rustc: support u128 discriminant ranges.

Author: eddyb

src/librustc/lib.rs

@@ -46,6 +46,7 @@
 #![feature(const_fn)]
 #![feature(core_intrinsics)]
 #![feature(drain_filter)]
+#![feature(i128)]
 #![feature(i128_type)]
 #![feature(inclusive_range)]
 #![feature(inclusive_range_syntax)]

src/librustc/ty/layout.rs

@@ -20,7 +20,7 @@ use syntax_pos::DUMMY_SP;
 
 use std::cmp;
 use std::fmt;
-use std::i64;
+use std::i128;
 use std::iter;
 use std::mem;
 use std::ops::{Add, Sub, Mul, AddAssign, Deref, RangeInclusive};
@@ -467,7 +467,7 @@ impl<'a, 'tcx> Integer {
     }
 
     /// Find the smallest Integer type which can represent the signed value.
-    pub fn fit_signed(x: i64) -> Integer {
+    pub fn fit_signed(x: i128) -> Integer {
         match x {
             -0x0000_0000_0000_0001...0x0000_0000_0000_0000 => I1,
             -0x0000_0000_0000_0080...0x0000_0000_0000_007f => I8,
@@ -479,7 +479,7 @@ impl<'a, 'tcx> Integer {
     }
 
     /// Find the smallest Integer type which can represent the unsigned value.
-    pub fn fit_unsigned(x: u64) -> Integer {
+    pub fn fit_unsigned(x: u128) -> Integer {
         match x {
             0...0x0000_0000_0000_0001 => I1,
             0...0x0000_0000_0000_00ff => I8,
@@ -495,7 +495,7 @@ impl<'a, 'tcx> Integer {
         let dl = cx.data_layout();
 
         let wanted = align.abi();
-        for &candidate in &[I8, I16, I32, I64] {
+        for &candidate in &[I8, I16, I32, I64, I128] {
             let ty = Int(candidate, false);
             if wanted == ty.align(dl).abi() && wanted == ty.size(dl).bytes() {
                 return Some(candidate);
@@ -522,19 +522,19 @@ impl<'a, 'tcx> Integer {
 
     /// Find the appropriate Integer type and signedness for the given
     /// signed discriminant range and #[repr] attribute.
-    /// N.B.: u64 values above i64::MAX will be treated as signed, but
+    /// N.B.: u128 values above i128::MAX will be treated as signed, but
     /// that shouldn't affect anything, other than maybe debuginfo.
     fn repr_discr(tcx: TyCtxt<'a, 'tcx, 'tcx>,
                   ty: Ty<'tcx>,
                   repr: &ReprOptions,
-                  min: i64,
-                  max: i64)
+                  min: i128,
+                  max: i128)
                   -> (Integer, bool) {
         // Theoretically, negative values could be larger in unsigned representation
         // than the unsigned representation of the signed minimum. However, if there
-        // are any negative values, the only valid unsigned representation is u64
-        // which can fit all i64 values, so the result remains unaffected.
-        let unsigned_fit = Integer::fit_unsigned(cmp::max(min as u64, max as u64));
+        // are any negative values, the only valid unsigned representation is u128
+        // which can fit all i128 values, so the result remains unaffected.
+        let unsigned_fit = Integer::fit_unsigned(cmp::max(min as u128, max as u128));
         let signed_fit = cmp::max(Integer::fit_signed(min), Integer::fit_signed(max));
 
         let mut min_from_extern = None;
@@ -782,11 +782,11 @@ pub enum Variants {
     Tagged {
         discr: Primitive,
         /// Inclusive wrap-around range of discriminant values, that is,
-        /// if min > max, it represents min..=u64::MAX followed by 0..=max.
+        /// if min > max, it represents min..=u128::MAX followed by 0..=max.
         // FIXME(eddyb) always use the shortest range, e.g. by finding
         // the largest space between two consecutive discriminants and
         // taking everything else as the (shortest) discriminant range.
-        discr_range: RangeInclusive<u64>,
+        discr_range: RangeInclusive<u128>,
         variants: Vec<CachedLayout>,
     },
 
@@ -1375,14 +1375,12 @@ impl<'a, 'tcx> CachedLayout {
                     }
                 }
 
-                let (mut min, mut max) = (i64::max_value(), i64::min_value());
+                let (mut min, mut max) = (i128::max_value(), i128::min_value());
                 for discr in def.discriminants(tcx) {
-                    let x = discr.to_u128_unchecked() as i64;
+                    let x = discr.to_u128_unchecked() as i128;
                     if x < min { min = x; }
                     if x > max { max = x; }
                 }
-                // FIXME: should handle i128? signed-value based impl is weird and hard to
-                // grok.
                 let (min_ity, signed) = Integer::repr_discr(tcx, ty, &def.repr, min, max);
 
                 let mut align = dl.aggregate_align;
@@ -1479,9 +1477,7 @@ impl<'a, 'tcx> CachedLayout {
                 tcx.intern_layout(CachedLayout {
                     variants: Variants::Tagged {
                         discr,
-
-                        // FIXME: should be u128?
-                        discr_range: (min as u64)..=(max as u64),
+                        discr_range: (min as u128)..=(max as u128),
                         variants
                     },
                     // FIXME(eddyb): using `FieldPlacement::Arbitrary` here results

src/librustc_trans/builder.rs

@@ -24,6 +24,7 @@ use rustc::session::{config, Session};
 
 use std::borrow::Cow;
 use std::ffi::CString;
+use std::ops::Range;
 use std::ptr;
 use syntax_pos::Span;
 
@@ -549,35 +550,26 @@ impl<'a, 'tcx> Builder<'a, 'tcx> {
     }
 
 
-    pub fn load_range_assert(&self, ptr: ValueRef, lo: u64,
-                             hi: u64, signed: llvm::Bool,
-                             align: Option<Align>) -> ValueRef {
-        let value = self.load(ptr, align);
-
+    pub fn range_metadata(&self, load: ValueRef, range: Range<u128>) {
         unsafe {
-            let t = llvm::LLVMGetElementType(llvm::LLVMTypeOf(ptr));
-            let min = llvm::LLVMConstInt(t, lo, signed);
-            let max = llvm::LLVMConstInt(t, hi, signed);
-
-            let v = [min, max];
+            let llty = val_ty(load);
+            let v = [
+                C_uint_big(llty, range.start),
+                C_uint_big(llty, range.end)
+            ];
 
-            llvm::LLVMSetMetadata(value, llvm::MD_range as c_uint,
+            llvm::LLVMSetMetadata(load, llvm::MD_range as c_uint,
                                   llvm::LLVMMDNodeInContext(self.ccx.llcx(),
                                                             v.as_ptr(),
                                                             v.len() as c_uint));
         }
-
-        value
     }
 
-    pub fn load_nonnull(&self, ptr: ValueRef, align: Option<Align>) -> ValueRef {
-        let value = self.load(ptr, align);
+    pub fn nonnull_metadata(&self, load: ValueRef) {
         unsafe {
-            llvm::LLVMSetMetadata(value, llvm::MD_nonnull as c_uint,
+            llvm::LLVMSetMetadata(load, llvm::MD_nonnull as c_uint,
                                   llvm::LLVMMDNodeInContext(self.ccx.llcx(), ptr::null(), 0));
         }
-
-        value
     }
 
     pub fn store(&self, val: ValueRef, ptr: ValueRef, align: Option<Align>) -> ValueRef {

src/librustc_trans/common.rs

@@ -178,9 +178,9 @@ pub fn C_uint(t: Type, i: u64) -> ValueRef {
     }
 }
 
-pub fn C_big_integral(t: Type, u: u128) -> ValueRef {
+pub fn C_uint_big(t: Type, u: u128) -> ValueRef {
     unsafe {
-        let words = [u as u64, u.wrapping_shr(64) as u64];
+        let words = [u as u64, (u >> 64) as u64];
         llvm::LLVMConstIntOfArbitraryPrecision(t.to_ref(), 2, words.as_ptr())
     }
 }

src/librustc_trans/meth.rs

@@ -40,7 +40,8 @@ impl<'a, 'tcx> VirtualIndex {
         debug!("get_fn({:?}, {:?})", Value(llvtable), self);
 
         let llvtable = bcx.pointercast(llvtable, fn_ty.llvm_type(bcx.ccx).ptr_to().ptr_to());
-        let ptr = bcx.load_nonnull(bcx.inbounds_gep(llvtable, &[C_usize(bcx.ccx, self.0)]), None);
+        let ptr = bcx.load(bcx.inbounds_gep(llvtable, &[C_usize(bcx.ccx, self.0)]), None);
+        bcx.nonnull_metadata(ptr);
         // Vtable loads are invariant
         bcx.set_invariant_load(ptr);
         ptr

src/librustc_trans/mir/block.rs

@@ -666,17 +666,18 @@ impl<'a, 'tcx> MirContext<'a, 'tcx> {
 
         if by_ref && !arg.is_indirect() {
             // Have to load the argument, maybe while casting it.
-            if arg.layout.ty == bcx.tcx().types.bool {
-                llval = bcx.load_range_assert(llval, 0, 2, llvm::False, None);
-                // We store bools as i8 so we need to truncate to i1.
-                llval = base::to_immediate(bcx, llval, arg.layout);
-            } else if let Some(ty) = arg.cast {
+            if let Some(ty) = arg.cast {
                 llval = bcx.load(bcx.pointercast(llval, ty.llvm_type(bcx.ccx).ptr_to()),
                                  (align | Alignment::Packed(arg.layout.align))
                                     .non_abi());
             } else {
                 llval = bcx.load(llval, align.non_abi());
             }
+            if arg.layout.ty == bcx.tcx().types.bool {
+                bcx.range_metadata(llval, 0..2);
+                // We store bools as i8 so we need to truncate to i1.
+                llval = base::to_immediate(bcx, llval, arg.layout);
+            }
         }
 
         llargs.push(llval);

src/librustc_trans/mir/constant.rs

@@ -28,7 +28,7 @@ use abi::{self, Abi};
 use callee;
 use builder::Builder;
 use common::{self, CrateContext, const_get_elt, val_ty};
-use common::{C_array, C_bool, C_bytes, C_int, C_uint, C_big_integral, C_u32, C_u64};
+use common::{C_array, C_bool, C_bytes, C_int, C_uint, C_uint_big, C_u32, C_u64};
 use common::{C_null, C_struct, C_str_slice, C_undef, C_usize, C_vector, C_fat_ptr};
 use common::const_to_opt_u128;
 use consts;
@@ -70,13 +70,13 @@ impl<'a, 'tcx> Const<'tcx> {
             I16(v) => (C_int(Type::i16(ccx), v as i64), tcx.types.i16),
             I32(v) => (C_int(Type::i32(ccx), v as i64), tcx.types.i32),
             I64(v) => (C_int(Type::i64(ccx), v as i64), tcx.types.i64),
-            I128(v) => (C_big_integral(Type::i128(ccx), v as u128), tcx.types.i128),
+            I128(v) => (C_uint_big(Type::i128(ccx), v as u128), tcx.types.i128),
             Isize(v) => (C_int(Type::isize(ccx), v.as_i64()), tcx.types.isize),
             U8(v) => (C_uint(Type::i8(ccx), v as u64), tcx.types.u8),
             U16(v) => (C_uint(Type::i16(ccx), v as u64), tcx.types.u16),
             U32(v) => (C_uint(Type::i32(ccx), v as u64), tcx.types.u32),
             U64(v) => (C_uint(Type::i64(ccx), v), tcx.types.u64),
-            U128(v) => (C_big_integral(Type::i128(ccx), v), tcx.types.u128),
+            U128(v) => (C_uint_big(Type::i128(ccx), v), tcx.types.u128),
             Usize(v) => (C_uint(Type::isize(ccx), v.as_u64()), tcx.types.usize),
         };
         Const { llval: llval, ty: ty }
@@ -994,7 +994,7 @@ unsafe fn cast_const_float_to_int(ccx: &CrateContext,
         let err = ConstEvalErr { span: span, kind: ErrKind::CannotCast };
         err.report(ccx.tcx(), span, "expression");
     }
-    C_big_integral(int_ty, cast_result.value)
+    C_uint_big(int_ty, cast_result.value)
 }
 
 unsafe fn cast_const_int_to_float(ccx: &CrateContext,

src/librustc_trans/mir/lvalue.rs

@@ -14,10 +14,9 @@ use rustc::ty::layout::{self, Align, TyLayout, LayoutOf};
 use rustc::mir;
 use rustc::mir::tcx::LvalueTy;
 use rustc_data_structures::indexed_vec::Idx;
-use abi;
 use base;
 use builder::Builder;
-use common::{self, CrateContext, C_usize, C_u8, C_u32, C_uint, C_int, C_null, val_ty};
+use common::{self, CrateContext, C_usize, C_u8, C_u32, C_uint, C_int, C_null};
 use consts;
 use type_of::LayoutLlvmExt;
 use type_::Type;
@@ -140,30 +139,7 @@ impl<'a, 'tcx> LvalueRef<'tcx> {
             return OperandRef::new_zst(bcx.ccx, self.layout);
         }
 
-        let val = if common::type_is_fat_ptr(bcx.ccx, self.layout.ty) {
-            let data = self.project_field(bcx, abi::FAT_PTR_ADDR);
-            let lldata = if self.layout.ty.is_region_ptr() || self.layout.ty.is_box() {
-                bcx.load_nonnull(data.llval, data.alignment.non_abi())
-            } else {
-                bcx.load(data.llval, data.alignment.non_abi())
-            };
-
-            let extra = self.project_field(bcx, abi::FAT_PTR_EXTRA);
-            let meta_ty = val_ty(extra.llval);
-            // If the 'extra' field is a pointer, it's a vtable, so use load_nonnull
-            // instead
-            let llextra = if meta_ty.element_type().kind() == llvm::TypeKind::Pointer {
-                bcx.load_nonnull(extra.llval, extra.alignment.non_abi())
-            } else {
-                bcx.load(extra.llval, extra.alignment.non_abi())
-            };
-
-            OperandValue::Pair(lldata, llextra)
-        } else if common::type_is_imm_pair(bcx.ccx, self.layout.ty) {
-            OperandValue::Pair(
-                self.project_field(bcx, 0).load(bcx).pack_if_pair(bcx).immediate(),
-                self.project_field(bcx, 1).load(bcx).pack_if_pair(bcx).immediate())
-        } else if self.layout.is_llvm_immediate() {
+        let val = if self.layout.is_llvm_immediate() {
             let mut const_llval = ptr::null_mut();
             unsafe {
                 let global = llvm::LLVMIsAGlobalVariable(self.llval);
@@ -174,22 +150,26 @@ impl<'a, 'tcx> LvalueRef<'tcx> {
 
             let llval = if !const_llval.is_null() {
                 const_llval
-            } else if self.layout.ty.is_bool() {
-                bcx.load_range_assert(self.llval, 0, 2, llvm::False,
-                    self.alignment.non_abi())
-            } else if self.layout.ty.is_char() {
-                // a char is a Unicode codepoint, and so takes values from 0
-                // to 0x10FFFF inclusive only.
-                bcx.load_range_assert(self.llval, 0, 0x10FFFF + 1, llvm::False,
-                    self.alignment.non_abi())
-            } else if self.layout.ty.is_region_ptr() ||
-                      self.layout.ty.is_box() ||
-                      self.layout.ty.is_fn() {
-                bcx.load_nonnull(self.llval, self.alignment.non_abi())
             } else {
-                bcx.load(self.llval, self.alignment.non_abi())
+                let load = bcx.load(self.llval, self.alignment.non_abi());
+                if self.layout.ty.is_bool() {
+                    bcx.range_metadata(load, 0..2);
+                } else if self.layout.ty.is_char() {
+                    // a char is a Unicode codepoint, and so takes values from 0
+                    // to 0x10FFFF inclusive only.
+                    bcx.range_metadata(load, 0..0x10FFFF+1);
+                } else if self.layout.ty.is_region_ptr() ||
+                        self.layout.ty.is_box() ||
+                        self.layout.ty.is_fn() {
+                    bcx.nonnull_metadata(load);
+                }
+                load
             };
             OperandValue::Immediate(base::to_immediate(bcx, llval, self.layout))
+        } else if common::type_is_imm_pair(bcx.ccx, self.layout.ty) {
+            OperandValue::Pair(
+                self.project_field(bcx, 0).load(bcx).pack_if_pair(bcx).immediate(),
+                self.project_field(bcx, 1).load(bcx).pack_if_pair(bcx).immediate())
         } else {
             OperandValue::Ref(self.llval, self.alignment)
         };
@@ -314,28 +294,26 @@ impl<'a, 'tcx> LvalueRef<'tcx> {
             layout::Variants::Tagged { ref discr_range, .. } => {
                 (discr_range.start, discr_range.end)
             }
-            _ => (0, u64::max_value()),
+            _ => (0, !0),
         };
         let max_next = max.wrapping_add(1);
         let bits = discr_scalar.size(bcx.ccx).bits();
-        assert!(bits <= 64);
-        let mask = !0u64 >> (64 - bits);
-        let lldiscr = match discr_scalar {
+        assert!(bits <= 128);
+        let mask = !0u128 >> (128 - bits);
+        let lldiscr = bcx.load(discr.llval, discr.alignment.non_abi());
+        match discr_scalar {
             // For a (max) discr of -1, max will be `-1 as usize`, which overflows.
             // However, that is fine here (it would still represent the full range),
             layout::Int(..) if max_next & mask != min & mask => {
                 // llvm::ConstantRange can deal with ranges that wrap around,
                 // so an overflow on (max + 1) is fine.
-                bcx.load_range_assert(discr.llval, min, max_next,
-                                      /* signed: */ llvm::True,
-                                      discr.alignment.non_abi())
+                bcx.range_metadata(lldiscr, min..max_next);
             }
             _ => {
                 // i.e., if the range is everything.  The lo==hi case would be
                 // rejected by the LLVM verifier (it would mean either an
                 // empty set, which is impossible, or the entire range of the
                 // type, which is pointless).
-                bcx.load(discr.llval, discr.alignment.non_abi())
             }
         };
         match self.layout.variants {