Known-small memcpy-s are still libcalls?

[nwf]

Continuing from #85 and looking at the NATIVE_LITTLE_ENDIAN option afforded by BLAKE2s's portable implementation (https://github.com/BLAKE2/libb2/blob/643decfbf8ae600c3387686754d74c84144950d1/src/blake2-impl.h#L32-L36), load32 now uses a memcpy to read four bytes.

And we dutifully turn that into an actual library call, rather than crushing it down into a clw instruction:

Disassembly of section .text.load32:

00000000 <load32>:
; {
       0: 7d 71         cincoffset      csp, csp, -16
       2: 06 e4         csc     cra, 8(csp)
       4: db 05 a5 fe   cmove   ca1, ca0
       8: 48 00         cincoffset      ca0, csp, 4
       a: 5b 25 45 00   csetbounds      ca0, ca0, 4
;   memcpy( &w, src, sizeof( w ) );
       e: 11 46         li      a2, 4

00000010 <.LBB1_1>:
      10: 97 03 00 00   auipcc  ct2, 0
                        00000010:  R_RISCV_CHERIOT_COMPARTMENT_HI       __library_import_libcalls_memcpy
                        00000010:  R_RISCV_RELAX        *ABS*
      14: 83 b3 03 00   clc     ct2, 0(ct2)
                        00000014:  R_RISCV_CHERIOT_COMPARTMENT_LO_I     .LBB1_1
                        00000014:  R_RISCV_RELAX        *ABS*
      18: 82 93         cjalr   ct2
;   return w;
      1a: 12 45         clw     a0, 4(csp)
      1c: a2 60         clc     cra, 8(csp)
      1e: 41 61         cincoffset      csp, csp, 16
      20: 82 80         cret

This, too, then fails to get inlined at -Oz and everything's still a bit sad.

[resistor]

@nwf Can you share more specifically what command line you're using that's producing that code? I'm not able to reproduce it, and am seeing codegen equivalent to #85 when using NATIVE_LITTLE_ENDIAN

[nwf]

The relevant bits of xmake output are, I think,

[ 43%]: cache compiling.release blake2s-ref.c
/cheri/out/cheriot/sdk/bin/clang -c -std=c2x -Qunused-arguments -target riscv32-unknown-unknown -mcpu=cheriot -mabi=cheriot -mxcheri-rvc -mrelax -fshort-wchar -nostdinc -Oz -g -ffunction-sections -fdata-sections -fomit-frame-pointer -fno-builtin -fno-exceptions -fno-asynchronous-unwind-tables -fno-c++-static-destructors -fno-rtti -Werror -I/cheri/source/cheriot/cheriot-rtos/sdk/include/c++-config -I/cheri/source/cheriot/cheriot-rtos/sdk/include/libc++ -I/cheri/source/cheriot/cheriot-rtos/sdk/include -I/cheri/source/cheriot/cheriot-rtos/sdk/boards/../include/platform/generic-riscv -DCHERIOT_NO_AMBIENT_MALLOC -DBLAKE2_API=__cheri_libcall -DNATIVE_LITTLE_ENDIAN -DTEMPORAL_SAFETY -DSOFTWARE_REVOKER -DSAIL -DRISCV_HTIF -DCPU_TIMER_HZ=2000 -DTICK_RATE_HZ=10 -DSIMULATION -DCONFIG_MSHWM -DDEVICE_EXISTS_clint -DDEVICE_EXISTS_shadow -DDEVICE_EXISTS_uart -DREVOKABLE_MEMORY_START=0x80000000 "-DCHERIOT_INTERRUPT_NAMES=FakeInterrupt=4, " -o ../../../../../build/cheriot/cheriot-rtos/b3sum/build.sail/.objs/libb2s/cheriot/cheriot/release/blake2s-ref.c.o blake2s-ref.c

and then

[ 72%]: linking library libb2s.library
/cheri/out/cheriot/sdk/bin/ld.lld --script=/cheri/source/cheriot/cheriot-rtos/sdk/library.ldscript --compartment --gc-sections --relax -o ../../../../../build/cheriot/cheriot-rtos/b3sum/build.sail/cheriot/cheriot/release/libb2s.library ../../../../../build/cheriot/cheriot-rtos/b3sum/build.sail/.objs/libb2s/cheriot/cheriot/release/blake2s-ref.c.o

Relative to upstream (BLAKE2/libb2@643decf), I have the following local changes to get things compiling:

1287f4a33af5% git diff
diff --git a/src/blake2-impl.h b/src/blake2-impl.h
index 115e66e..8e2c3af 100644
--- a/src/blake2-impl.h
+++ b/src/blake2-impl.h
@@ -21,7 +21,7 @@
 #include <stddef.h>
 #include <stdint.h>
 #include <string.h>
-#include "config.h"
+// #include "config.h"
 
 #define BLAKE2_IMPL_CAT(x,y) x ## y
 #define BLAKE2_IMPL_EVAL(x,y)  BLAKE2_IMPL_CAT(x,y)
@@ -154,7 +154,7 @@ static inline void secure_zero_memory(void *v, size_t n)
   explicit_memset(v, 0, n);
 #else
   memset(v, 0, n);
-  __asm__ __volatile__("" :: "r"(v) : "memory");
+  __asm__ __volatile__("" :: "C"(v) : "memory");
 #endif
 #endif
 }
diff --git a/src/blake2.h b/src/blake2.h
index 5ca17f6..00b44e4 100644
--- a/src/blake2.h
+++ b/src/blake2.h
@@ -31,6 +31,7 @@
   #define BLAKE2_DLL_PRIVATE
 #endif
 
+#if !defined(BLAKE2_API)
 #if defined(BLAKE2_DLL)
   #if defined(BLAKE2_DLL_EXPORTS) // defined if we are building the DLL
     #define BLAKE2_API BLAKE2_DLL_EXPORT
@@ -42,6 +43,7 @@
   #define BLAKE2_API
   #define BLAKE2_PRIVATE
 #endif
+#endif
 
 #if defined(__cplusplus)
 extern "C" {

[resistor]

The first reason this happens is because of the compiler flag -fno-builtin, which prevents LLVM from replacing the call to memcpy with a call to __builtin_memcpy.

Once that flag is dropped, a second reason comes into play: we're declaring memcpy with the cheri_libcall calling convention, which also blocks conversion to __builtin_memcpy.

[resistor]

With #90 this works properly when -fno-builtin is dropped.