Auto merge of rust-lang#133984 - DaniPopes:scmp-ucmp, r=scottmcm

Lower BinOp::Cmp to llvm.{s,u}cmp.* intrinsics

Lowers `mir::BinOp::Cmp` (`three_way_compare` intrinsic) to the corresponding LLVM `llvm.{s,u}cmp.i8.*` intrinsics.

These are the intrinsics mentioned in rust-lang#118310, which are now available in LLVM 19.

I couldn't find any follow-up PRs/discussions about this, please let me know if I missed something.

r? `@scottmcm`

compiler/rustc_codegen_llvm/src/builder.rs

@@ -14,6 +14,7 @@ use rustc_codegen_ssa::mir::place::PlaceRef;
 use rustc_codegen_ssa::traits::*;
 use rustc_data_structures::small_c_str::SmallCStr;
 use rustc_hir::def_id::DefId;
+use rustc_middle::bug;
 use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs;
 use rustc_middle::ty::layout::{
     FnAbiError, FnAbiOfHelpers, FnAbiRequest, HasTypingEnv, LayoutError, LayoutOfHelpers,
@@ -1066,6 +1067,35 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         unsafe { llvm::LLVMBuildFCmp(self.llbuilder, op as c_uint, lhs, rhs, UNNAMED) }
     }
 
+    fn three_way_compare(
+        &mut self,
+        ty: Ty<'tcx>,
+        lhs: Self::Value,
+        rhs: Self::Value,
+    ) -> Option<Self::Value> {
+        // FIXME: See comment on the definition of `three_way_compare`.
+        if crate::llvm_util::get_version() < (20, 0, 0) {
+            return None;
+        }
+
+        let name = match (ty.is_signed(), ty.primitive_size(self.tcx).bits()) {
+            (true, 8) => "llvm.scmp.i8.i8",
+            (true, 16) => "llvm.scmp.i8.i16",
+            (true, 32) => "llvm.scmp.i8.i32",
+            (true, 64) => "llvm.scmp.i8.i64",
+            (true, 128) => "llvm.scmp.i8.i128",
+
+            (false, 8) => "llvm.ucmp.i8.i8",
+            (false, 16) => "llvm.ucmp.i8.i16",
+            (false, 32) => "llvm.ucmp.i8.i32",
+            (false, 64) => "llvm.ucmp.i8.i64",
+            (false, 128) => "llvm.ucmp.i8.i128",
+
+            _ => bug!("three-way compare unsupported for type {ty:?}"),
+        };
+        Some(self.call_intrinsic(name, &[lhs, rhs]))
+    }
+
     /* Miscellaneous instructions */
     fn memcpy(
         &mut self,

compiler/rustc_codegen_llvm/src/context.rs

@@ -1108,6 +1108,18 @@ impl<'ll> CodegenCx<'ll, '_> {
         ifn!("llvm.usub.sat.i64", fn(t_i64, t_i64) -> t_i64);
         ifn!("llvm.usub.sat.i128", fn(t_i128, t_i128) -> t_i128);
 
+        ifn!("llvm.scmp.i8.i8", fn(t_i8, t_i8) -> t_i8);
+        ifn!("llvm.scmp.i8.i16", fn(t_i16, t_i16) -> t_i8);
+        ifn!("llvm.scmp.i8.i32", fn(t_i32, t_i32) -> t_i8);
+        ifn!("llvm.scmp.i8.i64", fn(t_i64, t_i64) -> t_i8);
+        ifn!("llvm.scmp.i8.i128", fn(t_i128, t_i128) -> t_i8);
+
+        ifn!("llvm.ucmp.i8.i8", fn(t_i8, t_i8) -> t_i8);
+        ifn!("llvm.ucmp.i8.i16", fn(t_i16, t_i16) -> t_i8);
+        ifn!("llvm.ucmp.i8.i32", fn(t_i32, t_i32) -> t_i8);
+        ifn!("llvm.ucmp.i8.i64", fn(t_i64, t_i64) -> t_i8);
+        ifn!("llvm.ucmp.i8.i128", fn(t_i128, t_i128) -> t_i8);
+
         ifn!("llvm.lifetime.start.p0i8", fn(t_i64, ptr) -> void);
         ifn!("llvm.lifetime.end.p0i8", fn(t_i64, ptr) -> void);
 

compiler/rustc_codegen_ssa/src/mir/rvalue.rs

@@ -992,6 +992,9 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             mir::BinOp::Cmp => {
                 use std::cmp::Ordering;
                 assert!(!is_float);
+                if let Some(value) = bx.three_way_compare(input_ty, lhs, rhs) {
+                    return value;
+                }
                 let pred = |op| base::bin_op_to_icmp_predicate(op, is_signed);
                 if bx.cx().tcx().sess.opts.optimize == OptLevel::No {
                     // FIXME: This actually generates tighter assembly, and is a classic trick

compiler/rustc_codegen_ssa/src/traits/builder.rs

@@ -385,6 +385,18 @@ pub trait BuilderMethods<'a, 'tcx>:
     fn icmp(&mut self, op: IntPredicate, lhs: Self::Value, rhs: Self::Value) -> Self::Value;
     fn fcmp(&mut self, op: RealPredicate, lhs: Self::Value, rhs: Self::Value) -> Self::Value;
 
+    /// Returns `-1` if `lhs < rhs`, `0` if `lhs == rhs`, and `1` if `lhs > rhs`.
+    // FIXME: Move the default implementation from `codegen_scalar_binop` into this method and
+    // remove the `Option` return once LLVM 20 is the minimum version.
+    fn three_way_compare(
+        &mut self,
+        _ty: Ty<'tcx>,
+        _lhs: Self::Value,
+        _rhs: Self::Value,
+    ) -> Option<Self::Value> {
+        None
+    }
+
     fn memcpy(
         &mut self,
         dst: Self::Value,

tests/codegen/integer-cmp.rs

@@ -4,7 +4,7 @@
 //@ revisions: llvm-pre-20 llvm-20
 //@ [llvm-20] min-llvm-version: 20
 //@ [llvm-pre-20] max-llvm-major-version: 19
-//@ compile-flags: -C opt-level=3
+//@ compile-flags: -C opt-level=3 -Zmerge-functions=disabled
 
 #![crate_type = "lib"]
 
@@ -13,7 +13,7 @@ use std::cmp::Ordering;
 // CHECK-LABEL: @cmp_signed
 #[no_mangle]
 pub fn cmp_signed(a: i64, b: i64) -> Ordering {
-    // llvm-20: @llvm.scmp.i8.i64
+    // llvm-20: call{{.*}} i8 @llvm.scmp.i8.i64
     // llvm-pre-20: icmp slt
     // llvm-pre-20: icmp ne
     // llvm-pre-20: zext i1
@@ -24,10 +24,39 @@ pub fn cmp_signed(a: i64, b: i64) -> Ordering {
 // CHECK-LABEL: @cmp_unsigned
 #[no_mangle]
 pub fn cmp_unsigned(a: u32, b: u32) -> Ordering {
-    // llvm-20: @llvm.ucmp.i8.i32
+    // llvm-20: call{{.*}} i8 @llvm.ucmp.i8.i32
     // llvm-pre-20: icmp ult
     // llvm-pre-20: icmp ne
     // llvm-pre-20: zext i1
     // llvm-pre-20: select i1
     a.cmp(&b)
 }
+
+// CHECK-LABEL: @cmp_char
+#[no_mangle]
+pub fn cmp_char(a: char, b: char) -> Ordering {
+    // llvm-20: call{{.*}} i8 @llvm.ucmp.i8.i32
+    // llvm-pre-20: icmp ult
+    // llvm-pre-20: icmp ne
+    // llvm-pre-20: zext i1
+    // llvm-pre-20: select i1
+    a.cmp(&b)
+}
+
+// CHECK-LABEL: @cmp_tuple
+#[no_mangle]
+pub fn cmp_tuple(a: (i16, u16), b: (i16, u16)) -> Ordering {
+    // llvm-20-DAG: call{{.*}} i8 @llvm.ucmp.i8.i16
+    // llvm-20-DAG: call{{.*}} i8 @llvm.scmp.i8.i16
+    // llvm-20: ret i8
+    // llvm-pre-20: icmp slt
+    // llvm-pre-20: icmp ne
+    // llvm-pre-20: zext i1
+    // llvm-pre-20: select i1
+    // llvm-pre-20: icmp ult
+    // llvm-pre-20: icmp ne
+    // llvm-pre-20: zext i1
+    // llvm-pre-20: select i1
+    // llvm-pre-20: select i1
+    a.cmp(&b)
+}