PPC: Display data values on hover for pools as well (#140)

* Fix missing dependency feature for objdiff-gui

* Update .gitignore

* PPC: Display data values on hover for pools as well

* Tooltip data display: Format floats and doubles better

Floats and doubles will now always be displayed with a decimal point and one digit after it, even if they are whole numbers. Floats will also have the f suffix. This is so you can tell the data type just by glancing at the value.

* Move big functions to bottom ppc.rs

* Clear pool relocs in volatile registers on function call

This fixes some false positives.

* Revert ObjArch API changes, add fake target symbol hack

Because we no longer have access to the actual symbol name via sections, guess_data_type can no longer detect the String data type for pooled references.

* Add hack to detect strings via the addi opcode

* Move hack to resolve placeholder symbol into process_code_symbol

* Merge reloc and fake_pool_reloc fields of ObjIns

objdiff-core/src/arch/mod.rs

@@ -36,13 +36,22 @@ pub enum DataType {
 
 impl DataType {
     pub fn display_bytes<Endian: ByteOrder>(&self, bytes: &[u8]) -> Option<String> {
-        // TODO: Attempt to interpret large symbols as arrays of a smaller type,
-        // fallback to intrepreting it as bytes.
-        // https://github.com/encounter/objdiff/issues/124
-        if self.required_len().is_some_and(|l| bytes.len() != l) {
-            log::warn!("Failed to display a symbol value for a symbol whose size doesn't match the instruction referencing it.");
+        if self.required_len().is_some_and(|l| bytes.len() < l) {
+            log::warn!("Failed to display a symbol value for a symbol whose size is too small for instruction referencing it.");
             return None;
         }
+        let mut bytes = bytes;
+        if self.required_len().is_some_and(|l| bytes.len() > l) {
+            // If the symbol's size is larger a single instance of this data type, we take just the
+            // bytes necessary for one of them in order to display the first element of the array.
+            bytes = &bytes[0..self.required_len().unwrap()];
+            // TODO: Attempt to interpret large symbols as arrays of a smaller type and show all
+            // elements of the array instead. https://github.com/encounter/objdiff/issues/124
+            // However, note that the stride of an array can not always be determined just by the
+            // data type guessed by the single instruction accessing it. There can also be arrays of
+            // structs that contain multiple elements of different types, so if other elements after
+            // the first one were to be displayed in this manner, they may be inaccurate.
+        }
 
         match self {
             DataType::Int8 => {
@@ -86,10 +95,10 @@ impl DataType {
                 }
             }
             DataType::Float => {
-                format!("Float: {}", Endian::read_f32(bytes))
+                format!("Float: {:?}f", Endian::read_f32(bytes))
             }
             DataType::Double => {
-                format!("Double: {}", Endian::read_f64(bytes))
+                format!("Double: {:?}", Endian::read_f64(bytes))
             }
             DataType::Bytes => {
                 format!("Bytes: {:#?}", bytes)

objdiff-core/src/arch/ppc.rs

@@ -1,4 +1,7 @@
-use std::{borrow::Cow, collections::BTreeMap};
+use std::{
+    borrow::Cow,
+    collections::{BTreeMap, HashMap},
+};
 
 use anyhow::{bail, ensure, Result};
 use byteorder::BigEndian;
@@ -7,7 +10,7 @@ use object::{
     elf, File, Object, ObjectSection, ObjectSymbol, Relocation, RelocationFlags, RelocationTarget,
     Symbol, SymbolKind,
 };
-use ppc750cl::{Argument, InsIter, Opcode, GPR};
+use ppc750cl::{Argument, InsIter, Opcode, ParsedIns, GPR};
 
 use crate::{
     arch::{DataType, ObjArch, ProcessCodeResult},
@@ -49,6 +52,8 @@ impl ObjArch for ObjArchPpc {
         let ins_count = code.len() / 4;
         let mut ops = Vec::<u16>::with_capacity(ins_count);
         let mut insts = Vec::<ObjIns>::with_capacity(ins_count);
+        let fake_pool_reloc_for_addr =
+            generate_fake_pool_reloc_for_addr_mapping(address, code, relocations);
         for (cur_addr, mut ins) in InsIter::new(code, address as u32) {
             let reloc = relocations.iter().find(|r| (r.address as u32 & !3) == cur_addr);
             if let Some(reloc) = reloc {
@@ -145,7 +150,7 @@ impl ObjArch for ObjArchPpc {
                 size: 4,
                 mnemonic: Cow::Borrowed(simplified.mnemonic),
                 args,
-                reloc: reloc.cloned(),
+                reloc: reloc.or(fake_pool_reloc_for_addr.get(&cur_addr)).cloned(),
                 op: ins.op as u16,
                 branch_dest,
                 line,
@@ -173,6 +178,7 @@ impl ObjArch for ObjArchPpc {
     fn display_reloc(&self, flags: RelocationFlags) -> Cow<'static, str> {
         match flags {
             RelocationFlags::Elf { r_type } => match r_type {
+                elf::R_PPC_NONE => Cow::Borrowed("R_PPC_NONE"), // We use this for fake pool relocs
                 elf::R_PPC_ADDR16_LO => Cow::Borrowed("R_PPC_ADDR16_LO"),
                 elf::R_PPC_ADDR16_HI => Cow::Borrowed("R_PPC_ADDR16_HI"),
                 elf::R_PPC_ADDR16_HA => Cow::Borrowed("R_PPC_ADDR16_HA"),
@@ -188,26 +194,22 @@ impl ObjArch for ObjArchPpc {
     }
 
     fn guess_data_type(&self, instruction: &ObjIns) -> Option<super::DataType> {
-        // Always shows the first string of the table. Not ideal, but it's really hard to find
-        // the actual string being referenced.
         if instruction.reloc.as_ref().is_some_and(|r| r.target.name.starts_with("@stringBase")) {
             return Some(DataType::String);
         }
 
-        match Opcode::from(instruction.op as u8) {
-            Opcode::Lbz | Opcode::Lbzu | Opcode::Lbzux | Opcode::Lbzx => Some(DataType::Int8),
-            Opcode::Lhz | Opcode::Lhzu | Opcode::Lhzux | Opcode::Lhzx => Some(DataType::Int16),
-            Opcode::Lha | Opcode::Lhau | Opcode::Lhaux | Opcode::Lhax => Some(DataType::Int16),
-            Opcode::Lwz | Opcode::Lwzu | Opcode::Lwzux | Opcode::Lwzx => Some(DataType::Int32),
-            Opcode::Lfs | Opcode::Lfsu | Opcode::Lfsux | Opcode::Lfsx => Some(DataType::Float),
-            Opcode::Lfd | Opcode::Lfdu | Opcode::Lfdux | Opcode::Lfdx => Some(DataType::Double),
-
-            Opcode::Stb | Opcode::Stbu | Opcode::Stbux | Opcode::Stbx => Some(DataType::Int8),
-            Opcode::Sth | Opcode::Sthu | Opcode::Sthux | Opcode::Sthx => Some(DataType::Int16),
-            Opcode::Stw | Opcode::Stwu | Opcode::Stwux | Opcode::Stwx => Some(DataType::Int32),
-            Opcode::Stfs | Opcode::Stfsu | Opcode::Stfsux | Opcode::Stfsx => Some(DataType::Float),
-            Opcode::Stfd | Opcode::Stfdu | Opcode::Stfdux | Opcode::Stfdx => Some(DataType::Double),
-            _ => None,
+        let op = Opcode::from(instruction.op as u8);
+        if let Some(ty) = guess_data_type_from_load_store_inst_op(op) {
+            Some(ty)
+        } else if op == Opcode::Addi {
+            // Assume that any addi instruction that references a local symbol is loading a string.
+            // This hack is not ideal and results in tons of false positives where it will show
+            // garbage strings (e.g. misinterpreting arrays, float literals, etc).
+            // But not all strings are in the @stringBase pool, so the condition above that checks
+            // the target symbol name would miss some.
+            Some(DataType::String)
+        } else {
+            None
         }
     }
 
@@ -381,3 +383,196 @@ fn make_symbol_ref(symbol: &Symbol) -> Result<ExtabSymbolRef> {
     let demangled_name = cwdemangle::demangle(&name, &cwdemangle::DemangleOptions::default());
     Ok(ExtabSymbolRef { original_index: symbol.index().0, name, demangled_name })
 }
+
+fn guess_data_type_from_load_store_inst_op(inst_op: Opcode) -> Option<DataType> {
+    match inst_op {
+        Opcode::Lbz | Opcode::Lbzu | Opcode::Lbzux | Opcode::Lbzx => Some(DataType::Int8),
+        Opcode::Lhz | Opcode::Lhzu | Opcode::Lhzux | Opcode::Lhzx => Some(DataType::Int16),
+        Opcode::Lha | Opcode::Lhau | Opcode::Lhaux | Opcode::Lhax => Some(DataType::Int16),
+        Opcode::Lwz | Opcode::Lwzu | Opcode::Lwzux | Opcode::Lwzx => Some(DataType::Int32),
+        Opcode::Lfs | Opcode::Lfsu | Opcode::Lfsux | Opcode::Lfsx => Some(DataType::Float),
+        Opcode::Lfd | Opcode::Lfdu | Opcode::Lfdux | Opcode::Lfdx => Some(DataType::Double),
+
+        Opcode::Stb | Opcode::Stbu | Opcode::Stbux | Opcode::Stbx => Some(DataType::Int8),
+        Opcode::Sth | Opcode::Sthu | Opcode::Sthux | Opcode::Sthx => Some(DataType::Int16),
+        Opcode::Stw | Opcode::Stwu | Opcode::Stwux | Opcode::Stwx => Some(DataType::Int32),
+        Opcode::Stfs | Opcode::Stfsu | Opcode::Stfsux | Opcode::Stfsx => Some(DataType::Float),
+        Opcode::Stfd | Opcode::Stfdu | Opcode::Stfdux | Opcode::Stfdx => Some(DataType::Double),
+        _ => None,
+    }
+}
+
+// Given an instruction, determine if it could accessing data at the address in a register.
+// If so, return the offset added to the register's address, the register containing that address,
+// and (optionally) which destination register the address is being copied into.
+fn get_offset_and_addr_gpr_for_possible_pool_reference(
+    opcode: Opcode,
+    simplified: &ParsedIns,
+) -> Option<(i16, GPR, Option<GPR>)> {
+    let args = &simplified.args;
+    if guess_data_type_from_load_store_inst_op(opcode).is_some() {
+        match (args[1], args[2]) {
+            (Argument::Offset(offset), Argument::GPR(addr_src_gpr)) => {
+                // e.g. lwz. Immediate offset.
+                Some((offset.0, addr_src_gpr, None))
+            }
+            (Argument::GPR(addr_src_gpr), Argument::GPR(_offset_gpr)) => {
+                // e.g. lwzx. The offset is in a register and was likely calculated from an index.
+                // Treat the offset as being 0 in this case to show the first element of the array.
+                // It may be possible to show all elements by figuring out the stride of the array
+                // from the calculations performed on the index before it's put into offset_gpr, but
+                // this would be much more complicated, so it's not currently done.
+                Some((0, addr_src_gpr, None))
+            }
+            _ => None,
+        }
+    } else {
+        // If it's not a load/store instruction, there's two more possibilities we need to handle.
+        // 1. It could be loading a pointer to a string.
+        // 2. It could be moving the relocation address plus an offset into a different register to
+        //    load from later.
+        // If either of these match, we also want to return the destination register that the
+        // address is being copied into so that we can detect any future references to that new
+        // register as well.
+        match (opcode, args[0], args[1], args[2]) {
+            (
+                Opcode::Addi,
+                Argument::GPR(addr_dst_gpr),
+                Argument::GPR(addr_src_gpr),
+                Argument::Simm(simm),
+            ) => Some((simm.0, addr_src_gpr, Some(addr_dst_gpr))),
+            (
+                Opcode::Or,
+                Argument::GPR(addr_dst_gpr),
+                Argument::GPR(addr_src_gpr),
+                Argument::None,
+            ) => Some((0, addr_src_gpr, Some(addr_dst_gpr))), // `mr` or `mr.`
+            _ => None,
+        }
+    }
+}
+
+// We create a fake relocation for an instruction, vaguely simulating what the actual relocation
+// might have looked like if it wasn't pooled. This is so minimal changes are needed to display
+// pooled accesses vs non-pooled accesses. We set the relocation type to R_PPC_NONE to indicate that
+// there isn't really a relocation here, as copying the pool relocation's type wouldn't make sense.
+// Also, if this instruction is accessing the middle of a symbol instead of the start, we add an
+// addend to indicate that.
+fn make_fake_pool_reloc(offset: i16, cur_addr: u32, pool_reloc: &ObjReloc) -> Option<ObjReloc> {
+    let offset_from_pool = pool_reloc.addend + offset as i64;
+    let target_address = pool_reloc.target.address.checked_add_signed(offset_from_pool)?;
+    let orig_section_index = pool_reloc.target.orig_section_index?;
+    // We also need to create a fake target symbol to go inside our fake relocation.
+    // This is because we don't have access to list of all symbols in this section, so we can't find
+    // the real symbol yet. Instead we make a placeholder that has the correct `orig_section_index`
+    // and `address` fields, and then later on when this information is displayed to the user, we
+    // can find the real symbol by searching through the object's section's symbols for one that
+    // contains this address.
+    let fake_target_symbol = ObjSymbol {
+        name: "".to_string(),
+        demangled_name: None,
+        address: target_address,
+        section_address: 0,
+        size: 0,
+        size_known: false,
+        kind: Default::default(),
+        flags: Default::default(),
+        orig_section_index: Some(orig_section_index),
+        virtual_address: None,
+        original_index: None,
+        bytes: vec![],
+    };
+    // The addend is also fake because we don't know yet if the `target_address` here is the exact
+    // start of the symbol or if it's in the middle of it.
+    let fake_addend = 0;
+    Some(ObjReloc {
+        flags: RelocationFlags::Elf { r_type: elf::R_PPC_NONE },
+        address: cur_addr as u64,
+        target: fake_target_symbol,
+        addend: fake_addend,
+    })
+}
+
+// Searches through all instructions in a function, determining which registers have the addresses
+// of pooled data relocations in them, finding which instructions load data from those addresses,
+// and constructing a mapping of the address of that instruction to a "fake pool relocation" that
+// simulates what that instruction's relocation would look like if data hadn't been pooled.
+// Limitations: This method currently only goes through the instructions in a function in linear
+// order, from start to finish. It does *not* follow any branches. This means that it could have
+// false positives or false negatives in determining which relocation is currently loaded in which
+// register at any given point in the function, as control flow is not respected.
+// There are currently no known examples of this method producing inaccurate results in reality, but
+// if examples are found, it may be possible to update this method to also follow all branches so
+// that it produces more accurate results.
+fn generate_fake_pool_reloc_for_addr_mapping(
+    address: u64,
+    code: &[u8],
+    relocations: &[ObjReloc],
+) -> HashMap<u32, ObjReloc> {
+    let mut active_pool_relocs = HashMap::new();
+    let mut pool_reloc_for_addr = HashMap::new();
+    for (cur_addr, ins) in InsIter::new(code, address as u32) {
+        let simplified = ins.simplified();
+        let reloc = relocations.iter().find(|r| (r.address as u32 & !3) == cur_addr);
+
+        if let Some(reloc) = reloc {
+            // This instruction has a real relocation, so it may be a pool load we want to keep
+            // track of.
+            let args = &simplified.args;
+            match (ins.op, args[0], args[1], args[2]) {
+                (
+                    Opcode::Addi,
+                    Argument::GPR(addr_dst_gpr),
+                    Argument::GPR(_addr_src_gpr),
+                    Argument::Simm(_simm),
+                ) => {
+                    active_pool_relocs.insert(addr_dst_gpr.0, reloc.clone()); // `lis` + `addi`
+                }
+                (
+                    Opcode::Ori,
+                    Argument::GPR(addr_dst_gpr),
+                    Argument::GPR(_addr_src_gpr),
+                    Argument::Uimm(_uimm),
+                ) => {
+                    active_pool_relocs.insert(addr_dst_gpr.0, reloc.clone()); // `lis` + `ori`
+                }
+                (Opcode::B, _, _, _) => {
+                    if simplified.mnemonic == "bl" {
+                        // When encountering a function call, clear any active pool relocations from
+                        // the volatile registers (r0, r3-r12), but not the nonvolatile registers.
+                        active_pool_relocs.remove(&0);
+                        for gpr in 3..12 {
+                            active_pool_relocs.remove(&gpr);
+                        }
+                    }
+                }
+                _ => {}
+            }
+        } else if let Some((offset, addr_src_gpr, addr_dst_gpr)) =
+            get_offset_and_addr_gpr_for_possible_pool_reference(ins.op, &simplified)
+        {
+            // This instruction doesn't have a real relocation, so it may be a reference to one of
+            // the already-loaded pools.
+            if let Some(pool_reloc) = active_pool_relocs.get(&addr_src_gpr.0) {
+                if let Some(fake_pool_reloc) = make_fake_pool_reloc(offset, cur_addr, pool_reloc) {
+                    pool_reloc_for_addr.insert(cur_addr, fake_pool_reloc);
+                }
+                if let Some(addr_dst_gpr) = addr_dst_gpr {
+                    // If the address of the pool relocation got copied into another register, we
+                    // need to keep track of it in that register too as future instructions may
+                    // reference the symbol indirectly via this new register, instead of the
+                    // register the symbol's address was originally loaded into.
+                    // For example, the start of the function might `lis` + `addi` the start of the
+                    // ...data pool into r25, and then later the start of a loop will `addi` r25
+                    // with the offset within the .data section of an array variable into r21.
+                    // Then the body of the loop will `lwzx` one of the array elements from r21.
+                    let mut new_reloc = pool_reloc.clone();
+                    new_reloc.addend += offset as i64;
+                    active_pool_relocs.insert(addr_dst_gpr.0, new_reloc);
+                }
+            }
+        }
+    }
+
+    pool_reloc_for_addr
+}

objdiff-core/src/diff/code.rs

@@ -9,7 +9,7 @@ use crate::{
         DiffObjConfig, ObjInsArgDiff, ObjInsBranchFrom, ObjInsBranchTo, ObjInsDiff, ObjInsDiffKind,
         ObjSymbolDiff,
     },
-    obj::{ObjInfo, ObjInsArg, ObjReloc, ObjSymbolFlags, SymbolRef},
+    obj::{ObjInfo, ObjInsArg, ObjReloc, ObjSection, ObjSymbol, ObjSymbolFlags, SymbolRef},
 };
 
 pub fn process_code_symbol(
@@ -21,14 +21,30 @@ pub fn process_code_symbol(
     let section = section.ok_or_else(|| anyhow!("Code symbol section not found"))?;
     let code = &section.data
         [symbol.section_address as usize..(symbol.section_address + symbol.size) as usize];
-    obj.arch.process_code(
+    let mut res = obj.arch.process_code(
         symbol.address,
         code,
         section.orig_index,
         &section.relocations,
         &section.line_info,
         config,
-    )
+    )?;
+
+    for inst in res.insts.iter_mut() {
+        if let Some(reloc) = &mut inst.reloc {
+            if reloc.target.size == 0 && reloc.target.name.is_empty() {
+                // Fake target symbol we added as a placeholder. We need to find the real one.
+                if let Some(real_target) =
+                    find_symbol_matching_fake_symbol_in_sections(&reloc.target, &obj.sections)
+                {
+                    reloc.addend = (reloc.target.address - real_target.address) as i64;
+                    reloc.target = real_target;
+                }
+            }
+        }
+    }
+
+    Ok(res)
 }
 
 pub fn no_diff_code(out: &ProcessCodeResult, symbol_ref: SymbolRef) -> Result<ObjSymbolDiff> {
@@ -369,3 +385,16 @@ fn compare_ins(
     }
     Ok(result)
 }
+
+fn find_symbol_matching_fake_symbol_in_sections(
+    fake_symbol: &ObjSymbol,
+    sections: &[ObjSection],
+) -> Option<ObjSymbol> {
+    let orig_section_index = fake_symbol.orig_section_index?;
+    let section = sections.iter().find(|s| s.orig_index == orig_section_index)?;
+    let real_symbol = section
+        .symbols
+        .iter()
+        .find(|s| s.size > 0 && (s.address..s.address + s.size).contains(&fake_symbol.address))?;
+    Some(real_symbol.clone())
+}