Missing range analysis for `llvm.ucmp` and `llvm.scmp` intrinsics

[scottmcm]

Found investigating rust-lang/rust#133984

Given this IR:

define noundef range(i8 -1, 3) i8 @rust_i16_partial_ord(i16 noundef %0, i16 noundef %1) unnamed_addr #0 {
  %7 = tail call noundef i8 @llvm.scmp.i8.i16(i16 %0, i16 %1)
  ret i8 %7
}

define noundef zeroext i1 @check_lt_direct_before_inlining(i16 noundef %0, i16 noundef %1, i16 noundef %2, i16 noundef %3) unnamed_addr #0 {
start:
  %_3.i4.i = tail call noundef i8 @rust_i16_partial_ord(i16 %0, i16 %2)
  switch i8 %_3.i4.i, label %bb4.i [
    i8 2, label %"_ZN4core5tuple65_$LT$impl$u20$core..cmp..PartialOrd$u20$for$u20$$LP$U$C$T$RP$$GT$2lt17h933a2b8cae739748E.exit"
    i8 0, label %bb5.i
  ]

bb5.i:                                            ; preds = %start
  %_0.i.i = icmp ult i16 %1, %3
  br label %"_ZN4core5tuple65_$LT$impl$u20$core..cmp..PartialOrd$u20$for$u20$$LP$U$C$T$RP$$GT$2lt17h933a2b8cae739748E.exit"

bb4.i:                                            ; preds = %start
  %4 = icmp slt i16 %0, %2
  br label %"_ZN4core5tuple65_$LT$impl$u20$core..cmp..PartialOrd$u20$for$u20$$LP$U$C$T$RP$$GT$2lt17h933a2b8cae739748E.exit"

"_ZN4core5tuple65_$LT$impl$u20$core..cmp..PartialOrd$u20$for$u20$$LP$U$C$T$RP$$GT$2lt17h933a2b8cae739748E.exit": ; preds = %start, %bb5.i, %bb4.i
  %_0.sroa.0.0.i = phi i1 [ %_0.i.i, %bb5.i ], [ %4, %bb4.i ], [ false, %start ]
  ret i1 %_0.sroa.0.0.i
}

Today https://llvm.godbolt.org/z/Wj8cPnK3n it optimizes to

define noundef zeroext i1 @check_lt_direct_before_inlining(i16 noundef %0, i16 noundef %1, i16 noundef %2, i16 noundef %3) unnamed_addr #0 {
start:
  %4 = tail call noundef range(i8 -1, 3) i8 @llvm.scmp.i8.i16(i16 %0, i16 %2)
  switch i8 %4, label %bb4.i [
    i8 2, label %"_ZN4core5tuple65_$LT$impl$u20$core..cmp..PartialOrd$u20$for$u20$$LP$U$C$T$RP$$GT$2lt17h933a2b8cae739748E.exit"
    i8 0, label %bb5.i
  ]

bb5.i:                                            ; preds = %start
  %_0.i.i = icmp ult i16 %1, %3
  br label %"_ZN4core5tuple65_$LT$impl$u20$core..cmp..PartialOrd$u20$for$u20$$LP$U$C$T$RP$$GT$2lt17h933a2b8cae739748E.exit"

bb4.i:                                            ; preds = %start
  %5 = icmp slt i16 %0, %2
  br label %"_ZN4core5tuple65_$LT$impl$u20$core..cmp..PartialOrd$u20$for$u20$$LP$U$C$T$RP$$GT$2lt17h933a2b8cae739748E.exit"

"_ZN4core5tuple65_$LT$impl$u20$core..cmp..PartialOrd$u20$for$u20$$LP$U$C$T$RP$$GT$2lt17h933a2b8cae739748E.exit": ; preds = %bb4.i, %bb5.i, %start
  %_0.sroa.0.0.i = phi i1 [ %_0.i.i, %bb5.i ], [ %5, %bb4.i ], [ false, %start ]
  ret i1 %_0.sroa.0.0.i
}

Note, in particular, the inlined intrinsic call at the beginning:

start:
  %4 = tail call noundef range(i8 -1, 3) i8 @llvm.scmp.i8.i16(i16 %0, i16 %2)

That's an overly-broad range for scmp, which is always in [-1, 2).

As a result, follow-up passes don't optimize away that dead i8 2 arm of the switch -- I guess they trust that range and don't look at what they know scmp actually returns.

So it would be good if either:

1. inlining was smarter about the output range it put on a known intrinsic like this, or
2. some pre-inlining optimization put the correct output range on scmp, as

 define noundef range(i8 -1, 3) i8 @rust_i16_partial_ord(i16 noundef %0, i16 noundef %1) unnamed_addr #0 {
-  %7 = tail call noundef i8 @llvm.scmp.i8.i16(i16 %0, i16 %1)
+  %7 = tail call noundef range(i8 -1, 2) i8 @llvm.scmp.i8.i16(i16 %0, i16 %1)
   ret i8 %7
 }

also fixes the problem (https://llvm.godbolt.org/z/h7a55WKxK).

[nikic]

I don't think the range attribute added by inlining is really the problem here. Even if you drop it, it still doesn't fold. We're just missing some range analysis for the cmp intrinsics. (It's supported by computeConstantRange, but that's insufficient here.)