ifeq ($(DEBUG_UART),1)
     CFLAGS+=-DDEBUG_UART
+    APP_NEEDS_STRING:=1
+endif
+
+ifeq ($(APP_NEEDS_STRING),1)
     APP_OBJS+=../src/string.o
 endif
 
@@ -288,7 +302,10 @@ ifeq ($(TZEN),1)
               endif
             endif
         endif
-        CFLAGS+=-DWOLFBOOT_SECURE_CALLS -Wstack-usage=19184
+        CFLAGS+=-DWOLFBOOT_SECURE_CALLS
+        ifneq ($(USE_CLANG),1)
+          CFLAGS+=-Wstack-usage=19184
+        endif
         LDFLAGS+=--specs=nosys.specs -u _printf_float
     endif
     ifeq ($(WOLFCRYPT_TZ_PSA),1)
@@ -899,15 +916,15 @@ delta-extra-data:LDFLAGS=-Wl,-Map=image.map
 
 image.bin: image.elf
 	@echo "\t[BIN] $@"
-	$(Q)$(OBJCOPY) $(OBJCOPY_FLAGS) -O binary $^ $@
+	$(Q)$(OBJCOPY) $(OBJCOPY_FLAGS) $(OBJCOPY_IMAGE_FLAGS) -O binary $^ $@
 
 image.hex: image.elf
 	@echo "\t[HEX] $@"
-	$(Q)$(OBJCOPY) $(OBJCOPY_FLAGS) -O ihex $^ $@
+	$(Q)$(OBJCOPY) $(OBJCOPY_FLAGS) $(OBJCOPY_IMAGE_FLAGS) -O ihex $^ $@
 
 image.srec: image.elf
 	@echo "\t[SREC] $@"
-	$(Q)$(OBJCOPY) $(OBJCOPY_FLAGS) -O srec $^ $@
+	$(Q)$(OBJCOPY) $(OBJCOPY_FLAGS) $(OBJCOPY_IMAGE_FLAGS) -O srec $^ $@
 
 APP_OBJS := $(patsubst $(WOLFBOOT_LIB_WOLFSSL)/%, \
     $(WOLFSSL_LOCAL_OBJDIR)/%, $(APP_OBJS))
@@ -917,7 +934,7 @@ ifeq ($(ELF_FLASH_SCATTER),1)
 SQUASHELF_TOOL = ../tools/squashelf/squashelf
 image-orig.elf: $(APP_OBJS) $(LSCRIPT)
 	@echo "\t[LD] $@"
-	$(Q)$(LD) $(LDFLAGS) $(APP_OBJS) $(LIBS) $(OUTPUT_FLAG) $@
+	$(Q)$(LD) $(LDFLAGS) $(LSCRIPT_FLAGS) $(APP_OBJS) $(LIBS) $(OUTPUT_FLAG) $@
 
 image.elf: image-orig.elf
 	@echo "\t[SQUASHELF] $@"
@@ -926,7 +943,7 @@ else
 # Default behavior when ELF_FLASH_SCATTER is not set
 image.elf: $(APP_OBJS) $(LSCRIPT)
 	@echo "\t[LD] $@"
-	$(Q)$(LD) $(LDFLAGS) $(APP_OBJS) $(LIBS) $(OUTPUT_FLAG) $@
+	$(Q)$(LD) $(LDFLAGS) $(LSCRIPT_FLAGS) $(APP_OBJS) $(LIBS) $(OUTPUT_FLAG) $@
 endif
 
 standalone: image.bin