Rollup merge of rust-lang#145071 - cuviper:min-llvm-20, r=nikic

Update the minimum external LLVM to 20

With this change, we'll have stable support for LLVM 20 and 21.
For reference, the previous increase to LLVM 19 was rust-lang#139275.

cc `@rust-lang/wg-llvm`
r? nikic

Author: Zalathar

compiler/rustc_codegen_llvm/src/builder.rs

@@ -1091,16 +1091,11 @@ impl<'a, 'll, 'tcx> BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         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;
-        }
-
+    ) -> Self::Value {
         let size = ty.primitive_size(self.tcx);
         let name = if ty.is_signed() { "llvm.scmp" } else { "llvm.ucmp" };
 
-        Some(self.call_intrinsic(name, &[self.type_i8(), self.type_ix(size.bits())], &[lhs, rhs]))
+        self.call_intrinsic(name, &[self.type_i8(), self.type_ix(size.bits())], &[lhs, rhs])
     }
 
     /* Miscellaneous instructions */

compiler/rustc_codegen_llvm/src/context.rs

@@ -172,35 +172,6 @@ pub(crate) unsafe fn create_module<'ll>(
     let mut target_data_layout = sess.target.data_layout.to_string();
     let llvm_version = llvm_util::get_version();
 
-    if llvm_version < (20, 0, 0) {
-        if sess.target.arch == "aarch64" || sess.target.arch.starts_with("arm64") {
-            // LLVM 20 defines three additional address spaces for alternate
-            // pointer kinds used in Windows.
-            // See https://github.com/llvm/llvm-project/pull/111879
-            target_data_layout =
-                target_data_layout.replace("-p270:32:32-p271:32:32-p272:64:64", "");
-        }
-        if sess.target.arch.starts_with("sparc") {
-            // LLVM 20 updates the sparc layout to correctly align 128 bit integers to 128 bit.
-            // See https://github.com/llvm/llvm-project/pull/106951
-            target_data_layout = target_data_layout.replace("-i128:128", "");
-        }
-        if sess.target.arch.starts_with("mips64") {
-            // LLVM 20 updates the mips64 layout to correctly align 128 bit integers to 128 bit.
-            // See https://github.com/llvm/llvm-project/pull/112084
-            target_data_layout = target_data_layout.replace("-i128:128", "");
-        }
-        if sess.target.arch.starts_with("powerpc64") {
-            // LLVM 20 updates the powerpc64 layout to correctly align 128 bit integers to 128 bit.
-            // See https://github.com/llvm/llvm-project/pull/118004
-            target_data_layout = target_data_layout.replace("-i128:128", "");
-        }
-        if sess.target.arch.starts_with("wasm32") || sess.target.arch.starts_with("wasm64") {
-            // LLVM 20 updates the wasm(32|64) layout to correctly align 128 bit integers to 128 bit.
-            // See https://github.com/llvm/llvm-project/pull/119204
-            target_data_layout = target_data_layout.replace("-i128:128", "");
-        }
-    }
     if llvm_version < (21, 0, 0) {
         if sess.target.arch == "nvptx64" {
             // LLVM 21 updated the default layout on nvptx: https://github.com/llvm/llvm-project/pull/124961

compiler/rustc_codegen_llvm/src/llvm_util.rs

@@ -246,9 +246,6 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea
         ("aarch64", "pmuv3") => Some(LLVMFeature::new("perfmon")),
         ("aarch64", "paca") => Some(LLVMFeature::new("pauth")),
         ("aarch64", "pacg") => Some(LLVMFeature::new("pauth")),
-        // Before LLVM 20 those two features were packaged together as b16b16
-        ("aarch64", "sve-b16b16") if get_version().0 < 20 => Some(LLVMFeature::new("b16b16")),
-        ("aarch64", "sme-b16b16") if get_version().0 < 20 => Some(LLVMFeature::new("b16b16")),
         ("aarch64", "flagm2") => Some(LLVMFeature::new("altnzcv")),
         // Rust ties fp and neon together.
         ("aarch64", "neon") => Some(LLVMFeature::with_dependencies(
@@ -262,57 +259,17 @@ pub(crate) fn to_llvm_features<'a>(sess: &Session, s: &'a str) -> Option<LLVMFea
         // Filter out features that are not supported by the current LLVM version
         ("aarch64", "fpmr") => None, // only existed in 18
         ("arm", "fp16") => Some(LLVMFeature::new("fullfp16")),
-        // NVPTX targets added in LLVM 20
-        ("nvptx64", "sm_100") if get_version().0 < 20 => None,
-        ("nvptx64", "sm_100a") if get_version().0 < 20 => None,
-        ("nvptx64", "sm_101") if get_version().0 < 20 => None,
-        ("nvptx64", "sm_101a") if get_version().0 < 20 => None,
-        ("nvptx64", "sm_120") if get_version().0 < 20 => None,
-        ("nvptx64", "sm_120a") if get_version().0 < 20 => None,
-        ("nvptx64", "ptx86") if get_version().0 < 20 => None,
-        ("nvptx64", "ptx87") if get_version().0 < 20 => None,
         // Filter out features that are not supported by the current LLVM version
-        ("loongarch64", "div32" | "lam-bh" | "lamcas" | "ld-seq-sa" | "scq")
-            if get_version().0 < 20 =>
-        {
-            None
-        }
         ("loongarch32" | "loongarch64", "32s") if get_version().0 < 21 => None,
-        // Filter out features that are not supported by the current LLVM version
-        ("riscv32" | "riscv64", "zacas" | "rva23u64" | "supm") if get_version().0 < 20 => None,
-        (
-            "s390x",
-            "message-security-assist-extension12"
-            | "concurrent-functions"
-            | "miscellaneous-extensions-4"
-            | "vector-enhancements-3"
-            | "vector-packed-decimal-enhancement-3",
-        ) if get_version().0 < 20 => None,
         // Enable the evex512 target feature if an avx512 target feature is enabled.
         ("x86", s) if s.starts_with("avx512") => Some(LLVMFeature::with_dependencies(
             s,
             smallvec![TargetFeatureFoldStrength::EnableOnly("evex512")],
         )),
-        // Support for `wide-arithmetic` will first land in LLVM 20 as part of
-        // llvm/llvm-project#111598
-        ("wasm32" | "wasm64", "wide-arithmetic") if get_version() < (20, 0, 0) => None,
         ("sparc", "leoncasa") => Some(LLVMFeature::new("hasleoncasa")),
-        // In LLVM 19, there is no `v8plus` feature and `v9` means "SPARC-V9 instruction available and SPARC-V8+ ABI used".
-        // https://github.com/llvm/llvm-project/blob/llvmorg-19.1.0/llvm/lib/Target/Sparc/MCTargetDesc/SparcELFObjectWriter.cpp#L27-L28
-        // Before LLVM 19, there was no `v8plus` feature and `v9` means "SPARC-V9 instruction available".
-        // https://github.com/llvm/llvm-project/blob/llvmorg-18.1.0/llvm/lib/Target/Sparc/MCTargetDesc/SparcELFObjectWriter.cpp#L26
-        ("sparc", "v8plus") if get_version().0 == 19 => Some(LLVMFeature::new("v9")),
         ("powerpc", "power8-crypto") => Some(LLVMFeature::new("crypto")),
-        // These new `amx` variants and `movrs` were introduced in LLVM20
-        ("x86", "amx-avx512" | "amx-fp8" | "amx-movrs" | "amx-tf32" | "amx-transpose")
-            if get_version().0 < 20 =>
-        {
-            None
-        }
-        ("x86", "movrs") if get_version().0 < 20 => None,
         ("x86", "avx10.1") => Some(LLVMFeature::new("avx10.1-512")),
-        ("x86", "avx10.2") if get_version().0 < 20 => None,
-        ("x86", "avx10.2") if get_version().0 >= 20 => Some(LLVMFeature::new("avx10.2-512")),
+        ("x86", "avx10.2") => Some(LLVMFeature::new("avx10.2-512")),
         ("x86", "apxf") => Some(LLVMFeature::with_dependencies(
             "egpr",
             smallvec![
@@ -716,17 +673,7 @@ pub(crate) fn global_llvm_features(
     };
 
     // Features implied by an implicit or explicit `--target`.
-    features.extend(
-        sess.target
-            .features
-            .split(',')
-            .filter(|v| !v.is_empty())
-            // Drop +v8plus feature introduced in LLVM 20.
-            // (Hard-coded target features do not go through `to_llvm_feature` since they already
-            // are LLVM feature names, hence we need a special case here.)
-            .filter(|v| *v != "+v8plus" || get_version() >= (20, 0, 0))
-            .map(String::from),
-    );
+    features.extend(sess.target.features.split(',').filter(|v| !v.is_empty()).map(String::from));
 
     if wants_wasm_eh(sess) && sess.panic_strategy() == PanicStrategy::Unwind {
         features.push("+exception-handling".into());

compiler/rustc_codegen_ssa/src/mir/rvalue.rs

@@ -901,36 +901,8 @@ 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(lhs_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
-                    // <https://graphics.stanford.edu/~seander/bithacks.html#CopyIntegerSign>
-                    // However, as of 2023-11 it optimizes worse in things like derived
-                    // `PartialOrd`, so only use it in debug for now. Once LLVM can handle it
-                    // better (see <https://github.com/llvm/llvm-project/issues/73417>), it'll
-                    // be worth trying it in optimized builds as well.
-                    let is_gt = bx.icmp(pred(mir::BinOp::Gt), lhs, rhs);
-                    let gtext = bx.zext(is_gt, bx.type_i8());
-                    let is_lt = bx.icmp(pred(mir::BinOp::Lt), lhs, rhs);
-                    let ltext = bx.zext(is_lt, bx.type_i8());
-                    bx.unchecked_ssub(gtext, ltext)
-                } else {
-                    // These operations are those expected by `tests/codegen-llvm/integer-cmp.rs`,
-                    // from <https://github.com/rust-lang/rust/pull/63767>.
-                    let is_lt = bx.icmp(pred(mir::BinOp::Lt), lhs, rhs);
-                    let is_ne = bx.icmp(pred(mir::BinOp::Ne), lhs, rhs);
-                    let ge = bx.select(
-                        is_ne,
-                        bx.cx().const_i8(Ordering::Greater as i8),
-                        bx.cx().const_i8(Ordering::Equal as i8),
-                    );
-                    bx.select(is_lt, bx.cx().const_i8(Ordering::Less as i8), ge)
-                }
+                bx.three_way_compare(lhs_ty, lhs, rhs)
             }
             mir::BinOp::AddWithOverflow
             | mir::BinOp::SubWithOverflow

compiler/rustc_codegen_ssa/src/traits/builder.rs

@@ -3,6 +3,7 @@ use std::ops::Deref;
 
 use rustc_abi::{Align, Scalar, Size, WrappingRange};
 use rustc_middle::middle::codegen_fn_attrs::CodegenFnAttrs;
+use rustc_middle::mir;
 use rustc_middle::ty::layout::{FnAbiOf, LayoutOf, TyAndLayout};
 use rustc_middle::ty::{AtomicOrdering, Instance, Ty};
 use rustc_session::config::OptLevel;
@@ -405,15 +406,41 @@ pub trait BuilderMethods<'a, 'tcx>:
     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
+        ty: Ty<'tcx>,
+        lhs: Self::Value,
+        rhs: Self::Value,
+    ) -> Self::Value {
+        // FIXME: This implementation was designed around LLVM's ability to optimize, but `cg_llvm`
+        // overrides this to just use `@llvm.scmp`/`ucmp` since LLVM 20. This default impl should be
+        // reevaluated with respect to the remaining backends like cg_gcc, whether they might use
+        // specialized implementations as well, or continue to use a generic implementation here.
+        use std::cmp::Ordering;
+        let pred = |op| crate::base::bin_op_to_icmp_predicate(op, ty.is_signed());
+        if self.cx().sess().opts.optimize == OptLevel::No {
+            // This actually generates tighter assembly, and is a classic trick:
+            // <https://graphics.stanford.edu/~seander/bithacks.html#CopyIntegerSign>.
+            // However, as of 2023-11 it optimized worse in LLVM in things like derived
+            // `PartialOrd`, so we were only using it in debug. Since LLVM now uses its own
+            // intrinsics, it may be be worth trying it in optimized builds for other backends.
+            let is_gt = self.icmp(pred(mir::BinOp::Gt), lhs, rhs);
+            let gtext = self.zext(is_gt, self.type_i8());
+            let is_lt = self.icmp(pred(mir::BinOp::Lt), lhs, rhs);
+            let ltext = self.zext(is_lt, self.type_i8());
+            self.unchecked_ssub(gtext, ltext)
+        } else {
+            // These operations were better optimized by LLVM, before `@llvm.scmp`/`ucmp` in 20.
+            // See <https://github.com/rust-lang/rust/pull/63767>.
+            let is_lt = self.icmp(pred(mir::BinOp::Lt), lhs, rhs);
+            let is_ne = self.icmp(pred(mir::BinOp::Ne), lhs, rhs);
+            let ge = self.select(
+                is_ne,
+                self.cx().const_i8(Ordering::Greater as i8),
+                self.cx().const_i8(Ordering::Equal as i8),
+            );
+            self.select(is_lt, self.cx().const_i8(Ordering::Less as i8), ge)
+        }
     }
 
     fn memcpy(