RISC-V linker relaxation in lld

On 2022-07-07, I added a RISC-V linker relaxation framework in ld.lld and implemented R_RISCV_ALIGN/R_RISCV_CALL/R_RISCV_CALL_PLT relaxation. The changes will be included in the next llvm-project release 15.0.0. This post describes the implementation.

See The dark side of RISC-V linker relaxation for more information about RISC-V linker relaxation.

Problems

ld.lld performs these steps (simplified):

  • Parse command line options
  • Find and scan input files (.o, .so, .a), interleaved with symbol resolution
  • Call LLVM LTO to get ELF object files
  • Global transforms (section based garbage collection, identical code folding, etc)
  • Create synthetic sections
  • Map input sections and synthetic (linker-generated) sections into output sections
  • Scan relocations
  • Finalize synthetic sections
  • Layout (addresses, thunks, SHT_RELR, symbol assignments)
  • Assign file offsets
  • Write header and sections

We need to find a place to insert the relaxation pass.

Relocation scanning

Relocation scanning makes dynamic relocation decisions and determines the sizes of .got, .got.plt, .plt, .rela.dyn, and .relr.dyn sections. Their address and size changes will affect subsequent sections and sections using certain linker script features. The one-pass relocation scanning scheme is tied to the whole ld.lld design and is difficult to change. Relocation scanning takes time and we want to perform it only when necessary.

Linker relaxation may make input sections smaller and nullify the current section layout. For a call code sequence, if the size decrease makes the destination closer enough to the relocated location, we need to rewrite the code sequence into a shorter form. This change may have a cascading effect and trigger further relaxation. E.g. in the following diagram consisting of three input sections, if call x (a pseudo instruction which expands to 8 bytes) in section B is shortened, B's size will decrease and call c in A may become a new candidate for relaxation.

1
A[... call c; ...] -- B[... call x; ...] -- C[c: ...]

Symbol values

The changed section layout may change symbol values. It is rare but an output section address can use a symbol value. In the following linker script example, the size change of .mid will change the address of .high.

1
2
3
4
5
SECTIONS {
.mid 0x10800 : { mid_start = .; *(.mid); mid_end = .; }
.high 0x110000+(mid_end-mid_start) : { *(.high) }
.high2 0x210000+SIZEOF(.mid) : { *(.high2) }
}

Design

Linker relaxation has to an iterative process. Since it interacts with address-dependent sections and symbol assignments, the main idea is to add linker relaxation to the layout phase. So we get:

  • Scan relocations
  • Finalize synthetic sections
  • Layout (relaxation, addresses, thunks, SHT_RELR, symbol assignments)
  • Assign file offsets

"Scan relocations" is kept. We add another relocation scanning pass to process all relaxable relocations. This pass computes multiple results.

  • for each relocated location, the replacement relocation type, the rewritten code sequence, and the number of bytes to delete
  • the size of each input code section
  • st_value and st_size for each symbol defined relative to the section

The results are used by script->assignAddresses() to compute the next layout: section addresses and symbol values. We repeat the process until the results converge.

In some uncommon cases an input section may expand in a later iteration. If we choose to shrink sections at the end of one iteration, the expansion will be difficult to handle. My idea is that the section size shrink and code sequence rewrites need to be postponed after the iteration fixed point is reached.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
template <class ELFT> void Writer<ELFT>::finalizeAddressDependentContent() {
...
uint32_t pass = 0;
for (;;) {
Create thunks or call relaxOnce;
++pass;

Report "not converged" if pass is too large;

Update address-dependent sections;
Assign addresses to sections and symbols;
}
if (!config->relocatable && config->emachine == EM_RISCV)
riscvFinalizeRelax(pass);
...
}

Two function calls are added to finalizeAddressDependentContent: relaxOnce and riscvFinalizeRelax. The RISC-V port implements relaxOnce which calls relax on all input code sections.

1
2
3
4
5
6
7
8
9
10
11
bool RISCV::relaxOnce(int pass) const {
...
bool changed = false;
for (OutputSection *osec : outputSections) {
if (!(osec->flags & SHF_EXECINSTR))
continue;
for (InputSection *sec : getInputSections(*osec, storage))
changed |= relax(*sec);
}
return changed;
}
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
static bool relax(InputSection &sec) {
Restore original st_value for symbols relative to this section.

std::fill_n(aux.relocTypes.get(), sec.relocations.size(), R_RISCV_NONE);
aux.writes.clear();
for (auto [i, r] : llvm::enumerate(sec.relocations)) {
const uint64_t loc = secAddr + r.offset - delta;
uint32_t &cur = aux.relocDeltas[i], remove = 0;
switch (r.type) {
case R_RISCV_ALIGN: {
remove = the number of bytes to delete;
break;
}
case R_RISCV_CALL:
case R_RISCV_CALL_PLT:
if (i + 1 != sec.relocations.size() &&
sec.relocations[i + 1].type == R_RISCV_RELAX)
relaxCall(sec, i, loc, r, remove);
break;
// Other relaxable relocation types
}

Update symbol st_value/st_size according to symbol anchors;

delta += remove;
if (delta != cur) {
cur = delta;
changed = true;
}
}

Update trailing symbol anchors;

sec.bytesDropped = delta;
return changed;
}

relax iterates over non-resolved relocations for this input section and sets remove to the number of bytes to delete. delta is the accumulated number of bytes to delete. It is stored in aux.relocDeltas[i] for processing in riscvFinalizeRelax.

Symbol anchors

Updating st_value and st_size for each symbol defined relative to the section uses a neat technique.

1
2
3
4
  ...
a:
.balign 16 # R_RISCV_ALIGN(r_addend=12)
b:

In this example, there is an R_RISCV_ALIGN relocation at the .balign location. Its offset equals of symbol a's st_value. If some bytes preceding a are deleted, a's st_value needs to be decreased by that number of bytes. b has a larger st_value and its st_value needs to additionally take into account the R_RISCV_ALIGN relaxation.

To compute all st_value of symbols relative to the current input section, we maintain two sorted lists: (a) relaxable relocations (b) st_value. For each symbol, find the relocation with the largest r_offset which is smaller than the symbol's st_value, then decrease st_value by r_offset. The interleave of st_value values and r_offset values is like the merge function of merge sort.

st_size can be computed similarly. Instead we interleave st_value+st_size values with r_offset values. After the final st_value+st_size is determined, decrease the sum by the final st_value to compute the final st_size. In the implementation, I just place all initial st_value and st_value+st_size values in one sorted list. Both are indicated by a SymbolAnchor object.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
struct SymbolAnchor {
uint64_t offset;
Defined *d;
bool end; // true for the anchor of st_value+st_size
};

if (remove) {
for (; sa.size() && sa[0].offset <= r.offset; sa = sa.slice(1)) {
if (sa[0].end)
sa[0].d->size = sa[0].offset - delta - sa[0].d->value;
else
sa[0].d->value -= delta;
}
}

Since we use the decrement amount (sa[0].d->value -= delta;), when starting the next iteration, we need to restore the original st_value.

Finalize relaxation

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
void elf::riscvFinalizeRelax(int passes) {
...
for (OutputSection *osec : outputSections) {
if (!(osec->flags & SHF_EXECINSTR))
continue;
for (InputSection *sec : getInputSections(*osec, storage)) {
RISCVRelaxAux &aux = *sec->relaxAux;
if (!aux.relocDeltas)
continue;

Allocate space for the new section content to `p`;
sec->rawData = makeArrayRef(p, newSize);

// Update section content: remove NOPs for R_RISCV_ALIGN and rewrite
// instructions for relaxed relocations.
for (size_t i = 0, e = rels.size(); i != e; ++i) {
uint32_t remove = aux.relocDeltas[i] - delta;
delta = aux.relocDeltas[i];
if (remove == 0)
continue;

// Copy from last location to the current relocated location.
const Relocation &r = rels[i];
uint64_t size = r.offset - offset;
memcpy(p, old.data() + offset, size);
p += size;
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
        // For R_RISCV_ALIGN, we will place `offset` in a location (among NOPs)
// to satisfy the alignment requirement. If `remove` is a multiple of 4,
// it is as if we have skipped some NOPs. Otherwise we are in the middle
// of a 4-byte NOP, and we need to rewrite the NOP sequence.
int64_t skip = 0;
if (r.type == R_RISCV_ALIGN) {
if (remove % 4 != 0) {
skip = r.addend - remove;
Rewrite `skip` bytes with nop and an optional trailing c.nop;
}
} else if (RelType newType = aux.relocTypes[i]) {
Rewrite code sequence;
}

p += skip;
offset = r.offset + skip + remove;
}
memcpy(p, old.data() + offset, old.size() - offset);

Subtract the previous relocDeltas value from the relocation offset.
For a pair of R_RISCV_CALL/R_RISCV_RELAX with the same offset, decrease
their r_offset by the same delta.
}
}
}

For each input code section, We iterate over its non-resolved relocations. For an R_RISCV_ALIGN associated with some bytes to delete, we copy all content from the previous location to r_offset, then skip some bytes for the next copy.

1
2
3
4
...         # Copy all content from the previous location to r_offset
.balign 8 # R_RISCV_ALIGN(r_addend=6)
# A prefix of the NOPs may be skipped for the next memcpy
addi a0, a0, 1

Say we need to delete 2 bytes. If we use [] to indicate the copied bytes, the current and the next copy patterns will look like:

1
2
3
4
5
old: ...]  NOP  NOP [NOP  NOP  NOP  NOP  ADDI ADDI ADDI ADDI ...]
old: next copy

new: ...] [NOP NOP NOP NOP ADDI ADDI ADDI ADDI ...]
new:

Let's check a call relaxation case. The call pseudo instruction expands to a pair of auipc and jalr.

1
call dest@plt  # R_RISCV_CALL_PLT, R_RISCV_RELAX

If auipc+jalr can be relaxed to a 4-byte jal, we ignore auipc, replace jalr with jal, and increment p and offset so that next memcpy will start copying from the first byte after jalr. The rewritten instruction starts at the first byte indicated by skip=4.

1
2
3
4
old: ...] AUIPC AUIPC AUIPC AUIPC JALR JALR JALR JALR [.........]
remove=4 skip=4 next copy

new: ...] JAL JAL JAL JAL [.........]

Here is a demonstration for a tail pseudo instruction which is relaxed to c.j.

1
tail dest@plt  # R_RISCV_CALL_PLT, R_RISCV_RELAX
1
2
3
4
old: ...] AUIPC AUIPC AUIPC AUIPC JALR JALR JALR JALR [.........]
remove=6 skip=2 next copy

new: ...] C.J C.J [.........]

Relaxable code sequences

Alignment relaxation

With 3 values we can compute the address of the relocated location: secAddr + r.offset - delta. delta is the asscumulated number of bytes to delete. It is subtracted from the original r_offset value.

The alignment is PowerOf2Ceil(r.addend + 2). The expected location after alignment is (loc + align - 1) & -align and therefore loc + r.addend - ((loc + align - 1) & -align) is the number of bytes to delete.

Call relaxation

These following pseudo instructions are available to call subroutines.

1
2
3
4
5
6
7
8
9
10
11
call a@plt          # @plt can be omitted. In ld.lld R_RISCV_CALL/R_RISCV_CALL_PLT are indistinguishable
# auipc ra, 0 # R_RISCV_CALL_PLT(a), R_RISCV_RELAX
# jalr ra, 0(ra)

tail a@plt
# auipc t1, 0 # R_RISCV_CALL_PLT(a), R_RISCV_RELAX
# jalr zero, 0(t1)

jump a, t0
# auipc t0, 0 # R_RISCV_CALL(a), R_RISCV_RELAX
# jalr zero, 0(t0)

Each expands to a pair of auipc and jalr.

  • call: ra is both a scratch register and the destination register
  • tail: t1 is a scratch register. x0 is the destination register
  • jump: the scratch register is specified. x0 is the destination register

The two instructions can be relaxed to one alternative instruction. There are 3 choices:

  • c.j: RVC, the destination register is x0, and the displacement is representable as an int12
  • c.jal: RV32C, the destination register is ra, and the displacement is representable as an int12
  • jal: the displacement is representable as an int21

The first two need to delete 6 bytes and rewrite 2 bytes while the third needs to delete 4 bytes and rewrite 4 bytes.

Local-exec TLS relaxation

See All about thread-local storage for more information about TLS.

Computing the address or storing a value into a TLS variable takes 3 instructions. If st_value(x) < 2048 (i.e. hi20(x) == 0), one instruction suffices.

1
2
3
4
5
6
7
lui rd, %tprel_hi(x)           # R_RISCV_TPREL_HI20, R_RISCV_RELAX
add rd, rd, tp, %tprel_add(x) # R_RISCV_TPREL_ADD, R_RISCV_RELAX
addi rd, rd, %tprel_lo(x) # R_RISCV_TPREL_LO12_I, R_RISCV_RELAX

=>

addi rd, tp, st_value(x)
1
2
3
4
5
6
7
lui rd, %tprel_hi(x)           # R_RISCV_TPREL_HI20, R_RISCV_RELAX
add rd, rd, tp, %tprel_add(x) # R_RISCV_TPREL_ADD, R_RISCV_RELAX
sw rs, st_value(x)(rd) # R_RISCV_TPREL_LO12_S, R_RISCV_RELAX

=>

sw rs, st_value(x)(rd)

Pending patch: https://reviews.llvm.org/D129425

lui relaxation

If ld.lld implements this, most absolute and PC-relative relocations need bookkeeping that they can candidates for relaxation. This may add quite a bit of overhead.

Relaxation against the Global Pointer

See "Relaxing Against the Global Pointer" on https://www.sifive.com/blog/all-aboard-part-3-linker-relaxation-in-riscv-toolchain.

I am of the opinion that this choice is short-sighted, so I created https://github.com/riscv-non-isa/riscv-elf-psabi-doc/issues/298 which was soon closed. However, I don't receive strong arguments supporting this scheme. I wish that interested users help me by making some measurement.