Missed optimizations when `b - a` is known nonnegative

[Explorer09]

This issue is considered a follow-up of #142283.

I just found out additional scenarios where b - a (both signed integer types) can be known nonnegative.

#include <stdint.h>
uint64_t func1(int32_t a, int32_t b) {
    if (b < a)
        a = b;
    return (int32_t)(b - a);
}
uint64_t func2(int32_t a, int32_t b) {
    if (b < a)
        a = b;
    return (uint32_t)(b - a);
}
uint64_t func3(int32_t a, int32_t b) {
    return (b < a ? 0 : (int32_t)(b - a));
}
uint64_t func4(int32_t a, int32_t b) {
    return (b < a ? 0 : (uint32_t)(b - a));
}

All four functions are equivalent. b - a should be known nonnegative in all four cases. In x86-64, for example, it would make shorter code by zero-extending instead of sign-extending b - a.

The optimizer should also be able to recognize b - b and optimize to constant 0, but that's not part of the issue I'm filing now.

[RKSimon]

define range(i64 -2147483648, 2147483648) i64 @func1(i32 noundef %a, i32 noundef %b) {
entry:
  %spec.select = tail call i32 @llvm.smin.i32(i32 %b, i32 %a)
  %sub = sub nsw i32 %b, %spec.select
  %conv = sext i32 %sub to i64
  ret i64 %conv
}

define range(i64 0, 4294967296) i64 @func2(i32 noundef %a, i32 noundef %b) {
entry:
  %spec.select = tail call i32 @llvm.smin.i32(i32 %b, i32 %a)
  %sub = sub nsw i32 %b, %spec.select
  %conv = zext i32 %sub to i64
  ret i64 %conv
}

define range(i64 -2147483648, 2147483648) i64 @func3(i32 noundef %a, i32 noundef %b) {
entry:
  %cmp = icmp slt i32 %b, %a
  %sub = sub nsw i32 %b, %a
  %cond = select i1 %cmp, i32 0, i32 %sub
  %conv = sext i32 %cond to i64
  ret i64 %conv
}

define range(i64 0, 4294967296) i64 @func4(i32 noundef %a, i32 noundef %b) {
entry:
  %cmp = icmp slt i32 %b, %a
  %sub = sub nsw i32 %b, %a
  %cond = select i1 %cmp, i32 0, i32 %sub
  %conv = zext i32 %cond to i64
  ret i64 %conv
}

declare i32 @llvm.smin.i32(i32, i32) #1

[jerryyiransun]

I'd like to work on this issue, can I please get assigned to it? @EugeneZelenko