[libc-commits] [libc] [libc][NFC] Adjust use of off_t internally (PR #68269)

[Guillaume Chatelet via libc-commits]

================
@@ -19,15 +19,18 @@ namespace LIBC_NAMESPACE {
 
 LLVM_LIBC_FUNCTION(ssize_t, pread,
                    (int fd, void *buf, size_t count, off_t offset)) {
-#ifdef LIBC_TARGET_ARCH_IS_RISCV32
-  static_assert(sizeof(off_t) == 8);
-  ssize_t ret = LIBC_NAMESPACE::syscall_impl<ssize_t>(
-      SYS_pread64, fd, buf, count, (long)offset,
-      (long)(((uint64_t)(offset)) >> 32));
-#else
-  ssize_t ret = LIBC_NAMESPACE::syscall_impl<ssize_t>(SYS_pread64, fd, buf,
-                                                      count, offset);
-#endif
+  ssize_t ret;
+  if constexpr (sizeof(off_t) == 8 && sizeof(size_t) == 4) {
+    // This is a 32-bit system with a 64-bit offset.
+    long offset_low = static_cast<long>(offset);
+    long offset_high = static_cast<long>(offset >> 32);
----------------
gchatelet wrote:

Bystander comment, is this supposed to split offset 8 bytes into two 4 bytes ?

If so I find this code suspicious for several reasons:
 - why test `sizeof(size_t) == 4` if we're actually using the `long` type in the syscall?
 - to get the low bits I would expect to see a mask, I see a cast and so the semantic seems unclear,
 - what happens to the sign bit in case `offset` is negative?

I'm not saying the code is wrong, but it's not obviously correct either.

https://github.com/llvm/llvm-project/pull/68269