Missed Optimization: -(x & c1) >> shamt | -(x & c1) —> x & c1 ? -1 : 0 when 2^shamt > c1 and shamt ≤ bitwidth(x)/2

[Cancelll]

define i64 @src(i32 %arg0) {
  %1 = sub i32 0, %arg0
  %2 = and i32 %1, 63
  %3 = zext nneg i32 %2 to i64
  %4 = sub nsw i64 0, %3
  %5 = lshr i64 %4, 8
  %6 = or i64 %5, %4
  ret i64 %6
}

define i64 @tgt(i32 %arg0) {
  %1 = and i32 %arg0, 63
  %2 = icmp ne i32 %1, 0
  %3 = sext i1 %2 to i64
  ret i64 %3
}

Alive2: https://alive2.llvm.org/ce/z/yLQgJp
Example: https://godbolt.org/z/hEKWGoeq3

Found this pattern in function bn_rshift_fixed_top in openssl/crypto/bn/sn_shift.c

[dtcxzyw]

I don't know how to generalize this specific fold :(

[ZhangYudongengineer]

if possible I want to try my best to solve this issue.

[el-ev]

@ZhangYudongengineer Assigned. Please check llvm.org/docs/InstCombineContributorGuide.html before making changes.

[ZhangYudongengineer]

@ZhangYudongengineer Assigned. Please check llvm.org/docs/InstCombineContributorGuide.html before making changes.

ok，thx trusting me.