Merge from rustc

4 files changed, 69 insertions, 52 deletions

diff --git a/compiler/rustc_codegen_ssa/src/base.rs b/compiler/rustc_codegen_ssa/src/base.rs
index f7f2bfca838..410b5d27c57 100644
--- a/compiler/rustc_codegen_ssa/src/base.rs
+++ b/compiler/rustc_codegen_ssa/src/base.rs
@@ -5,7 +5,7 @@ use crate::back::write::{
     compute_per_cgu_lto_type, start_async_codegen, submit_codegened_module_to_llvm,
     submit_post_lto_module_to_llvm, submit_pre_lto_module_to_llvm, ComputedLtoType, OngoingCodegen,
 };
-use crate::common::{IntPredicate, RealPredicate, TypeKind};
+use crate::common::{self, IntPredicate, RealPredicate, TypeKind};
 use crate::errors;
 use crate::meth;
 use crate::mir;
@@ -33,7 +33,7 @@ use rustc_middle::mir::mono::{CodegenUnit, CodegenUnitNameBuilder, MonoItem};
 use rustc_middle::query::Providers;
 use rustc_middle::ty::layout::{HasTyCtxt, LayoutOf, TyAndLayout};
 use rustc_middle::ty::{self, Instance, Ty, TyCtxt};
-use rustc_session::config::{self, CrateType, EntryFnType, OutputType};
+use rustc_session::config::{self, CrateType, EntryFnType, OptLevel, OutputType};
 use rustc_session::Session;
 use rustc_span::symbol::sym;
 use rustc_span::Symbol;
@@ -300,14 +300,35 @@ pub fn coerce_unsized_into<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
     }
 }
 
-pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
+/// Returns `rhs` sufficiently masked, truncated, and/or extended so that
+/// it can be used to shift `lhs`.
+///
+/// Shifts in MIR are all allowed to have mismatched LHS & RHS types.
+/// The shift methods in `BuilderMethods`, however, are fully homogeneous
+/// (both parameters and the return type are all the same type).
+///
+/// If `is_unchecked` is false, this masks the RHS to ensure it stays in-bounds,
+/// as the `BuilderMethods` shifts are UB for out-of-bounds shift amounts.
+/// For 32- and 64-bit types, this matches the semantics
+/// of Java. (See related discussion on #1877 and #10183.)
+///
+/// If `is_unchecked` is true, this does no masking, and adds sufficient `assume`
+/// calls or operation flags to preserve as much freedom to optimize as possible.
+pub fn build_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
     bx: &mut Bx,
     lhs: Bx::Value,
-    rhs: Bx::Value,
+    mut rhs: Bx::Value,
+    is_unchecked: bool,
 ) -> Bx::Value {
     // Shifts may have any size int on the rhs
     let mut rhs_llty = bx.cx().val_ty(rhs);
     let mut lhs_llty = bx.cx().val_ty(lhs);
+
+    let mask = common::shift_mask_val(bx, lhs_llty, rhs_llty, false);
+    if !is_unchecked {
+        rhs = bx.and(rhs, mask);
+    }
+
     if bx.cx().type_kind(rhs_llty) == TypeKind::Vector {
         rhs_llty = bx.cx().element_type(rhs_llty)
     }
@@ -317,6 +338,12 @@ pub fn cast_shift_expr_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
     let rhs_sz = bx.cx().int_width(rhs_llty);
     let lhs_sz = bx.cx().int_width(lhs_llty);
     if lhs_sz < rhs_sz {
+        if is_unchecked && bx.sess().opts.optimize != OptLevel::No {
+            // FIXME: Use `trunc nuw` once that's available
+            let inrange = bx.icmp(IntPredicate::IntULE, rhs, mask);
+            bx.assume(inrange);
+        }
+
         bx.trunc(rhs, lhs_llty)
     } else if lhs_sz > rhs_sz {
         // We zero-extend even if the RHS is signed. So e.g. `(x: i32) << -1i8` will zero-extend the
diff --git a/compiler/rustc_codegen_ssa/src/common.rs b/compiler/rustc_codegen_ssa/src/common.rs
index 71fca403def..b41739867c7 100644
--- a/compiler/rustc_codegen_ssa/src/common.rs
+++ b/compiler/rustc_codegen_ssa/src/common.rs
@@ -3,10 +3,9 @@
 use rustc_hir::LangItem;
 use rustc_middle::mir;
 use rustc_middle::ty::Instance;
-use rustc_middle::ty::{self, layout::TyAndLayout, Ty, TyCtxt};
+use rustc_middle::ty::{self, layout::TyAndLayout, TyCtxt};
 use rustc_span::Span;
 
-use crate::base;
 use crate::traits::*;
 
 #[derive(Copy, Clone)]
@@ -128,44 +127,6 @@ pub fn build_langcall<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
     (bx.fn_abi_of_instance(instance, ty::List::empty()), bx.get_fn_addr(instance), instance)
 }
 
-// To avoid UB from LLVM, these two functions mask RHS with an
-// appropriate mask unconditionally (i.e., the fallback behavior for
-// all shifts). For 32- and 64-bit types, this matches the semantics
-// of Java. (See related discussion on #1877 and #10183.)
-
-pub fn build_masked_lshift<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
-    bx: &mut Bx,
-    lhs: Bx::Value,
-    rhs: Bx::Value,
-) -> Bx::Value {
-    let rhs = base::cast_shift_expr_rhs(bx, lhs, rhs);
-    // #1877, #10183: Ensure that input is always valid
-    let rhs = shift_mask_rhs(bx, rhs);
-    bx.shl(lhs, rhs)
-}
-
-pub fn build_masked_rshift<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
-    bx: &mut Bx,
-    lhs_t: Ty<'tcx>,
-    lhs: Bx::Value,
-    rhs: Bx::Value,
-) -> Bx::Value {
-    let rhs = base::cast_shift_expr_rhs(bx, lhs, rhs);
-    // #1877, #10183: Ensure that input is always valid
-    let rhs = shift_mask_rhs(bx, rhs);
-    let is_signed = lhs_t.is_signed();
-    if is_signed { bx.ashr(lhs, rhs) } else { bx.lshr(lhs, rhs) }
-}
-
-fn shift_mask_rhs<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
-    bx: &mut Bx,
-    rhs: Bx::Value,
-) -> Bx::Value {
-    let rhs_llty = bx.val_ty(rhs);
-    let shift_val = shift_mask_val(bx, rhs_llty, rhs_llty, false);
-    bx.and(rhs, shift_val)
-}
-
 pub fn shift_mask_val<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
     bx: &mut Bx,
     llty: Bx::Type,
diff --git a/compiler/rustc_codegen_ssa/src/mir/rvalue.rs b/compiler/rustc_codegen_ssa/src/mir/rvalue.rs
index 0af84ff067a..6f7b98a262d 100644
--- a/compiler/rustc_codegen_ssa/src/mir/rvalue.rs
+++ b/compiler/rustc_codegen_ssa/src/mir/rvalue.rs
@@ -3,10 +3,11 @@ use super::place::PlaceRef;
 use super::{FunctionCx, LocalRef};
 
 use crate::base;
-use crate::common::{self, IntPredicate};
+use crate::common::IntPredicate;
 use crate::traits::*;
 use crate::MemFlags;
 
+use rustc_hir as hir;
 use rustc_middle::mir;
 use rustc_middle::mir::Operand;
 use rustc_middle::ty::cast::{CastTy, IntTy};
@@ -508,7 +509,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                     // Since int2ptr can have arbitrary integer types as input (so we have to do
                     // sign extension and all that), it is currently best handled in the same code
                     // path as the other integer-to-X casts.
-                    | mir::CastKind::PointerFromExposedAddress => {
+                    | mir::CastKind::PointerWithExposedProvenance => {
                         assert!(bx.cx().is_backend_immediate(cast));
                         let ll_t_out = bx.cx().immediate_backend_type(cast);
                         if operand.layout.abi.is_uninhabited() {
@@ -860,14 +861,12 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                     bx.inbounds_gep(llty, lhs, &[rhs])
                 }
             }
-            mir::BinOp::Shl => common::build_masked_lshift(bx, lhs, rhs),
-            mir::BinOp::ShlUnchecked => {
-                let rhs = base::cast_shift_expr_rhs(bx, lhs, rhs);
+            mir::BinOp::Shl | mir::BinOp::ShlUnchecked => {
+                let rhs = base::build_shift_expr_rhs(bx, lhs, rhs, op == mir::BinOp::ShlUnchecked);
                 bx.shl(lhs, rhs)
             }
-            mir::BinOp::Shr => common::build_masked_rshift(bx, input_ty, lhs, rhs),
-            mir::BinOp::ShrUnchecked => {
-                let rhs = base::cast_shift_expr_rhs(bx, lhs, rhs);
+            mir::BinOp::Shr | mir::BinOp::ShrUnchecked => {
+                let rhs = base::build_shift_expr_rhs(bx, lhs, rhs, op == mir::BinOp::ShrUnchecked);
                 if is_signed { bx.ashr(lhs, rhs) } else { bx.lshr(lhs, rhs) }
             }
             mir::BinOp::Ne
@@ -882,6 +881,35 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                     bx.icmp(base::bin_op_to_icmp_predicate(op.to_hir_binop(), is_signed), lhs, rhs)
                 }
             }
+            mir::BinOp::Cmp => {
+                use std::cmp::Ordering;
+                debug_assert!(!is_float);
+                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(hir::BinOpKind::Gt), lhs, rhs);
+                    let gtext = bx.zext(is_gt, bx.type_i8());
+                    let is_lt = bx.icmp(pred(hir::BinOpKind::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/integer-cmp.rs`,
+                    // from <https://github.com/rust-lang/rust/pull/63767>.
+                    let is_lt = bx.icmp(pred(hir::BinOpKind::Lt), lhs, rhs);
+                    let is_ne = bx.icmp(pred(hir::BinOpKind::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)
+                }
+            }
         }
     }
 
diff --git a/compiler/rustc_codegen_ssa/src/traits/consts.rs b/compiler/rustc_codegen_ssa/src/traits/consts.rs
index 4dff9c7684f..8cb17a5b37a 100644
--- a/compiler/rustc_codegen_ssa/src/traits/consts.rs
+++ b/compiler/rustc_codegen_ssa/src/traits/consts.rs
@@ -19,6 +19,7 @@ pub trait ConstMethods<'tcx>: BackendTypes {
     fn const_bool(&self, val: bool) -> Self::Value;
     fn const_i16(&self, i: i16) -> Self::Value;
     fn const_i32(&self, i: i32) -> Self::Value;
+    fn const_i8(&self, i: i8) -> Self::Value;
     fn const_u32(&self, i: u32) -> Self::Value;
     fn const_u64(&self, i: u64) -> Self::Value;
     fn const_u128(&self, i: u128) -> Self::Value;