Auto merge of #142771 - dianqk:mir-stmt-debuginfo, r=<try>

Introduce debuginfo to statements in MIR

The PR introduces support for debug information within dead statements. Currently, only the reference statement is supported, which is sufficient to fix #128081.

I don't modify Stable MIR, as I don't think we need debug information when using it.

This PR represents the debug information for the dead reference statement via `#dbg_value`. For example, `let _foo_b = &foo.b` becomes `#dbg_value(ptr %foo, !22, !DIExpression(DW_OP_plus_uconst, 4, DW_OP_stack_value), !26)`. You can see this here: https://rust.godbolt.org/z/d43js6adv.

The general principle for handling debug information is to never provide less debug information than the optimized LLVM IR.

The current rules for dropping debug information in this PR are:

- If the LLVM IR cannot represent a reference address, it's replaced with poison or simply dropped. For example, see: https://rust.godbolt.org/z/shGqPec8W. I'm using poison in all such cases now.
- All debuginfos is dropped when merging multiple successor BBs. An example is available here: https://rust.godbolt.org/z/TE1q3Wq6M.

> I haven't drop debuginfos in `MatchBranchSimplification`, because LLVM also pick one branch for it.

For [the perf result](#142771 (comment)):

I expected this to introduce some regressions;  however, the results mixed the effects of inlining. Looking at the doc profile, this is a clear optimization. One potential regression I'm investigating is `serde-1.0.219-debug-full`.

Author: rust-bors[bot]

compiler/rustc_borrowck/src/diagnostics/mutability_errors.rs

@@ -1146,7 +1146,6 @@ impl<'infcx, 'tcx> MirBorrowckCtxt<'_, 'infcx, 'tcx> {
                 let opt_assignment_rhs_span =
                     self.find_assignments(local).first().map(|&location| {
                         if let Some(mir::Statement {
-                            source_info: _,
                             kind:
                                 mir::StatementKind::Assign(box (
                                     _,

compiler/rustc_codegen_gcc/src/debuginfo.rs

@@ -26,14 +26,17 @@ impl<'a, 'gcc, 'tcx> DebugInfoBuilderMethods for Builder<'a, 'gcc, 'tcx> {
         &mut self,
         _dbg_var: Self::DIVariable,
         _dbg_loc: Self::DILocation,
-        _variable_alloca: Self::Value,
+        is_declared: bool,
+        val: Self::Value,
         _direct_offset: Size,
         _indirect_offsets: &[Size],
         _fragment: Option<Range<Size>>,
     ) {
         // FIXME(tempdragon): Not sure if this is correct, probably wrong but still keep it here.
         #[cfg(feature = "master")]
-        _variable_alloca.set_location(_dbg_loc);
+        if is_declared {
+            val.set_location(_dbg_loc);
+        }
     }
 
     fn insert_reference_to_gdb_debug_scripts_section_global(&mut self) {

compiler/rustc_codegen_llvm/src/debuginfo/dwarf_const.rs

@@ -35,3 +35,6 @@ declare_constant!(DW_OP_plus_uconst: u64);
 /// Double-checked by a static assertion in `RustWrapper.cpp`.
 #[allow(non_upper_case_globals)]
 pub(crate) const DW_OP_LLVM_fragment: u64 = 0x1000;
+// It describes the actual value of a source variable which might not exist in registers or in memory.
+#[allow(non_upper_case_globals)]
+pub(crate) const DW_OP_stack_value: u64 = 0x9f;

compiler/rustc_codegen_llvm/src/debuginfo/mod.rs

@@ -160,19 +160,23 @@ impl<'ll> DebugInfoBuilderMethods for Builder<'_, 'll, '_> {
         &mut self,
         dbg_var: &'ll DIVariable,
         dbg_loc: &'ll DILocation,
-        variable_alloca: Self::Value,
+        is_declared: bool,
+        val: Self::Value,
         direct_offset: Size,
         indirect_offsets: &[Size],
         fragment: Option<Range<Size>>,
     ) {
-        use dwarf_const::{DW_OP_LLVM_fragment, DW_OP_deref, DW_OP_plus_uconst};
+        use dwarf_const::{DW_OP_LLVM_fragment, DW_OP_deref, DW_OP_plus_uconst, DW_OP_stack_value};
 
         // Convert the direct and indirect offsets and fragment byte range to address ops.
         let mut addr_ops = SmallVec::<[u64; 8]>::new();
 
         if direct_offset.bytes() > 0 {
             addr_ops.push(DW_OP_plus_uconst);
             addr_ops.push(direct_offset.bytes() as u64);
+            if !is_declared {
+                addr_ops.push(DW_OP_stack_value);
+            }
         }
         for &offset in indirect_offsets {
             addr_ops.push(DW_OP_deref);
@@ -189,17 +193,30 @@ impl<'ll> DebugInfoBuilderMethods for Builder<'_, 'll, '_> {
             addr_ops.push((fragment.end - fragment.start).bits() as u64);
         }
 
-        unsafe {
-            // FIXME(eddyb) replace `llvm.dbg.declare` with `llvm.dbg.addr`.
-            llvm::LLVMRustDIBuilderInsertDeclareAtEnd(
-                DIB(self.cx()),
-                variable_alloca,
-                dbg_var,
-                addr_ops.as_ptr(),
-                addr_ops.len() as c_uint,
-                dbg_loc,
-                self.llbb(),
-            );
+        if is_declared {
+            unsafe {
+                llvm::LLVMRustDIBuilderInsertDeclareAtEnd(
+                    DIB(self.cx()),
+                    val,
+                    dbg_var,
+                    addr_ops.as_ptr(),
+                    addr_ops.len() as c_uint,
+                    dbg_loc,
+                    self.llbb(),
+                );
+            }
+        } else {
+            unsafe {
+                llvm::LLVMRustDIBuilderInsertDbgValueAtEnd(
+                    DIB(self.cx()),
+                    val,
+                    dbg_var,
+                    addr_ops.as_ptr(),
+                    addr_ops.len() as c_uint,
+                    dbg_loc,
+                    self.llbb(),
+                );
+            }
         }
     }
 

compiler/rustc_codegen_llvm/src/llvm/ffi.rs

@@ -2308,6 +2308,16 @@ unsafe extern "C" {
         InsertAtEnd: &'a BasicBlock,
     );
 
+    pub(crate) fn LLVMRustDIBuilderInsertDbgValueAtEnd<'a>(
+        Builder: &DIBuilder<'a>,
+        Val: &'a Value,
+        VarInfo: &'a DIVariable,
+        AddrOps: *const u64,
+        AddrOpsCount: c_uint,
+        DL: &'a DILocation,
+        InsertAtEnd: &'a BasicBlock,
+    );
+
     pub(crate) fn LLVMRustDIBuilderCreateEnumerator<'a>(
         Builder: &DIBuilder<'a>,
         Name: *const c_char,

compiler/rustc_codegen_ssa/src/mir/block.rs

@@ -1252,6 +1252,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             for statement in &data.statements {
                 self.codegen_statement(bx, statement);
             }
+            self.codegen_stmt_debuginfos(bx, &data.after_last_stmt_debuginfos);
 
             let merging_succ = self.codegen_terminator(bx, bb, data.terminator());
             if let MergingSucc::False = merging_succ {

compiler/rustc_codegen_ssa/src/mir/debuginfo.rs

@@ -253,6 +253,52 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
         spill_slot
     }
 
+    pub(crate) fn debug_new_value_to_local(
+        &self,
+        bx: &mut Bx,
+        local: mir::Local,
+        base: PlaceValue<Bx::Value>,
+        layout: TyAndLayout<'tcx>,
+        projection: &[mir::PlaceElem<'tcx>],
+    ) {
+        let full_debug_info = bx.sess().opts.debuginfo == DebugInfo::Full;
+        if !full_debug_info {
+            return;
+        }
+
+        let vars = match &self.per_local_var_debug_info {
+            Some(per_local) => &per_local[local],
+            None => return,
+        };
+
+        for var in vars.iter().cloned() {
+            self.debug_new_value_to_local_as_var(bx, base, layout, projection, var);
+        }
+    }
+
+    fn debug_new_value_to_local_as_var(
+        &self,
+        bx: &mut Bx,
+        base: PlaceValue<Bx::Value>,
+        layout: TyAndLayout<'tcx>,
+        projection: &[mir::PlaceElem<'tcx>],
+        var: PerLocalVarDebugInfo<'tcx, Bx::DIVariable>,
+    ) {
+        let Some(dbg_var) = var.dbg_var else { return };
+        let Some(dbg_loc) = self.dbg_loc(var.source_info) else { return };
+        let DebugInfoOffset { direct_offset, indirect_offsets, result: _ } =
+            calculate_debuginfo_offset(bx, projection, layout);
+        bx.dbg_var_addr(
+            dbg_var,
+            dbg_loc,
+            false,
+            base.llval,
+            direct_offset,
+            &indirect_offsets,
+            var.fragment,
+        );
+    }
+
     /// Apply debuginfo and/or name, after creating the `alloca` for a local,
     /// or initializing the local with an operand (whichever applies).
     pub(crate) fn debug_introduce_local(&self, bx: &mut Bx, local: mir::Local) {
@@ -421,6 +467,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             bx.dbg_var_addr(
                 dbg_var,
                 dbg_loc,
+                true,
                 alloca.val.llval,
                 Size::ZERO,
                 &[Size::ZERO],
@@ -430,6 +477,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             bx.dbg_var_addr(
                 dbg_var,
                 dbg_loc,
+                true,
                 base.val.llval,
                 direct_offset,
                 &indirect_offsets,
@@ -455,7 +503,7 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
                 let base = FunctionCx::spill_operand_to_stack(operand, Some(name), bx);
                 bx.clear_dbg_loc();
 
-                bx.dbg_var_addr(dbg_var, dbg_loc, base.val.llval, Size::ZERO, &[], fragment);
+                bx.dbg_var_addr(dbg_var, dbg_loc, true, base.val.llval, Size::ZERO, &[], fragment);
             }
         }
     }

compiler/rustc_codegen_ssa/src/mir/statement.rs

@@ -1,13 +1,17 @@
-use rustc_middle::mir::{self, NonDivergingIntrinsic};
-use rustc_middle::span_bug;
+use rustc_middle::mir::{self, NonDivergingIntrinsic, StmtDebugInfo};
+use rustc_middle::{bug, span_bug};
 use tracing::instrument;
 
 use super::{FunctionCx, LocalRef};
+use crate::common::TypeKind;
+use crate::mir::operand::OperandValue;
+use crate::mir::place::PlaceRef;
 use crate::traits::*;
 
 impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
     #[instrument(level = "debug", skip(self, bx))]
     pub(crate) fn codegen_statement(&mut self, bx: &mut Bx, statement: &mir::Statement<'tcx>) {
+        self.codegen_stmt_debuginfos(bx, &statement.debuginfos);
         self.set_debug_loc(bx, statement.source_info);
         match statement.kind {
             mir::StatementKind::Assign(box (ref place, ref rvalue)) => {
@@ -101,4 +105,66 @@ impl<'a, 'tcx, Bx: BuilderMethods<'a, 'tcx>> FunctionCx<'a, 'tcx, Bx> {
             | mir::StatementKind::Nop => {}
         }
     }
+
+    pub(crate) fn codegen_stmt_debuginfo(&mut self, bx: &mut Bx, debuginfo: &StmtDebugInfo<'tcx>) {
+        match debuginfo {
+            StmtDebugInfo::AssignRef(dest, place) => {
+                let place_ref = match self.locals[place.local] {
+                    LocalRef::Place(place_ref) | LocalRef::UnsizedPlace(place_ref) => {
+                        Some(place_ref)
+                    }
+                    LocalRef::Operand(operand_ref) => match operand_ref.val {
+                        OperandValue::Immediate(v) => {
+                            Some(PlaceRef::new_sized(v, operand_ref.layout))
+                        }
+                        OperandValue::Ref(_)
+                        | OperandValue::Pair(_, _)
+                        | OperandValue::ZeroSized => None,
+                    },
+                    LocalRef::PendingOperand => None,
+                }
+                .filter(|place_ref| {
+                    // Drop unsupported projections.
+                    // FIXME: Add a test case.
+                    place.projection.iter().all(|p| p.can_use_in_debuginfo()) &&
+                    // Only pointers can calculate addresses.
+                    bx.type_kind(bx.val_ty(place_ref.val.llval)) == TypeKind::Pointer
+                });
+                let (val, layout, projection) =
+                    match (place_ref, place.is_indirect_first_projection()) {
+                        (Some(place_ref), false) => {
+                            (place_ref.val, place_ref.layout, place.projection.as_slice())
+                        }
+                        (Some(place_ref), true) => {
+                            let projected_ty =
+                                place_ref.layout.ty.builtin_deref(true).unwrap_or_else(|| {
+                                    bug!("deref of non-pointer {:?}", place_ref)
+                                });
+                            let layout = bx.cx().layout_of(projected_ty);
+                            (place_ref.val, layout, &place.projection[1..])
+                        }
+                        _ => {
+                            // If the address cannot be computed, use poison to indicate that the value has been optimized out.
+                            let ty = self.monomorphize(self.mir.local_decls[*dest].ty);
+                            let layout = bx.cx().layout_of(ty);
+                            let to_backend_ty = bx.cx().immediate_backend_type(layout);
+                            let place_ref =
+                                PlaceRef::new_sized(bx.cx().const_poison(to_backend_ty), layout);
+                            (place_ref.val, layout, [].as_slice())
+                        }
+                    };
+                self.debug_new_value_to_local(bx, *dest, val, layout, projection);
+            }
+        }
+    }
+
+    pub(crate) fn codegen_stmt_debuginfos(
+        &mut self,
+        bx: &mut Bx,
+        debuginfos: &[StmtDebugInfo<'tcx>],
+    ) {
+        for debuginfo in debuginfos {
+            self.codegen_stmt_debuginfo(bx, debuginfo);
+        }
+    }
 }

compiler/rustc_codegen_ssa/src/traits/debuginfo.rs

@@ -71,7 +71,8 @@ pub trait DebugInfoBuilderMethods: BackendTypes {
         &mut self,
         dbg_var: Self::DIVariable,
         dbg_loc: Self::DILocation,
-        variable_alloca: Self::Value,
+        is_declared: bool,
+        val: Self::Value,
         direct_offset: Size,
         // NB: each offset implies a deref (i.e. they're steps in a pointer chain).
         indirect_offsets: &[Size],

compiler/rustc_llvm/llvm-wrapper/RustWrapper.cpp

@@ -58,6 +58,7 @@ using namespace llvm::object;
 // This opcode is an LLVM detail that could hypothetically change (?), so
 // verify that the hard-coded value in `dwarf_const.rs` still agrees with LLVM.
 static_assert(dwarf::DW_OP_LLVM_fragment == 0x1000);
+static_assert(dwarf::DW_OP_stack_value == 0x9f);
 
 // LLVMAtomicOrdering is already an enum - don't create another
 // one.
@@ -1241,6 +1242,18 @@ LLVMRustDIBuilderInsertDeclareAtEnd(LLVMDIBuilderRef Builder, LLVMValueRef V,
       DebugLoc(cast<MDNode>(unwrap(DL))), unwrap(InsertAtEnd));
 }
 
+extern "C" void
+LLVMRustDIBuilderInsertDbgValueAtEnd(LLVMDIBuilderRef Builder, LLVMValueRef V,
+                                     LLVMMetadataRef VarInfo, uint64_t *AddrOps,
+                                     unsigned AddrOpsCount, LLVMMetadataRef DL,
+                                     LLVMBasicBlockRef InsertAtEnd) {
+  unwrap(Builder)->insertDbgValueIntrinsic(
+      unwrap(V), unwrap<DILocalVariable>(VarInfo),
+      unwrap(Builder)->createExpression(
+          llvm::ArrayRef<uint64_t>(AddrOps, AddrOpsCount)),
+      DebugLoc(cast<MDNode>(unwrap(DL))), unwrap(InsertAtEnd));
+}
+
 extern "C" LLVMMetadataRef
 LLVMRustDIBuilderCreateEnumerator(LLVMDIBuilderRef Builder, const char *Name,
                                   size_t NameLen, const uint64_t Value[2],