Add support for hard float on M33 via CMake variable PICO_HARD_FLOAT_ABI=1

cmake/preload/toolchains/pico_arm_cortex_m33_clang.cmake

@@ -1,8 +1,17 @@
 set(CMAKE_SYSTEM_PROCESSOR cortex-m33)
 
-# these are all the directories under LLVM embedded toolchain for ARM (newlib or pibolibc) and under llvm_libc
-set(PICO_CLANG_RUNTIMES armv8m.main_soft_nofp armv8m.main_soft_nofp_unaligned armv8m.main-unknown-none-eabi)
+set(PICO_COMMON_LANG_FLAGS "-mcpu=cortex-m33 --target=armv8m.main-none-eabi -march=armv8m.main+fp+dsp")
+
+if (PICO_HARD_FLOAT_ABI)
+    set(PICO_COMMON_LANG_FLAGS "${PICO_COMMON_LANG_FLAGS} -mfloat-abi=hard")
+    # todo - doesn't seem to be a hard_fp variant for google atm?
+    # these are all the directories under LLVM embedded toolchain for ARM (newlib or pibolibc)
+    set(PICO_CLANG_RUNTIMES armv8m.main_hard_fp armv8m.main_hard_fp_unaligned)
+else()
+    set(PICO_COMMON_LANG_FLAGS "${PICO_COMMON_LANG_FLAGS} -mfloat-abi=softfp")
+    # these are all the directories under LLVM embedded toolchain for ARM (newlib or pibolibc) and under llvm_libc
+    set(PICO_CLANG_RUNTIMES armv8m.main_soft_nofp armv8m.main_soft_nofp_unaligned armv8m.main-unknown-none-eabi)
+endif()
 
-set(PICO_COMMON_LANG_FLAGS "-mcpu=cortex-m33 --target=armv8m.main-none-eabi -mfloat-abi=softfp -march=armv8m.main+fp+dsp")
 set(PICO_DISASM_OBJDUMP_ARGS --mcpu=cortex-m33 --arch=armv8m.main+fp+dsp)
 include(${CMAKE_CURRENT_LIST_DIR}/util/pico_arm_clang_common.cmake)

cmake/preload/toolchains/pico_arm_cortex_m33_gcc.cmake

@@ -2,7 +2,13 @@ set(CMAKE_SYSTEM_PROCESSOR cortex-m33)
 set(PICO_DEFAULT_GCC_TRIPLE arm-none-eabi)
 
 set(PICO_COMMON_LANG_FLAGS " -mcpu=cortex-m33 -mthumb -march=armv8-m.main+fp+dsp")
-set(PICO_COMMON_LANG_FLAGS "${PICO_COMMON_LANG_FLAGS} -mfloat-abi=softfp")
+# PICO_CMAKE_CONFIG: PICO_HARD_FLOAT_ABI, use hard floating point ABI, type=bool, default=0, group=build, docref=cmake-toolchain-config
+if (PICO_HARD_FLOAT_ABI)
+    set(PICO_COMMON_LANG_FLAGS "${PICO_COMMON_LANG_FLAGS} -mfloat-abi=hard")
+else()
+    set(PICO_COMMON_LANG_FLAGS "${PICO_COMMON_LANG_FLAGS} -mfloat-abi=softfp")
+endif()
+
 # PICO_CMAKE_CONFIG: PICO_NO_CMSE, Disable CMSE compiler extensions, type=bool, default=0, group=build, docref=cmake-toolchain-config
 if (NOT PICO_NO_CMSE)
     set(PICO_COMMON_LANG_FLAGS "${PICO_COMMON_LANG_FLAGS} -mcmse")

cmake/preload/toolchains/util/pico_arm_clang_common.cmake

@@ -85,7 +85,9 @@ if (PICO_CLIB STREQUAL "llvm_libc")
     # TODO: Remove -nostdlib++ once we include libc++ in the toolchain.
     # TODO: Move -nostartfiles to the appropriate library.
     foreach(TYPE IN ITEMS EXE SHARED MODULE)
-        set(CMAKE_${TYPE}_LINKER_FLAGS_INIT "-nostdlib++ -nostartfiles")
+        # note --unwindlib=none is only needed on recent compiler/lib versions, however just produces a
+        # warning on earlier versions, so not attempting a version check for now
+        set(CMAKE_${TYPE}_LINKER_FLAGS_INIT "-nostdlib++ -nostartfiles --unwindlib=none")
     endforeach()
 else()
     if (NOT PICO_COMPILER_SYSROOT)

src/rp2_common/pico_double/CMakeLists.txt

@@ -22,7 +22,6 @@ if (NOT TARGET pico_double)
     target_link_libraries(pico_double INTERFACE
             $<IF:$<BOOL:$<TARGET_PROPERTY:PICO_TARGET_DOUBLE_IMPL>>,$<TARGET_PROPERTY:PICO_TARGET_DOUBLE_IMPL>,${PICO_DEFAULT_DOUBLE_IMPL}>)
 
-    # todo maybe this should be pico_double_dcp (note FreeRTOS has a dependency on the LIB_PICO_DOUBLE_FOO)
     pico_add_library(pico_double_pico)
     if (PICO_RP2040)
         target_sources(pico_double_pico INTERFACE
@@ -31,17 +30,20 @@ if (NOT TARGET pico_double)
                 ${CMAKE_CURRENT_LIST_DIR}/double_math.c
                 ${CMAKE_CURRENT_LIST_DIR}/double_v1_rom_shim_rp2040.S
         )
+        target_link_libraries(pico_double_pico INTERFACE pico_bootrom pico_double_headers hardware_divider)
     elseif(NOT PICO_RISCV)
-        target_sources(pico_double_pico INTERFACE
+        pico_add_library(pico_double_pico_dcp)
+        target_sources(pico_double_pico_dcp INTERFACE
                 ${CMAKE_CURRENT_LIST_DIR}/double_math.c
                 ${CMAKE_CURRENT_LIST_DIR}/double_aeabi_dcp.S
                 ${CMAKE_CURRENT_LIST_DIR}/double_fma_dcp.S
                 ${CMAKE_CURRENT_LIST_DIR}/double_sci_m33.S
                 ${CMAKE_CURRENT_LIST_DIR}/double_conv_m33.S
                 )
+        target_link_libraries(pico_double_pico_dcp INTERFACE pico_double_headers)
+        target_link_libraries(pico_double_pico INTERFACE pico_double_pico_dcp)
     endif()
 
-    target_link_libraries(pico_double_pico INTERFACE pico_bootrom pico_double_headers hardware_divider)
 
     pico_add_library(pico_double_none)
     target_sources(pico_double_none INTERFACE
@@ -119,7 +121,7 @@ if (NOT TARGET pico_double)
         pico_wrap_function(${TARGET} fma)
     endfunction()
 
-    wrap_double_functions(pico_double_pico)
+    wrap_double_functions(pico_double_pico_dcp)
     wrap_double_functions(pico_double_none)
 
     macro(pico_set_double_implementation TARGET IMPL)

src/rp2_common/pico_double/double_aeabi_dcp.S

@@ -114,7 +114,14 @@ saving_func wrapper __aeabi_dmul
 
 double_section ddiv_fast
 saving_func regular ddiv_fast
+#if defined(__ARM_PCS_VFP)
+  vmov r0,r1,d0
+  vmov r2,r3,d1
+#endif
   dcp_ddiv_fast_m r0,r1,r0,r1,r2,r3,r0,r1,r2,r3,r12
+#if defined(__ARM_PCS_VFP)
+  vmov d0,r0,r1
+#endif
   saving_func_return
 
 double_wrapper_section __aeabi_ddiv
@@ -125,40 +132,99 @@ saving_func wrapper __aeabi_ddiv
 
 double_section sqrt_fast
 saving_func regular sqrt_fast
+#if defined(__ARM_PCS_VFP)
+  vmov r0,r1,d0
+#endif
   dcp_dsqrt_fast_m r0,r1,r0,r1,r0,r1,r2,r3,r12
+#if defined(__ARM_PCS_VFP)
+  vmov d0,r0,r1
+#endif
   saving_func_return
 
 double_wrapper_section sqrt
 saving_func wrapper sqrt
+#if defined(__ARM_PCS_VFP)
+  vmov r0,r1,d0
+#endif
 @ with correct rounding
   dcp_dsqrt_m r0,r1,r0,r1,r0,r1,r2,r3,r12
+#if defined(__ARM_PCS_VFP)
+  vmov d0,r0,r1
+#endif
   saving_func_return
 
 double_section dclassify
 saving_func regular dclassify
+#if defined(__ARM_PCS_VFP)
+  vmov r0,r1,d0
+#endif
   dcp_dclassify_m apsr_nzcv,r0,r1
   saving_func_return
 
 // ============== CONVERSION FUNCTIONS ===============
 
 double_wrapper_section __aeabi_d2f
+#if defined(__ARM_PCS_VFP)
+saving_func wrapper __aeabi_d2f
+#else
 saving_func wrapper __aeabi_d2f double2float
+#endif
 @ with rounding
   dcp_double2float_m r0,r0,r1
   saving_func_return
 
+#if defined(__ARM_PCS_VFP)
+double_section double2float
+regular_func double2float
+  push {lr}
+  vmov r0,r1,d0
+  bl WRAPPER_FUNC_NAME(__aeabi_d2f)
+  vmov s0,r0
+  pop {pc}
+#endif
+
 double_wrapper_section __aeabi_i2d
+#if defined(__ARM_PCS_VFP)
+saving_func wrapper __aeabi_i2d
+#else
 saving_func wrapper __aeabi_i2d int2double
+#endif
   dcp_int2double_m r0,r1,r0
   saving_func_return
 
+#if defined(__ARM_PCS_VFP)
+double_section int2double
+regular_func int2double
+  push {lr}
+  bl WRAPPER_FUNC_NAME(__aeabi_i2d)
+  vmov d0,r0,r1
+  pop {pc}
+#endif
+
 double_wrapper_section __aeabi_ui2d
+#if defined(__ARM_PCS_VFP)
+saving_func wrapper __aeabi_ui2d
+#else
 saving_func wrapper __aeabi_ui2d uint2double
+#endif
   dcp_uint2double_m r0,r1,r0
   saving_func_return
 
+#if defined(__ARM_PCS_VFP)
+double_section uint2double
+regular_func uint2double
+  push {lr}
+  bl WRAPPER_FUNC_NAME(__aeabi_ui2d)
+  vmov d0,r0,r1
+  pop {pc}
+#endif
+
 double_section double2fix_z
 saving_func regular double2fix_z
+#if defined(__ARM_PCS_VFP)
+  mov r2, r0
+  vmov r0, r1, d0
+#endif
   ubfx r3, r1, #20, #11
   adds r3, r2
   beq 1f // very small; we don't care that we might make a denormal
@@ -173,6 +239,10 @@ saving_func regular double2fix_z
 
 double_section double2ufix
 saving_func regular double2ufix_z double2ufix
+#if defined(__ARM_PCS_VFP)
+  mov r2, r0
+  vmov r0, r1, d0
+#endif
 double2ufix_z_entry:
   ubfx r3, r1, #20, #11
   adds r3, r2
@@ -188,6 +258,10 @@ double2ufix_z_entry:
 
 double_section double2fix
 saving_func regular double2fix
+#if defined(__ARM_PCS_VFP)
+  mov r2, r0
+  vmov r0, r1, d0
+#endif
   ubfx r3, r1, #20, #11
   cbz r3, 2f // 0 or denormal
   adds r3, r2
@@ -207,6 +281,9 @@ saving_func_return
 
 double_section double2int
 saving_func regular double2int
+#if defined(__ARM_PCS_VFP)
+  vmov r0, r1, d0
+#endif
 double2int_entry:
   lsls r2, r1, #1
   bcc double2int_z_entry // positive is ok for int64_z
@@ -230,29 +307,62 @@ double2int_entry:
 saving_func_return
 
 double_wrapper_section __aeabi_d2iz
+#if defined(__ARM_PCS_VFP)
+saving_func wrapper __aeabi_d2iz
+#else
 saving_func wrapper __aeabi_d2iz double2int_z
+#endif
 double2int_z_entry:
 @ with truncation towards 0
   dcp_double2int_m r0,r0,r1
   // note: this works with either saved or not saved call as it is just a `bx lr`
   saving_func_return
 
+#if defined(__ARM_PCS_VFP)
+double_section double2int_z
+regular_func double2int_z
+  vmov r0,r1,d0
+  push {r4,lr}
+  bl WRAPPER_FUNC_NAME(__aeabi_d2iz)
+  pop {r4,pc}
+#endif
+
 double_wrapper_section __aeabi_d2uiz
+#if defined(__ARM_PCS_VFP)
+saving_func wrapper __aeabi_d2uiz
+#else
 saving_func wrapper __aeabi_d2uiz double2uint double2uint_z
+#endif
 double2uint_z_entry:
 @ with truncation towards 0
   dcp_double2uint_m r0,r0,r1
   saving_func_return
 
+#if defined(__ARM_PCS_VFP)
+double_section double2uint
+regular_func double2uint
+regular_func double2uint_z
+  vmov r0,r1,d0
+  push {r4,lr}
+  bl WRAPPER_FUNC_NAME(__aeabi_d2uiz)
+  pop {r4,pc}
+#endif
+
 double_section double2int_r
 saving_func regular double2int_r
 @ with rounding
+#if defined(__ARM_PCS_VFP)
+  vmov r0,r1,d0
+#endif
   dcp_double2int_r_m r0,r0,r1
   saving_func_return
 
 double_section double2uint_r
 saving_func regular double2uint_r
 @ with rounding
+#if defined(__ARM_PCS_VFP)
+  vmov r0,r1,d0
+#endif
   dcp_double2uint_r_m r0,r0,r1
   saving_func_return