Incorrect folding of two __builtin_memcpy_inline into memmove

[miyuki]

Consider the following code implementing a 16-byte memmove:

void test(void *src, void *dest) {
  char temp[16];
  __builtin_memcpy_inline(temp, src, 16);
  __builtin_memcpy_inline(dest, temp, 16);
}

The point of memcpy_inline builtin is to avoid calls to library functions so that the builtin can be used to implement library functions such as memcpy and memmove. When targeting Arm M-profile Clang compiles this code into a call to __aeabi_memmove:

$clang -target arm-arm-none-eabi -mcpu=cortex-m7 -O2 -S test.c

produces

test:
        push    {r7, lr}
        mov     r7, sp
        mov     r3, r0
        movs    r2, #16
        mov     r0, r1
        mov     r1, r3
        bl      __aeabi_memmove
        pop     {r7, pc}

The folding happens at LLVM IR level, in MemCpyOptPass:

*** IR Dump After ADCEPass on test ***
; Function Attrs: nounwind
define dso_local void @test(ptr noundef %0, ptr noundef %1) local_unnamed_addr #0 {
  %3 = alloca [16 x i8], align 1
  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %3)
  call void @llvm.memcpy.inline.p0.p0.i64(ptr nonnull align 1 %3, ptr align 1 %0, i64 16, i1 false)
  call void @llvm.memcpy.inline.p0.p0.i64(ptr align 1 %1, ptr nonnull align 1 %3, i64 16, i1 false)
  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %3)
  ret void
}
*** IR Dump After MemCpyOptPass on test ***
; Function Attrs: nounwind
define dso_local void @test(ptr noundef %0, ptr noundef %1) local_unnamed_addr #0 {
  %3 = alloca [16 x i8], align 1
  call void @llvm.lifetime.start.p0(i64 16, ptr nonnull %3)
  call void @llvm.memcpy.inline.p0.p0.i64(ptr nonnull align 1 %3, ptr align 1 %0, i64 16, i1 false)
  call void @llvm.memmove.p0.p0.i64(ptr align 1 %1, ptr align 1 %0, i64 16, i1 false)
  call void @llvm.lifetime.end.p0(i64 16, ptr nonnull %3)
  ret void
}

[miyuki]

https://reviews.llvm.org/D147162