[lld][LoongArch] GOT indirection to PC relative optimization #123743
+76
−4
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
|
@llvm/pr-subscribers-backend-loongarch @llvm/pr-subscribers-lld-elf Author: Zhaoxin Yang (ylzsx) ChangesIn LoongArch, this optimization is only supported when relaxation is enabled.
If the original code sequence can be relaxed into a single instruction FIXME: Althouth the optimization has been performed, the GOT entries still exists, similarly to AArch64. Eliminating the entries may be require additional marking in the common code. Full diff: https://github.com/llvm/llvm-project/pull/123743.diff 2 Files Affected:
diff --git a/lld/ELF/Arch/LoongArch.cpp b/lld/ELF/Arch/LoongArch.cpp
index 5f49b23e8ffb1a..6ae45e109a6dec 100644
--- a/lld/ELF/Arch/LoongArch.cpp
+++ b/lld/ELF/Arch/LoongArch.cpp
@@ -47,6 +47,8 @@ class LoongArch final : public TargetInfo {
void tlsIeToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const;
void tlsdescToIe(uint8_t *loc, const Relocation &rel, uint64_t val) const;
void tlsdescToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const;
+ bool tryGotToPCRel(uint8_t *loc, const Relocation &rHi20,
+ const Relocation &rLo12, uint64_t secAddr) const;
};
} // end anonymous namespace
@@ -1150,6 +1152,58 @@ void LoongArch::tlsdescToLe(uint8_t *loc, const Relocation &rel,
}
}
+// Try GOT indirection to PC relative optimization when relaxation is enabled.
+// From:
+// * pcalau12i $a0, %got_pc_hi20(sym_got)
+// * ld.w/d $a0, $a0, %got_pc_lo12(sym_got)
+// To:
+// * pcalau12i $a0, %pc_hi20(sym)
+// * addi.w/d $a0, $a0, %pc_lo12(sym)
+//
+// FIXME: Althouth the optimization has been performed, the GOT entries still
+// exists, similarly to AArch64. Eliminating the entries may be require
+// additional marking in the common code.
+bool LoongArch::tryGotToPCRel(uint8_t *loc, const Relocation &rHi20,
+ const Relocation &rLo12, uint64_t secAddr) const {
+ if (!rHi20.sym->isDefined() || rHi20.sym->isPreemptible ||
+ rHi20.sym->isGnuIFunc() ||
+ (ctx.arg.isPic && !cast<Defined>(*rHi20.sym).section))
+ return false;
+
+ Symbol &sym = *rHi20.sym;
+ uint64_t symLocal = sym.getVA(ctx) + rHi20.addend;
+ // Check if the address difference is within +/-2GB range.
+ // For simplicity, the range mentioned here is an approximate estimate and is
+ // not fully equivalent to the entire region that PC-relative addressing can
+ // cover.
+ int64_t pageOffset =
+ getLoongArchPage(symLocal) - getLoongArchPage(secAddr + rHi20.offset);
+ if (!isInt<20>(pageOffset >> 12))
+ return false;
+
+ Relocation newRHi20 = {RE_LOONGARCH_PAGE_PC, R_LARCH_PCALA_HI20, rHi20.offset,
+ rHi20.addend, &sym};
+ Relocation newRLo12 = {R_ABS, R_LARCH_PCALA_LO12, rLo12.offset, rLo12.addend,
+ &sym};
+
+ const uint32_t currInsn = read32le(loc);
+ const uint32_t nextInsn = read32le(loc + 4);
+ // Check if use the same register.
+ if (getD5(currInsn) != getJ5(nextInsn) || getJ5(nextInsn) != getD5(nextInsn))
+ return false;
+
+ uint64_t pageDelta =
+ getLoongArchPageDelta(symLocal, secAddr + rHi20.offset, rHi20.type);
+ // pcalau12i $a0, %pc_hi20
+ write32le(loc, insn(PCALAU12I, getD5(currInsn), 0, 0));
+ relocate(loc, newRHi20, pageDelta);
+ // addi.w/d $a0, $a0, %pc_lo12
+ write32le(loc + 4, insn(ctx.arg.is64 ? ADDI_D : ADDI_W, getD5(nextInsn),
+ getJ5(nextInsn), 0));
+ relocate(loc + 4, newRLo12, SignExtend64(symLocal, 64));
+ return true;
+}
+
// During TLSDESC GD_TO_IE, the converted code sequence always includes an
// instruction related to the Lo12 relocation (ld.[wd]). To obtain correct val
// in `getRelocTargetVA`, expr of this instruction should be adjusted to
@@ -1259,6 +1313,22 @@ void LoongArch::relocateAlloc(InputSectionBase &sec, uint8_t *buf) const {
tlsdescToLe(loc, rel, val);
}
continue;
+ case RE_LOONGARCH_GOT_PAGE_PC:
+ // In LoongArch, we try GOT indirection to PC relative optimization only
+ // when relaxation is enabled. This approach avoids determining whether
+ // relocation types are paired and whether the destination register of
+ // pcalau12i is only used by the immediately following instruction.
+ // Moreover, if the original code sequence can be relaxed to a single
+ // instruction `pcaddi`, the first instruction will be removed and it will
+ // not reach here.
+ if (isPairRelaxable(relocs, i) && rel.type == R_LARCH_GOT_PC_HI20 &&
+ relocs[i + 2].type == R_LARCH_GOT_PC_LO12 &&
+ tryGotToPCRel(loc, rel, relocs[i + 2], secAddr)) {
+ i = i + 3; // skip relocations R_LARCH_RELAX, R_LARCH_GOT_PC_LO12,
+ // R_LARCH_RELAX
+ continue;
+ }
+ break;
default:
break;
}
diff --git a/lld/test/ELF/loongarch-relax-pc-hi20-lo12.s b/lld/test/ELF/loongarch-relax-pc-hi20-lo12.s
index 760fe77d774e30..ae3b29e14fb3c1 100644
--- a/lld/test/ELF/loongarch-relax-pc-hi20-lo12.s
+++ b/lld/test/ELF/loongarch-relax-pc-hi20-lo12.s
@@ -30,24 +30,26 @@
## offset = 0x410000 - 0x10000: 0x400 pages, page offset 0
# NORELAX32-NEXT: 10000: pcalau12i $a0, 1024
# NORELAX32-NEXT: addi.w $a0, $a0, 0
+## Not relaxation, convertion to PCRel.
# NORELAX32-NEXT: pcalau12i $a0, 1024
-# NORELAX32-NEXT: ld.w $a0, $a0, 4
+# NORELAX32-NEXT: addi.w $a0, $a0, 0
# NORELAX32-NEXT: pcalau12i $a0, 1024
# NORELAX32-NEXT: addi.w $a0, $a0, 0
# NORELAX32-NEXT: pcalau12i $a0, 1024
-# NORELAX32-NEXT: ld.w $a0, $a0, 4
+# NORELAX32-NEXT: addi.w $a0, $a0, 0
# NORELAX64-LABEL: <_start>:
## offset exceed range of pcaddi
## offset = 0x410000 - 0x10000: 0x400 pages, page offset 0
# NORELAX64-NEXT: 10000: pcalau12i $a0, 1024
# NORELAX64-NEXT: addi.d $a0, $a0, 0
+## Not relaxation, convertion to PCRel.
# NORELAX64-NEXT: pcalau12i $a0, 1024
-# NORELAX64-NEXT: ld.d $a0, $a0, 8
+# NORELAX64-NEXT: addi.d $a0, $a0, 0
# NORELAX64-NEXT: pcalau12i $a0, 1024
# NORELAX64-NEXT: addi.d $a0, $a0, 0
# NORELAX64-NEXT: pcalau12i $a0, 1024
-# NORELAX64-NEXT: ld.d $a0, $a0, 8
+# NORELAX64-NEXT: addi.d $a0, $a0, 0
.section .text
.global _start
|
|
@llvm/pr-subscribers-lld Author: Zhaoxin Yang (ylzsx) ChangesIn LoongArch, this optimization is only supported when relaxation is enabled.
If the original code sequence can be relaxed into a single instruction FIXME: Althouth the optimization has been performed, the GOT entries still exists, similarly to AArch64. Eliminating the entries may be require additional marking in the common code. Full diff: https://github.com/llvm/llvm-project/pull/123743.diff 2 Files Affected:
diff --git a/lld/ELF/Arch/LoongArch.cpp b/lld/ELF/Arch/LoongArch.cpp
index 5f49b23e8ffb1a..6ae45e109a6dec 100644
--- a/lld/ELF/Arch/LoongArch.cpp
+++ b/lld/ELF/Arch/LoongArch.cpp
@@ -47,6 +47,8 @@ class LoongArch final : public TargetInfo {
void tlsIeToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const;
void tlsdescToIe(uint8_t *loc, const Relocation &rel, uint64_t val) const;
void tlsdescToLe(uint8_t *loc, const Relocation &rel, uint64_t val) const;
+ bool tryGotToPCRel(uint8_t *loc, const Relocation &rHi20,
+ const Relocation &rLo12, uint64_t secAddr) const;
};
} // end anonymous namespace
@@ -1150,6 +1152,58 @@ void LoongArch::tlsdescToLe(uint8_t *loc, const Relocation &rel,
}
}
+// Try GOT indirection to PC relative optimization when relaxation is enabled.
+// From:
+// * pcalau12i $a0, %got_pc_hi20(sym_got)
+// * ld.w/d $a0, $a0, %got_pc_lo12(sym_got)
+// To:
+// * pcalau12i $a0, %pc_hi20(sym)
+// * addi.w/d $a0, $a0, %pc_lo12(sym)
+//
+// FIXME: Althouth the optimization has been performed, the GOT entries still
+// exists, similarly to AArch64. Eliminating the entries may be require
+// additional marking in the common code.
+bool LoongArch::tryGotToPCRel(uint8_t *loc, const Relocation &rHi20,
+ const Relocation &rLo12, uint64_t secAddr) const {
+ if (!rHi20.sym->isDefined() || rHi20.sym->isPreemptible ||
+ rHi20.sym->isGnuIFunc() ||
+ (ctx.arg.isPic && !cast<Defined>(*rHi20.sym).section))
+ return false;
+
+ Symbol &sym = *rHi20.sym;
+ uint64_t symLocal = sym.getVA(ctx) + rHi20.addend;
+ // Check if the address difference is within +/-2GB range.
+ // For simplicity, the range mentioned here is an approximate estimate and is
+ // not fully equivalent to the entire region that PC-relative addressing can
+ // cover.
+ int64_t pageOffset =
+ getLoongArchPage(symLocal) - getLoongArchPage(secAddr + rHi20.offset);
+ if (!isInt<20>(pageOffset >> 12))
+ return false;
+
+ Relocation newRHi20 = {RE_LOONGARCH_PAGE_PC, R_LARCH_PCALA_HI20, rHi20.offset,
+ rHi20.addend, &sym};
+ Relocation newRLo12 = {R_ABS, R_LARCH_PCALA_LO12, rLo12.offset, rLo12.addend,
+ &sym};
+
+ const uint32_t currInsn = read32le(loc);
+ const uint32_t nextInsn = read32le(loc + 4);
+ // Check if use the same register.
+ if (getD5(currInsn) != getJ5(nextInsn) || getJ5(nextInsn) != getD5(nextInsn))
+ return false;
+
+ uint64_t pageDelta =
+ getLoongArchPageDelta(symLocal, secAddr + rHi20.offset, rHi20.type);
+ // pcalau12i $a0, %pc_hi20
+ write32le(loc, insn(PCALAU12I, getD5(currInsn), 0, 0));
+ relocate(loc, newRHi20, pageDelta);
+ // addi.w/d $a0, $a0, %pc_lo12
+ write32le(loc + 4, insn(ctx.arg.is64 ? ADDI_D : ADDI_W, getD5(nextInsn),
+ getJ5(nextInsn), 0));
+ relocate(loc + 4, newRLo12, SignExtend64(symLocal, 64));
+ return true;
+}
+
// During TLSDESC GD_TO_IE, the converted code sequence always includes an
// instruction related to the Lo12 relocation (ld.[wd]). To obtain correct val
// in `getRelocTargetVA`, expr of this instruction should be adjusted to
@@ -1259,6 +1313,22 @@ void LoongArch::relocateAlloc(InputSectionBase &sec, uint8_t *buf) const {
tlsdescToLe(loc, rel, val);
}
continue;
+ case RE_LOONGARCH_GOT_PAGE_PC:
+ // In LoongArch, we try GOT indirection to PC relative optimization only
+ // when relaxation is enabled. This approach avoids determining whether
+ // relocation types are paired and whether the destination register of
+ // pcalau12i is only used by the immediately following instruction.
+ // Moreover, if the original code sequence can be relaxed to a single
+ // instruction `pcaddi`, the first instruction will be removed and it will
+ // not reach here.
+ if (isPairRelaxable(relocs, i) && rel.type == R_LARCH_GOT_PC_HI20 &&
+ relocs[i + 2].type == R_LARCH_GOT_PC_LO12 &&
+ tryGotToPCRel(loc, rel, relocs[i + 2], secAddr)) {
+ i = i + 3; // skip relocations R_LARCH_RELAX, R_LARCH_GOT_PC_LO12,
+ // R_LARCH_RELAX
+ continue;
+ }
+ break;
default:
break;
}
diff --git a/lld/test/ELF/loongarch-relax-pc-hi20-lo12.s b/lld/test/ELF/loongarch-relax-pc-hi20-lo12.s
index 760fe77d774e30..ae3b29e14fb3c1 100644
--- a/lld/test/ELF/loongarch-relax-pc-hi20-lo12.s
+++ b/lld/test/ELF/loongarch-relax-pc-hi20-lo12.s
@@ -30,24 +30,26 @@
## offset = 0x410000 - 0x10000: 0x400 pages, page offset 0
# NORELAX32-NEXT: 10000: pcalau12i $a0, 1024
# NORELAX32-NEXT: addi.w $a0, $a0, 0
+## Not relaxation, convertion to PCRel.
# NORELAX32-NEXT: pcalau12i $a0, 1024
-# NORELAX32-NEXT: ld.w $a0, $a0, 4
+# NORELAX32-NEXT: addi.w $a0, $a0, 0
# NORELAX32-NEXT: pcalau12i $a0, 1024
# NORELAX32-NEXT: addi.w $a0, $a0, 0
# NORELAX32-NEXT: pcalau12i $a0, 1024
-# NORELAX32-NEXT: ld.w $a0, $a0, 4
+# NORELAX32-NEXT: addi.w $a0, $a0, 0
# NORELAX64-LABEL: <_start>:
## offset exceed range of pcaddi
## offset = 0x410000 - 0x10000: 0x400 pages, page offset 0
# NORELAX64-NEXT: 10000: pcalau12i $a0, 1024
# NORELAX64-NEXT: addi.d $a0, $a0, 0
+## Not relaxation, convertion to PCRel.
# NORELAX64-NEXT: pcalau12i $a0, 1024
-# NORELAX64-NEXT: ld.d $a0, $a0, 8
+# NORELAX64-NEXT: addi.d $a0, $a0, 0
# NORELAX64-NEXT: pcalau12i $a0, 1024
# NORELAX64-NEXT: addi.d $a0, $a0, 0
# NORELAX64-NEXT: pcalau12i $a0, 1024
-# NORELAX64-NEXT: ld.d $a0, $a0, 8
+# NORELAX64-NEXT: addi.d $a0, $a0, 0
.section .text
.global _start
|
In LoongArch, this optimization is only supported when relaxation is enabled. From: * pcalau12i $a0, %got_pc_hi20(sym_got) * ld.w/d $a0, $a0, %got_pc_lo12(sym_got) To: * pcalau12i $a0, %pc_hi20(sym) * addi.w/d $a0, $a0, %pc_lo12(sym) If the original code sequence can be relaxed into a single instruction `pcaddi`, this patch will not be taken (see https://). The implementation related to `got` is split into two locations because the `relax()` function is part of an iteration fixed-point algorithm. We should minimize it to achieve better linker performance. FIXME: Althouth the optimization has been performed, the GOT entries still exists, similarly to AArch64. Eliminating the entries may be require additional marking in the common code.
MaskRay
reviewed
Jan 22, 2025
| // additional marking in the common code. | ||
| bool LoongArch::tryGotToPCRel(uint8_t *loc, const Relocation &rHi20, | ||
| const Relocation &rLo12, uint64_t secAddr) const { | ||
| if (!rHi20.sym->isDefined() || rHi20.sym->isPreemptible || |
There was a problem hiding this comment.
We need symbol tests to test each condition here. aarch64-adrp-ldr-got-symbols.s has an example
There was a problem hiding this comment.
Thanks. I have added this test in a previous patch(#123566).
MaskRay
reviewed
Jan 22, 2025
lld/ELF/Arch/LoongArch.cpp
Outdated
| // * pcalau12i $a0, %pc_hi20(sym) | ||
| // * addi.w/d $a0, $a0, %pc_lo12(sym) | ||
| // | ||
| // FIXME: Althouth the optimization has been performed, the GOT entries still |
There was a problem hiding this comment.
Delete FIXME. While the GOT entries are not eliminated, this might actually be desired to reduce code complexity.
There was a problem hiding this comment.
Thanks, I will change it as follows：
Note: Althouth the optimization has been performed, the GOT entries still exists, similarly to AArch64. Eliminating the entries will increase code complexity.
|
Can you drop the trailing |
ylzsx
changed the title
ylzsx
force-pushed
the
users/ylzsx/r-got-to-pcrel
branch
from
30a9eb5 to
e3fc1d6
Compare
ylzsx
force-pushed
the
users/ylzsx/r-tlsdesc-to-iele-relax
branch
from
e024b7c to
99a1e07
Compare
ylzsx
force-pushed
the
users/ylzsx/r-got-to-pcrel
branch
from
e3fc1d6 to
47d84d9
Compare
|
cc @xen0n |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
In LoongArch, this optimization is only supported when relaxation is enabled.
From:
To:
If the original code sequence can be relaxed into a single instruction
pcaddi, this patch will not be taken (see #123566).The implementation related to
gotis split into two locations because therelax()function is part of an iteration fixed-point algorithm. We should minimize it to achieve better linker performance.Note: Althouth the optimization has been performed, the GOT entries still exists, similarly to AArch64. Eliminating the entries will increase code complexity.