[PATCH] D90218: Fix PR47973: Addressing integer division edge case with INT_MIN

[Ayshe Kuran via Phabricator via llvm-commits]

akuran added inline comments.


================
Comment at: compiler-rt/lib/builtins/int_div_impl.inc:77
+  fixint_t s_b = b >> N;                            // s_b = b < 0 ? -1 : 0
+  fixuint_t a_u = (fixuint_t)(a ^ s_a) + (-s_a);    // negate if s_a == -1
+  fixuint_t b_u = (fixuint_t)(b ^ s_b) + (-s_b);    // negate if s_b == -1
----------------
sepavloff wrote:
> What is the purpose of using addition+negation instead of subtraction?
(fixuint_t)a^s_a will always be some unsigned value, and with just subtraction we have <unsigned> - s_a. And when s_a = -1, s_a will get promoted to an unsigned value UINT_MAX, resulting in <unsigned> - UINT_MAX which can result in undefined behaviour. By negating s_a first we'd ensure that the operation there is always just either +0 or +1.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D90218/new/

https://reviews.llvm.org/D90218