Auto merge of #87695 - tmiasko:gep-opaque-pointers, r=nagisa

bors · bors · commit d54fbb98555c · 2021-08-04T18:22:29.000Z
Prepare GEP building for opaque pointers
diff --git a/compiler/rustc_codegen_llvm/src/builder.rs b/compiler/rustc_codegen_llvm/src/builder.rs
@@ -497,9 +497,10 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
             OperandValue::Immediate(self.to_immediate(llval, place.layout))
         } else if let abi::Abi::ScalarPair(ref a, ref b) = place.layout.abi {
             let b_offset = a.value.size(self).align_to(b.value.align(self).abi);
+            let pair_ty = place.layout.llvm_type(self);
 
             let mut load = |i, scalar: &abi::Scalar, align| {
-                let llptr = self.struct_gep(place.llval, i as u64);
+                let llptr = self.struct_gep(pair_ty, place.llval, i as u64);
                 let llty = place.layout.scalar_pair_element_llvm_type(self, i, false);
                 let load = self.load(llty, llptr, align);
                 scalar_load_metadata(self, load, scalar);
@@ -543,7 +544,11 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
             .val
             .store(&mut body_bx, PlaceRef::new_sized_aligned(current, cg_elem.layout, align));
 
-        let next = body_bx.inbounds_gep(current, &[self.const_usize(1)]);
+        let next = body_bx.inbounds_gep(
+            self.backend_type(cg_elem.layout),
+            current,
+            &[self.const_usize(1)],
+        );
         body_bx.br(header_bx.llbb());
         header_bx.add_incoming_to_phi(current, next, body_bx.llbb());
 
@@ -639,10 +644,11 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         }
     }
 
-    fn gep(&mut self, ptr: &'ll Value, indices: &[&'ll Value]) -> &'ll Value {
+    fn gep(&mut self, ty: &'ll Type, ptr: &'ll Value, indices: &[&'ll Value]) -> &'ll Value {
         unsafe {
-            llvm::LLVMBuildGEP(
+            llvm::LLVMBuildGEP2(
                 self.llbuilder,
+                ty,
                 ptr,
                 indices.as_ptr(),
                 indices.len() as c_uint,
@@ -651,10 +657,16 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         }
     }
 
-    fn inbounds_gep(&mut self, ptr: &'ll Value, indices: &[&'ll Value]) -> &'ll Value {
+    fn inbounds_gep(
+        &mut self,
+        ty: &'ll Type,
+        ptr: &'ll Value,
+        indices: &[&'ll Value],
+    ) -> &'ll Value {
         unsafe {
-            llvm::LLVMBuildInBoundsGEP(
+            llvm::LLVMBuildInBoundsGEP2(
                 self.llbuilder,
+                ty,
                 ptr,
                 indices.as_ptr(),
                 indices.len() as c_uint,
@@ -663,9 +675,9 @@ impl BuilderMethods<'a, 'tcx> for Builder<'a, 'll, 'tcx> {
         }
     }
 
-    fn struct_gep(&mut self, ptr: &'ll Value, idx: u64) -> &'ll Value {
+    fn struct_gep(&mut self, ty: &'ll Type, ptr: &'ll Value, idx: u64) -> &'ll Value {
         assert_eq!(idx as c_uint as u64, idx);
-        unsafe { llvm::LLVMBuildStructGEP(self.llbuilder, ptr, idx as c_uint, UNNAMED) }
+        unsafe { llvm::LLVMBuildStructGEP2(self.llbuilder, ty, ptr, idx as c_uint, UNNAMED) }
     }
 
     /* Casts */
diff --git a/compiler/rustc_codegen_llvm/src/common.rs b/compiler/rustc_codegen_llvm/src/common.rs
@@ -268,7 +268,8 @@ impl ConstMethods<'tcx> for CodegenCx<'ll, 'tcx> {
                     }
                 };
                 let llval = unsafe {
-                    llvm::LLVMConstInBoundsGEP(
+                    llvm::LLVMRustConstInBoundsGEP2(
+                        self.type_i8(),
                         self.const_bitcast(base_addr, self.type_i8p_ext(base_addr_space)),
                         &self.const_usize(offset.bytes()),
                         1,
@@ -303,7 +304,8 @@ impl ConstMethods<'tcx> for CodegenCx<'ll, 'tcx> {
             let base_addr = self.static_addr_of(init, alloc.align, None);
 
             let llval = unsafe {
-                llvm::LLVMConstInBoundsGEP(
+                llvm::LLVMRustConstInBoundsGEP2(
+                    self.type_i8(),
                     self.const_bitcast(base_addr, self.type_i8p()),
                     &self.const_usize(offset.bytes()),
                     1,
diff --git a/compiler/rustc_codegen_llvm/src/debuginfo/gdb.rs b/compiler/rustc_codegen_llvm/src/debuginfo/gdb.rs
@@ -15,12 +15,11 @@ use rustc_span::symbol::sym;
 /// .debug_gdb_scripts global is referenced, so it isn't removed by the linker.
 pub fn insert_reference_to_gdb_debug_scripts_section_global(bx: &mut Builder<'_, '_, '_>) {
     if needs_gdb_debug_scripts_section(bx) {
-        let gdb_debug_scripts_section = get_or_insert_gdb_debug_scripts_section_global(bx);
+        let gdb_debug_scripts_section =
+            bx.const_bitcast(get_or_insert_gdb_debug_scripts_section_global(bx), bx.type_i8p());
         // Load just the first byte as that's all that's necessary to force
         // LLVM to keep around the reference to the global.
-        let indices = [bx.const_i32(0), bx.const_i32(0)];
-        let element = bx.inbounds_gep(gdb_debug_scripts_section, &indices);
-        let volative_load_instruction = bx.volatile_load(bx.type_i8(), element);
+        let volative_load_instruction = bx.volatile_load(bx.type_i8(), gdb_debug_scripts_section);
         unsafe {
             llvm::LLVMSetAlignment(volative_load_instruction, 1);
         }
diff --git a/compiler/rustc_codegen_llvm/src/intrinsic.rs b/compiler/rustc_codegen_llvm/src/intrinsic.rs
@@ -686,11 +686,19 @@ fn codegen_emcc_try(
         // create an alloca and pass a pointer to that.
         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
         let i8_align = bx.tcx().data_layout.i8_align.abi;
-        let catch_data =
-            catch.alloca(bx.type_struct(&[bx.type_i8p(), bx.type_bool()], false), ptr_align);
-        let catch_data_0 = catch.inbounds_gep(catch_data, &[bx.const_usize(0), bx.const_usize(0)]);
+        let catch_data_type = bx.type_struct(&[bx.type_i8p(), bx.type_bool()], false);
+        let catch_data = catch.alloca(catch_data_type, ptr_align);
+        let catch_data_0 = catch.inbounds_gep(
+            catch_data_type,
+            catch_data,
+            &[bx.const_usize(0), bx.const_usize(0)],
+        );
         catch.store(ptr, catch_data_0, ptr_align);
-        let catch_data_1 = catch.inbounds_gep(catch_data, &[bx.const_usize(0), bx.const_usize(1)]);
+        let catch_data_1 = catch.inbounds_gep(
+            catch_data_type,
+            catch_data,
+            &[bx.const_usize(0), bx.const_usize(1)],
+        );
         catch.store(is_rust_panic, catch_data_1, i8_align);
         let catch_data = catch.bitcast(catch_data, bx.type_i8p());
 
diff --git a/compiler/rustc_codegen_llvm/src/llvm/ffi.rs b/compiler/rustc_codegen_llvm/src/llvm/ffi.rs
@@ -1011,7 +1011,8 @@ extern "C" {
     pub fn LLVMConstVector(ScalarConstantVals: *const &Value, Size: c_uint) -> &Value;
 
     // Constant expressions
-    pub fn LLVMConstInBoundsGEP(
+    pub fn LLVMRustConstInBoundsGEP2(
+        ty: &'a Type,
         ConstantVal: &'a Value,
         ConstantIndices: *const &'a Value,
         NumIndices: c_uint,
@@ -1394,22 +1395,25 @@ extern "C" {
 
     pub fn LLVMBuildStore(B: &Builder<'a>, Val: &'a Value, Ptr: &'a Value) -> &'a Value;
 
-    pub fn LLVMBuildGEP(
+    pub fn LLVMBuildGEP2(
         B: &Builder<'a>,
+        Ty: &'a Type,
         Pointer: &'a Value,
         Indices: *const &'a Value,
         NumIndices: c_uint,
         Name: *const c_char,
     ) -> &'a Value;
-    pub fn LLVMBuildInBoundsGEP(
+    pub fn LLVMBuildInBoundsGEP2(
         B: &Builder<'a>,
+        Ty: &'a Type,
         Pointer: &'a Value,
         Indices: *const &'a Value,
         NumIndices: c_uint,
         Name: *const c_char,
     ) -> &'a Value;
-    pub fn LLVMBuildStructGEP(
+    pub fn LLVMBuildStructGEP2(
         B: &Builder<'a>,
+        Ty: &'a Type,
         Pointer: &'a Value,
         Idx: c_uint,
         Name: *const c_char,
diff --git a/compiler/rustc_codegen_llvm/src/va_arg.rs b/compiler/rustc_codegen_llvm/src/va_arg.rs
@@ -50,12 +50,12 @@ fn emit_direct_ptr_va_arg(
 
     let aligned_size = size.align_to(slot_size).bytes() as i32;
     let full_direct_size = bx.cx().const_i32(aligned_size);
-    let next = bx.inbounds_gep(addr, &[full_direct_size]);
+    let next = bx.inbounds_gep(bx.type_i8(), addr, &[full_direct_size]);
     bx.store(next, va_list_addr, bx.tcx().data_layout.pointer_align.abi);
 
     if size.bytes() < slot_size.bytes() && bx.tcx().sess.target.endian == Endian::Big {
         let adjusted_size = bx.cx().const_i32((slot_size.bytes() - size.bytes()) as i32);
-        let adjusted = bx.inbounds_gep(addr, &[adjusted_size]);
+        let adjusted = bx.inbounds_gep(bx.type_i8(), addr, &[adjusted_size]);
         (bx.bitcast(adjusted, bx.cx().type_ptr_to(llty)), addr_align)
     } else {
         (bx.bitcast(addr, bx.cx().type_ptr_to(llty)), addr_align)
@@ -98,6 +98,7 @@ fn emit_aapcs_va_arg(
     // Implementation of the AAPCS64 calling convention for va_args see
     // https://github.com/ARM-software/abi-aa/blob/master/aapcs64/aapcs64.rst
     let va_list_addr = list.immediate();
+    let va_list_ty = list.deref(bx.cx).layout.llvm_type(bx);
     let layout = bx.cx.layout_of(target_ty);
 
     let mut maybe_reg = bx.build_sibling_block("va_arg.maybe_reg");
@@ -109,11 +110,11 @@ fn emit_aapcs_va_arg(
 
     let gr_type = target_ty.is_any_ptr() || target_ty.is_integral();
     let (reg_off, reg_top_index, slot_size) = if gr_type {
-        let gr_offs = bx.struct_gep(va_list_addr, 7);
+        let gr_offs = bx.struct_gep(va_list_ty, va_list_addr, 7);
         let nreg = (layout.size.bytes() + 7) / 8;
         (gr_offs, 3, nreg * 8)
     } else {
-        let vr_off = bx.struct_gep(va_list_addr, 9);
+        let vr_off = bx.struct_gep(va_list_ty, va_list_addr, 9);
         let nreg = (layout.size.bytes() + 15) / 16;
         (vr_off, 5, nreg * 16)
     };
@@ -141,15 +142,15 @@ fn emit_aapcs_va_arg(
     maybe_reg.cond_br(use_stack, &on_stack.llbb(), &in_reg.llbb());
 
     let top_type = bx.type_i8p();
-    let top = in_reg.struct_gep(va_list_addr, reg_top_index);
+    let top = in_reg.struct_gep(va_list_ty, va_list_addr, reg_top_index);
     let top = in_reg.load(top_type, top, bx.tcx().data_layout.pointer_align.abi);
 
     // reg_value = *(@top + reg_off_v);
-    let mut reg_addr = in_reg.gep(top, &[reg_off_v]);
+    let mut reg_addr = in_reg.gep(bx.type_i8(), top, &[reg_off_v]);
     if bx.tcx().sess.target.endian == Endian::Big && layout.size.bytes() != slot_size {
         // On big-endian systems the value is right-aligned in its slot.
         let offset = bx.const_i32((slot_size - layout.size.bytes()) as i32);
-        reg_addr = in_reg.gep(reg_addr, &[offset]);
+        reg_addr = in_reg.gep(bx.type_i8(), reg_addr, &[offset]);
     }
     let reg_type = layout.llvm_type(bx);
     let reg_addr = in_reg.bitcast(reg_addr, bx.cx.type_ptr_to(reg_type));
diff --git a/compiler/rustc_codegen_ssa/src/base.rs b/compiler/rustc_codegen_ssa/src/base.rs
@@ -168,8 +168,11 @@ pub fn unsized_info<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>>(
                 let ptr_ty = cx.type_i8p();
                 let ptr_align = cx.tcx().data_layout.pointer_align.abi;
                 let llvtable = bx.pointercast(old_info, bx.type_ptr_to(ptr_ty));
-                let gep =
-                    bx.inbounds_gep(llvtable, &[bx.const_usize(u64::try_from(entry_idx).unwrap())]);
+                let gep = bx.inbounds_gep(
+                    ptr_ty,
+                    llvtable,
+                    &[bx.const_usize(u64::try_from(entry_idx).unwrap())],
+                );
                 let new_vptr = bx.load(ptr_ty, gep, ptr_align);
                 bx.nonnull_metadata(new_vptr);
                 // Vtable loads are invariant.
diff --git a/compiler/rustc_codegen_ssa/src/meth.rs b/compiler/rustc_codegen_ssa/src/meth.rs
@@ -23,7 +23,7 @@ impl<'a, 'tcx> VirtualIndex {
         let llty = bx.fn_ptr_backend_type(fn_abi);
         let llvtable = bx.pointercast(llvtable, bx.type_ptr_to(llty));
         let ptr_align = bx.tcx().data_layout.pointer_align.abi;
-        let gep = bx.inbounds_gep(llvtable, &[bx.const_usize(self.0)]);
+        let gep = bx.inbounds_gep(llty, llvtable, &[bx.const_usize(self.0)]);
         let ptr = bx.load(llty, gep, ptr_align);
         bx.nonnull_metadata(ptr);
         // Vtable loads are invariant.
@@ -42,7 +42,7 @@ impl<'a, 'tcx> VirtualIndex {
         let llty = bx.type_isize();
         let llvtable = bx.pointercast(llvtable, bx.type_ptr_to(llty));
         let usize_align = bx.tcx().data_layout.pointer_align.abi;
-        let gep = bx.inbounds_gep(llvtable, &[bx.const_usize(self.0)]);
+        let gep = bx.inbounds_gep(llty, llvtable, &[bx.const_usize(self.0)]);
         let ptr = bx.load(llty, gep, usize_align);
         // Vtable loads are invariant.
         bx.set_invariant_load(ptr);
diff --git a/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs b/compiler/rustc_codegen_ssa/src/mir/intrinsic.rs
@@ -116,14 +116,18 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 OperandRef::from_const(bx, value, ret_ty).immediate_or_packed_pair(bx)
             }
             sym::offset => {
+                let ty = substs.type_at(0);
+                let layout = bx.layout_of(ty);
                 let ptr = args[0].immediate();
                 let offset = args[1].immediate();
-                bx.inbounds_gep(ptr, &[offset])
+                bx.inbounds_gep(bx.backend_type(layout), ptr, &[offset])
             }
             sym::arith_offset => {
+                let ty = substs.type_at(0);
+                let layout = bx.layout_of(ty);
                 let ptr = args[0].immediate();
                 let offset = args[1].immediate();
-                bx.gep(ptr, &[offset])
+                bx.gep(bx.backend_type(layout), ptr, &[offset])
             }
             sym::copy => {
                 copy_intrinsic(
diff --git a/compiler/rustc_codegen_ssa/src/mir/operand.rs b/compiler/rustc_codegen_ssa/src/mir/operand.rs
@@ -311,14 +311,15 @@ impl<'a, 'tcx, V: CodegenObject> OperandValue<V> {
                     Abi::ScalarPair(ref a, ref b) => (a, b),
                     _ => bug!("store_with_flags: invalid ScalarPair layout: {:#?}", dest.layout),
                 };
+                let ty = bx.backend_type(dest.layout);
                 let b_offset = a_scalar.value.size(bx).align_to(b_scalar.value.align(bx).abi);
 
-                let llptr = bx.struct_gep(dest.llval, 0);
+                let llptr = bx.struct_gep(ty, dest.llval, 0);
                 let val = bx.from_immediate(a);
                 let align = dest.align;
                 bx.store_with_flags(val, llptr, align, flags);
 
-                let llptr = bx.struct_gep(dest.llval, 1);
+                let llptr = bx.struct_gep(ty, dest.llval, 1);
                 let val = bx.from_immediate(b);
                 let align = dest.align.restrict_for_offset(b_offset);
                 bx.store_with_flags(val, llptr, align, flags);
diff --git a/compiler/rustc_codegen_ssa/src/mir/place.rs b/compiler/rustc_codegen_ssa/src/mir/place.rs
@@ -103,12 +103,13 @@ impl<'a, 'tcx, V: CodegenObject> PlaceRef<'tcx, V> {
                     if offset == a.value.size(bx.cx()).align_to(b.value.align(bx.cx()).abi) =>
                 {
                     // Offset matches second field.
-                    bx.struct_gep(self.llval, 1)
+                    let ty = bx.backend_type(self.layout);
+                    bx.struct_gep(ty, self.llval, 1)
                 }
                 Abi::Scalar(_) | Abi::ScalarPair(..) | Abi::Vector { .. } if field.is_zst() => {
                     // ZST fields are not included in Scalar, ScalarPair, and Vector layouts, so manually offset the pointer.
                     let byte_ptr = bx.pointercast(self.llval, bx.cx().type_i8p());
-                    bx.gep(byte_ptr, &[bx.const_usize(offset.bytes())])
+                    bx.gep(bx.cx().type_i8(), byte_ptr, &[bx.const_usize(offset.bytes())])
                 }
                 Abi::Scalar(_) | Abi::ScalarPair(..) => {
                     // All fields of Scalar and ScalarPair layouts must have been handled by this point.
@@ -119,7 +120,10 @@ impl<'a, 'tcx, V: CodegenObject> PlaceRef<'tcx, V> {
                         self.layout
                     );
                 }
-                _ => bx.struct_gep(self.llval, bx.cx().backend_field_index(self.layout, ix)),
+                _ => {
+                    let ty = bx.backend_type(self.layout);
+                    bx.struct_gep(ty, self.llval, bx.cx().backend_field_index(self.layout, ix))
+                }
             };
             PlaceRef {
                 // HACK(eddyb): have to bitcast pointers until LLVM removes pointee types.
@@ -185,7 +189,7 @@ impl<'a, 'tcx, V: CodegenObject> PlaceRef<'tcx, V> {
 
         // Cast and adjust pointer.
         let byte_ptr = bx.pointercast(self.llval, bx.cx().type_i8p());
-        let byte_ptr = bx.gep(byte_ptr, &[offset]);
+        let byte_ptr = bx.gep(bx.cx().type_i8(), byte_ptr, &[offset]);
 
         // Finally, cast back to the type expected.
         let ll_fty = bx.cx().backend_type(field);
@@ -380,7 +384,11 @@ impl<'a, 'tcx, V: CodegenObject> PlaceRef<'tcx, V> {
         };
 
         PlaceRef {
-            llval: bx.inbounds_gep(self.llval, &[bx.cx().const_usize(0), llindex]),
+            llval: bx.inbounds_gep(
+                bx.cx().backend_type(self.layout),
+                self.llval,
+                &[bx.cx().const_usize(0), llindex],
+            ),
             llextra: None,
             layout,
             align: self.align.restrict_for_offset(offset),
diff --git a/compiler/rustc_codegen_ssa/src/mir/rvalue.rs b/compiler/rustc_codegen_ssa/src/mir/rvalue.rs
@@ -636,7 +636,14 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             mir::BinOp::BitOr => bx.or(lhs, rhs),
             mir::BinOp::BitAnd => bx.and(lhs, rhs),
             mir::BinOp::BitXor => bx.xor(lhs, rhs),
-            mir::BinOp::Offset => bx.inbounds_gep(lhs, &[rhs]),
+            mir::BinOp::Offset => {
+                let pointee_type = input_ty
+                    .builtin_deref(true)
+                    .unwrap_or_else(|| bug!("deref of non-pointer {:?}", input_ty))
+                    .ty;
+                let llty = bx.cx().backend_type(bx.cx().layout_of(pointee_type));
+                bx.inbounds_gep(llty, lhs, &[rhs])
+            }
             mir::BinOp::Shl => common::build_unchecked_lshift(bx, lhs, rhs),
             mir::BinOp::Shr => common::build_unchecked_rshift(bx, input_ty, lhs, rhs),
             mir::BinOp::Ne
diff --git a/compiler/rustc_codegen_ssa/src/traits/builder.rs b/compiler/rustc_codegen_ssa/src/traits/builder.rs
@@ -176,9 +176,14 @@ pub trait BuilderMethods<'a, 'tcx>:
         size: Size,
     );
 
-    fn gep(&mut self, ptr: Self::Value, indices: &[Self::Value]) -> Self::Value;
-    fn inbounds_gep(&mut self, ptr: Self::Value, indices: &[Self::Value]) -> Self::Value;
-    fn struct_gep(&mut self, ptr: Self::Value, idx: u64) -> Self::Value;
+    fn gep(&mut self, ty: Self::Type, ptr: Self::Value, indices: &[Self::Value]) -> Self::Value;
+    fn inbounds_gep(
+        &mut self,
+        ty: Self::Type,
+        ptr: Self::Value,
+        indices: &[Self::Value],
+    ) -> Self::Value;
+    fn struct_gep(&mut self, ty: Self::Type, ptr: Self::Value, idx: u64) -> Self::Value;
 
     fn trunc(&mut self, val: Self::Value, dest_ty: Self::Type) -> Self::Value;
     fn sext(&mut self, val: Self::Value, dest_ty: Self::Type) -> Self::Value;
diff --git a/compiler/rustc_llvm/llvm-wrapper/RustWrapper.cpp b/compiler/rustc_llvm/llvm-wrapper/RustWrapper.cpp
@@ -1551,6 +1551,16 @@ extern "C" void LLVMRustSetLinkage(LLVMValueRef V,
   LLVMSetLinkage(V, fromRust(RustLinkage));
 }
 
+extern "C" LLVMValueRef LLVMRustConstInBoundsGEP2(LLVMTypeRef Ty,
+                                                  LLVMValueRef ConstantVal,
+                                                  LLVMValueRef *ConstantIndices,
+                                                  unsigned NumIndices) {
+  ArrayRef<Constant *> IdxList(unwrap<Constant>(ConstantIndices, NumIndices),
+                               NumIndices);
+  Constant *Val = unwrap<Constant>(ConstantVal);
+  return wrap(ConstantExpr::getInBoundsGetElementPtr(unwrap(Ty), Val, IdxList));
+}
+
 // Returns true if both high and low were successfully set. Fails in case constant wasn’t any of
 // the common sizes (1, 8, 16, 32, 64, 128 bits)
 extern "C" bool LLVMRustConstInt128Get(LLVMValueRef CV, bool sext, uint64_t *high, uint64_t *low)
