[PATCH v2] Use readelf -WSrs instead of objdump -thr so that llvm-readelf can be used as a replacement
objdump -h relies heavily on BFD internals and the BFD flags are difficult to emulate in llvm-objdump. llvm-objdump -h has a different output (https://reviews.llvm.org/D57146)
Switch to readelf, which is generally better than objdump when dumping ELF section/symbol information.
Signed-off-by: Fangrui Song maskray@google.com --- Makefile | 18 +++--- scripts/checkrom.py | 2 +- scripts/layoutrom.py | 143 ++++++++++++++++++++----------------------- 3 files changed, 78 insertions(+), 85 deletions(-) -- Changes v1->v2 * Deleted unrelated debug lines
diff --git a/Makefile b/Makefile index 118dec0..c56d4b7 100644 --- a/Makefile +++ b/Makefile @@ -20,7 +20,7 @@ endif AS=$(CROSS_PREFIX)as LD=$(CROSS_PREFIX)ld OBJCOPY=$(CROSS_PREFIX)objcopy -OBJDUMP=$(CROSS_PREFIX)objdump +READELF=$(CROSS_PREFIX)readelf STRIP=$(CROSS_PREFIX)strip PYTHON=python CPP=cpp @@ -161,14 +161,14 @@ $(OUT)romlayout.o: src/romlayout.S $(OUT)autoconf.h $(OUT)asm-offsets.h
$(OUT)romlayout16.lds: $(OUT)ccode32flat.o $(OUT)code32seg.o $(OUT)ccode16.o $(OUT)romlayout.o src/version.c scripts/layoutrom.py scripts/buildversion.py @echo " Building ld scripts" - $(Q)$(PYTHON) ./scripts/buildversion.py -e "$(EXTRAVERSION)" -t "$(CC);$(AS);$(LD);$(OBJCOPY);$(OBJDUMP);$(STRIP)" $(OUT)autoversion.h + $(Q)$(PYTHON) ./scripts/buildversion.py -e "$(EXTRAVERSION)" -t "$(CC);$(AS);$(LD);$(OBJCOPY);$(READELF);$(STRIP)" $(OUT)autoversion.h $(Q)$(CC) $(CFLAGS32FLAT) -c src/version.c -o $(OUT)version.o $(Q)$(LD) $(LD32BIT_FLAG) -r $(OUT)ccode32flat.o $(OUT)version.o -o $(OUT)code32flat.o $(Q)$(LD) $(LD32BIT_FLAG) -r $(OUT)ccode16.o $(OUT)romlayout.o -o $(OUT)code16.o - $(Q)$(OBJDUMP) -thr $(OUT)code32flat.o > $(OUT)code32flat.o.objdump - $(Q)$(OBJDUMP) -thr $(OUT)code32seg.o > $(OUT)code32seg.o.objdump - $(Q)$(OBJDUMP) -thr $(OUT)code16.o > $(OUT)code16.o.objdump - $(Q)$(PYTHON) ./scripts/layoutrom.py $(OUT)code16.o.objdump $(OUT)code32seg.o.objdump $(OUT)code32flat.o.objdump $(OUT)$(KCONFIG_AUTOHEADER) $(OUT)romlayout16.lds $(OUT)romlayout32seg.lds $(OUT)romlayout32flat.lds + $(Q)$(READELF) -WSrs $(OUT)code32flat.o > $(OUT)code32flat.o.readelf + $(Q)$(READELF) -WSrs $(OUT)code32seg.o > $(OUT)code32seg.o.readelf + $(Q)$(READELF) -WSrs $(OUT)code16.o > $(OUT)code16.o.readelf + $(Q)$(PYTHON) ./scripts/layoutrom.py $(OUT)code16.o.readelf $(OUT)code32seg.o.readelf $(OUT)code32flat.o.readelf $(OUT)$(KCONFIG_AUTOHEADER) $(OUT)romlayout16.lds $(OUT)romlayout32seg.lds $(OUT)romlayout32flat.lds
# These are actually built by scripts/layoutrom.py above, but by pulling them # into an extra rule we prevent make -j from spawning layoutrom.py 4 times. @@ -193,9 +193,9 @@ $(OUT)rom.o: $(OUT)rom16.strip.o $(OUT)rom32seg.strip.o $(OUT)code32flat.o $(OUT $(OUT)bios.bin.prep: $(OUT)rom.o scripts/checkrom.py @echo " Prepping $@" $(Q)rm -f $(OUT)bios.bin $(OUT)Csm16.bin $(OUT)bios.bin.elf - $(Q)$(OBJDUMP) -thr $< > $<.objdump + $(Q)$(READELF) -WSrs $< > $<.readelf $(Q)$(OBJCOPY) -O binary $< $(OUT)bios.bin.raw - $(Q)$(PYTHON) ./scripts/checkrom.py $<.objdump $(CONFIG_ROM_SIZE) $(OUT)bios.bin.raw $(OUT)bios.bin.prep + $(Q)$(PYTHON) ./scripts/checkrom.py $<.readelf $(CONFIG_ROM_SIZE) $(OUT)bios.bin.raw $(OUT)bios.bin.prep
$(OUT)bios.bin: $(OUT)bios.bin.prep @echo " Creating $@" @@ -237,7 +237,7 @@ $(OUT)vgaentry.o: vgasrc/vgaentry.S $(OUT)autoconf.h $(OUT)asm-offsets.h
$(OUT)vgarom.o: $(OUT)vgaccode16.o $(OUT)vgaentry.o $(OUT)vgasrc/vgalayout.lds vgasrc/vgaversion.c scripts/buildversion.py @echo " Linking $@" - $(Q)$(PYTHON) ./scripts/buildversion.py -e "$(EXTRAVERSION)" -t "$(CC);$(AS);$(LD);$(OBJCOPY);$(OBJDUMP);$(STRIP)" $(OUT)autovgaversion.h + $(Q)$(PYTHON) ./scripts/buildversion.py -e "$(EXTRAVERSION)" -t "$(CC);$(AS);$(LD);$(OBJCOPY);$(READELF);$(STRIP)" $(OUT)autovgaversion.h $(Q)$(CC) $(CFLAGS16) -c vgasrc/vgaversion.c -o $(OUT)vgaversion.o $(Q)$(LD) --gc-sections -T $(OUT)vgasrc/vgalayout.lds $(OUT)vgaccode16.o $(OUT)vgaentry.o $(OUT)vgaversion.o -o $@
diff --git a/scripts/checkrom.py b/scripts/checkrom.py index aced5e2..721444d 100755 --- a/scripts/checkrom.py +++ b/scripts/checkrom.py @@ -23,7 +23,7 @@ def main():
# Read in symbols objinfofile = open(objinfo, 'r') - symbols = layoutrom.parseObjDump(objinfofile, 'in')[1] + symbols = layoutrom.parseReadElf(objinfofile, 'in')[1]
# Read in raw file f = open(rawfile, 'rb') diff --git a/scripts/layoutrom.py b/scripts/layoutrom.py index 94d4412..350800e 100755 --- a/scripts/layoutrom.py +++ b/scripts/layoutrom.py @@ -544,87 +544,80 @@ class Reloc: class Symbol: name = offset = section = None
-# Read in output from objdump -def parseObjDump(file, fileid): +def parseReadElf(file, fileid): # sections = [section, ...] sections = [] sectionmap = {} # symbols[symbolname] = symbol symbols = {} + lines = file.readlines()
- state = None - for line in file.readlines(): - line = line.rstrip() - if line == 'Sections:': - state = 'section' - continue - if line == 'SYMBOL TABLE:': - state = 'symbol' - continue - if line.startswith('RELOCATION RECORDS FOR ['): - sectionname = line[24:-2] - if sectionname.startswith('.debug_'): - # Skip debugging sections (to reduce parsing time) - state = None - continue - state = 'reloc' - relocsection = sectionmap[sectionname] - continue + i = 0 + while not lines[i].startswith('Section Headers'): + i += 1 + i += 3 # Skip [Nr] and [ 0] + while ']' in lines[i]: + parts = lines[i][lines[i].find(']')+1:].split() + if len(parts) == 10: + name, _, _, _, size, _, _, _, _, align = parts + else: # not SHF_ALLOC + name, _, _, _, size, _, _, _, align = parts + section = Section() + section.name = name + section.size = int(size, 16) + section.align = int(align) + section.fileid = fileid + section.relocs = [] + sections.append(section) + sectionmap[name] = section + i += 1
- if state == 'section': - try: - idx, name, size, vma, lma, fileoff, align = line.split() - if align[:3] != '2**': - continue - section = Section() - section.name = name - section.size = int(size, 16) - section.align = 2**int(align[3:]) - section.fileid = fileid - section.relocs = [] - sections.append(section) - sectionmap[name] = section - except ValueError: - pass + while not lines[i].startswith('Symbol table '): + i += 1 + i += 3 # Skip Num: and 0: + while i < len(lines) and ':' in lines[i]: + colon = lines[i].find(':') + parts = lines[i].split() + try: + ndx = int(parts[6]) + name = sections[ndx-1].name + except ValueError: # ABS or UND + ndx = 0 + name = '' + if parts[3] != 'SECTION': + name = parts[7] + symbol = Symbol() + symbol.size = int(parts[2]) + symbol.offset = int(parts[1], 16) + symbol.name = name + symbol.section = sections[ndx-1] if ndx > 0 else None + symbols[name] = symbol + i += 1 + + i = 0 + while i < len(lines): + if not lines[i].startswith('Relocation section'): + i += 1 continue - if state == 'symbol': - try: - parts = line[17:].split() - if len(parts) == 3: - sectionname, size, name = parts - elif len(parts) == 4 and parts[2] == '.hidden': - sectionname, size, hidden, name = parts - else: - continue - symbol = Symbol() - symbol.size = int(size, 16) - symbol.offset = int(line[:8], 16) - symbol.name = name - symbol.section = sectionmap.get(sectionname) - symbols[name] = symbol - except ValueError: - pass + rel_name = lines[i][20:lines[i].index("'", 20)] + i += 2 + if rel_name.startswith('.rel.debug_'): + # Skip debugging sections (to reduce parsing time) continue - if state == 'reloc': - try: - off, type, symbolname = line.split() - reloc = Reloc() - reloc.offset = int(off, 16) - reloc.type = type - reloc.symbolname = symbolname - reloc.symbol = symbols.get(symbolname) - if reloc.symbol is None: - # Some binutils (2.20.1) give section name instead - # of a symbol - create a dummy symbol. - reloc.symbol = symbol = Symbol() - symbol.size = 0 - symbol.offset = 0 - symbol.name = symbolname - symbol.section = sectionmap.get(symbolname) - symbols[symbolname] = symbol - relocsection.relocs.append(reloc) - except ValueError: - pass + assert rel_name.startswith('.rel') + relocated = sectionmap[rel_name[4:]] + while lines[i][0].isalnum(): + line = lines[i] + off, _, typ, _, symbolname = line.split() + reloc = Reloc() + reloc.offset = int(off, 16) + reloc.type = typ + reloc.symbolname = symbolname + reloc.symbol = symbols.get(symbolname) + assert reloc.symbol is not None + relocated.relocs.append(reloc) + i += 1 + return sections, symbols
# Parser for constants in simple C header files. @@ -653,9 +646,9 @@ def main(): infile32flat = open(in32flat, 'r')
# infoX = (sections, symbols) - info16 = parseObjDump(infile16, '16') - info32seg = parseObjDump(infile32seg, '32seg') - info32flat = parseObjDump(infile32flat, '32flat') + info16 = parseReadElf(infile16, '16') + info32seg = parseReadElf(infile32seg, '32seg') + info32flat = parseReadElf(infile32flat, '32flat')
# Read kconfig config file config = scanconfig(cfgfile)
On Wed, Apr 01, 2020 at 06:12:09PM -0700, Fangrui Song wrote:
objdump -h relies heavily on BFD internals and the BFD flags are difficult to emulate in llvm-objdump. llvm-objdump -h has a different output (https://reviews.llvm.org/D57146)
Switch to readelf, which is generally better than objdump when dumping ELF section/symbol information.
Signed-off-by: Fangrui Song maskray@google.com
At a high-level, this change makes sense to me. What Linux distros/versions have you tested this on? (Or, do you have reason to believe the readelf output has been stable for many years?)
-Kevin
On 2020-04-06, Kevin O'Connor wrote:
On Wed, Apr 01, 2020 at 06:12:09PM -0700, Fangrui Song wrote:
objdump -h relies heavily on BFD internals and the BFD flags are difficult to emulate in llvm-objdump. llvm-objdump -h has a different output (https://reviews.llvm.org/D57146)
Switch to readelf, which is generally better than objdump when dumping ELF section/symbol information.
Signed-off-by: Fangrui Song maskray@google.com
At a high-level, this change makes sense to me. What Linux distros/versions have you tested this on? (Or, do you have reason to believe the readelf output has been stable for many years?)
-Kevin
Tested on Debian testing.
readelf -S -r -s output is stable. It just prints raw ELF fields. objdump -t -h -r output is not. The output goes through various BFD abstractions. "CONTENTS, ALLOC, LOAD, READONLY, DATA" -> these are really BFD_* constants. The genuine ELF values are SHT_PROGBITS, SHF_ALLOC, non-SHF_WRITE, etc.
The LMA column is not really recorded in ELF. objdump internally does some translation like sh_addr-p_vaddr+p_paddr.
I made llvm-objdump's output closer to objdump as long as it is reasonable. https://github.com/llvm/llvm-project/commits/master/llvm/tools/llvm-objdump/... These BFD flags may be difficult to implement and may not provide enough value. I actually claim myself as a maintainer of llvm-{readobj,objdump,...} and have reported many binutils issues as well.
llvm-readelf, on the other side, has great compatibility with GNU readelf.
-
Fangrui Song -
Kevin O'Connor