ixdiff

ixdiff compares the assembly of two binaries to show how compiler or build-flag changes affected the generated code: which functions changed, by how much, and what the instruction-level differences are.

Install

go install github.com/loov/ixdiff@latest

Prebuilt binaries for linux, macOS, and Windows are attached to releases.

Usage

Summary of everything that changed:

$ ixdiff tsgo.v0 tsgo.v3
functions: 5044 identical, 232 changed (+17229 relocations), 5 added, 8 removed
total text size delta: +6000 bytes

instruction delta by opcode:
    +362 MOV
    +244 LDR
     +93 BL
    ...

package delta:
       bytes    insts  changed  added  removed  package
       +2784     +699       15      0        2  syscall
        +992     +247        2      2        0  fmt
    ...

top 100 by size delta:
       bytes    insts state     function
       +1216     +304 changed   syscall.forkExec
        +608     +152 added     fmt.wrapErrorf
    ...

"Relocations" counts functions whose bytes differ only because code or data moved; they are not real changes and are excluded from all statistics.

Diff one function (--fn accepts substrings and suggests near misses; repeatable):

$ ixdiff --fn net/url.parseHost tsgo.v0 tsgo.v3
--- net/url.parseHost (1888 bytes)
+++ net/url.parseHost (1904 bytes)
@@ -10046b65c +10046c8ac @@
 10046b65c: ADD  $240, RSP, R2
 10046b660: ORR  $1, ZR, R3
 10046b664: MOVD R3, R4
+10046c8b8: ORR  $14, ZR, R5
 10046b668: CALL fmt.errorf(SB)
 10046b66c: CBNZ R0, L22

Every line carries the address of the instruction it came from, so lines can be cross-referenced with objdump or profiler output. Added and removed functions print their full listing.

The same diff as two columns, old on the left and new on the right:

$ ixdiff --side-by-side --fn net/url.parseHost tsgo.v0 tsgo.v3
--- net/url.parseHost (1888 bytes)
+++ net/url.parseHost (1904 bytes)
@@ -10046b65c +10046c8ac @@
10046b65c: ADD  $240, RSP, R2    10046c8ac: ADD  $240, RSP, R2
10046b660: ORR  $1, ZR, R3       10046c8b0: ORR  $1, ZR, R3
10046b664: MOVD R3, R4           10046c8b4: MOVD R3, R4
                               > 10046c8b8: ORR  $14, ZR, R5
10046b668: CALL fmt.errorf(SB)   10046c8bc: CALL fmt.errorf(SB)
10046b66c: CBNZ R0, L22          10046c8c0: CBNZ R0, L22

Other flags:

Flag	Effect
--top N, --sort size|insts|name	ranking table length and order
--filter net/url, --filter ~^runtime\.	scope the summary by substring or regexp (repeatable)
--state changed|added|removed	limit tables to functions in this state (repeatable)
--all	follow the summary with a diff of every ranked function
--side-by-side	two-column diff view
--blocks	match basic blocks before diffing, tolerating block reordering
--mask-sp	ignore stack-offset shifts caused by frame-size changes
--json	machine-readable output
--color auto|always|never	diff coloring (auto respects NO_COLOR)
--version	print the module version and VCS revision

How it works

1. Function ranges come from symbol tables plus the Go pclntab; pairs are matched by name, with renamed functions (closure renumbering, generic respecialization, near-identical bodies) re-paired among the added/removed leftovers.
2. Byte-identical pairs are skipped outright. Pairs whose bytes differ only in relocation fields (verified against the symbol tables, so a retargeted call still counts as a change) are classified without disassembly.
3. The rest are disassembled and normalized: call targets are symbolized, intra-function branches become labels derived from an instruction alignment (so inserting code doesn't renumber every label), and address-dependent operands — IP/ADRP/AUIPC-relative data references, ppc64 ADDIS/ADD address-materialization pairs, address-valued immediates — are masked.
4. Normalized instructions are diffed per function; opcode histograms aggregate into the summary.

What is deliberately not masked: real immediates, register allocation, instruction selection, and (by default) stack offsets — those are the differences an optimization comparison is looking for.

Limitations

• Instruction-level masking exists for amd64, arm64, arm (32-bit), 386, riscv64, loong64, s390x, ppc64/ppc64le, and wasm only; other architectures (notably the mips family, which has no pure-Go disassembler) are unsupported. On s390x, pc-relative data references (larl and friends) are always masked rather than resolved to symbol names, and the relocation-only fast path does not apply.
• Non-Go binaries work only as well as their symbol tables; duplicate static symbols (cgo) keep the last definition.
• The pclntab header magic list needs an update when a new Go release revs the format.

Development

make test   # go test -race ./...
make lint   # go vet + golangci-lint (includes staticcheck)
make check  # both