[lld][LoongArch] Support relaxation during TLSDESC GD/LD to IE/LE conversion #123730

ylzsx · 2025-01-21T11:00:17Z

Complement #123715. When relaxation enable, remove redundant NOPs.

Support relaxation optimization for two types of code sequences. ``` From: pcalau12i $a0, %pc_hi20(sym) R_LARCH_PCALA_HI20, R_LARCH_RELAX addi.w/d $a0, $a0, %pc_lo12(sym) R_LARCH_PCALA_LO12, R_LARCH_RELAX To: pcaddi $a0, %pc_lo12(sym) R_LARCH_PCREL20_S2 From: pcalau12i $a0, %got_pc_hi20(sym_got) R_LARCH_GOT_PC_HI20, R_LARCH_RELAX ld.w/d $a0, $a0, %got_pc_lo12(sym_got) R_LARCH_GOT_PC_LO12, R_LARCH_RELAX To: pcaddi $a0, %got_pc_hi20(sym_got) R_LARCH_PCREL20_S2 ```

Similar to aarch64-adrp-ldr-got-symbols.s.

Dependency on #123017

Instructions with relocation `R_LARCH_CALL36` may be relax as follows: ``` From: pcaddu18i $dest, %call36(foo) R_LARCH_CALL36, R_LARCH_RELAX jirl $r, $dest, 0 To: b/bl foo # bl if r=$ra, b if r=$zero R_LARCH_B26 ```

In local-exec form, the code sequence is converted as follows: ``` From: lu12i.w $rd, %le_hi20_r(sym) R_LARCH_TLS_LE_HI20_R, R_LARCH_RELAX add.w/d $rd, $rd, $tp, %le_add_r(sym) R_LARCH_TLS_LE_ADD_R, R_LARCH_RELAX addi/ld/st.w/d $rd, $rd, %le_lo12_r(sym) R_LARCH_TLS_LE_LO12_R, R_LARCH_RELAX To: addi/ld/st.w/d $rd, $tp, %le_lo12_r(sym) R_LARCH_TLS_LE_LO12_R ``` In global-dynamic or local-dynamic, the code sequence is converted as follows: ``` From: pcalau12i $a0, %ld_pc_hi20(sym) | %gd_pc_hi20(sym) R_LARCH_TLS_GD_PC_HI20 | R_LARCH_TLS_LD_PC_HI20, R_LARCH_RELAX addi.w/d $a0, $a0, %got_pc_lo12(sym) | %got_pc_lo12(sym) R_LARCH_GOT_PC_LO12, R_LARCH_RELAX To: pcaddi $a0, %got_pc_lo12(sym) | %got_pc_lo12(sym) R_LARCH_TLS_GD_PCREL20_S2 | R_LARCH_TLS_LD_PCREL20_S2 ``` Note: For initial-exec form, since it involves the conversion from IE to LE, we will implement it in a future patch.

…-relocs.s

Original code sequence: * pcalau12i $a0, %desc_pc_hi20(sym_desc) * addi.d $a0, $a0, %desc_pc_lo12(sym_desc) * ld.d $ra, $a0, %desc_ld(sym_desc) * jirl $ra, $ra, %desc_call(sym_desc) Cannot convert to LE/IE, but relax: * pcaddi $a0, %desc_pcrel_20(sym_desc) * ld.d $ra, $a0, %desc_ld(sym_desc) * jirl $ra, $ra, %desc_call(sym_desc) FIXME: The conversion of TLSDESC GD/LD to LE/IE will implement in a future patch.

llvm-objdump -dr cannot print the relocations, however objudmp works well. It also occurs for RISCV.

Original code sequence: * pcalau12i $a0, %ie_pc_hi20(sym) * ld.d $a0, $a0, %ie_pc_lo12(sym) The code sequence converted is as follows: * lu12i.w $a0, %ie_pc_hi20(sym) # le_hi20 != 0, otherwise NOP * ori $a0 $a0, %ie_pc_lo12(sym) FIXME: When relaxation enables, redundant NOP can be removed. This will be implemented in a future patch. Note: In the normal or medium code model, original code sequence with relocations can appear interleaved, because converted code sequence calculates the absolute offset. However, in extreme code model, to identify the current code model, the first four instructions with relocations must appear consecutively.

Complement https://. When relaxation enable, remove redundant NOPs.

Support TLSDESC to initial-exec or local-exec optimizations. Introduce a new hook RE_LOONGARCH_RELAX_TLS_GD_TO_IE_PAGE_PC and use existing R_RELAX_TLS_GD_TO_IE_ABS to support TLSDESC => IE, while use existing R_RELAX_TLS_GD_TO_LE to support TLSDESC => LE. In normal or medium code model, there are two forms of code sequences: * pcalau12i $a0, %desc_pc_hi20(sym_desc) * addi.d $a0, $a0, %desc_pc_lo12(sym_desc) * ld.d $ra, $a0, %desc_ld(sym_desc) * jirl $ra, $ra, %desc_call(sym_desc) ------ * pcaddi $a0, %desc_pcrel_20(sym_desc) * ld.d $ra, $a0, %desc_ld(sym_desc) * jirl $ra, $ra, %desc_call(sym_desc) The code sequence obtained is as follows: * pcalau12i $a0, %ie_pc_hi20(sym_ie) * ld.[wd] $a0, $a0, %ie_pc_lo12(sym_ie) Simplicity, whether tlsdescToIe or tlsdescToLe, we always tend to convert the preceding instructions to NOPs, due to both forms of code sequence (corresponding to relocation combinations: R_LARCH_TLS_DESC_PC_HI20+R_LARCH_TLS_DESC_PC_LO12 and R_LARCH_TLS_DESC_PCREL20_S2) have same process. FIXME: When relaxation enables, redundant NOPs can be removed. It will be implemented in a future patch. Note: All forms of TLSDESC code sequences should not appear interleaved in the normal, medium or extreme code model, which compilers do not generate and lld is unsupported. This is thanks to the guard in PostRASchedulerList.cpp in llvm. ``` Calls are not scheduling boundaries before register allocation, but post-ra we don't gain anything by scheduling across calls since we don't need to worry about register pressure. ```

llvmbot · 2025-01-21T11:00:49Z

@llvm/pr-subscribers-backend-loongarch
@llvm/pr-subscribers-lld-elf

@llvm/pr-subscribers-lld

Author: Zhaoxin Yang (ylzsx)

Changes

Complement #123715. When relaxation enable, remove redundant NOPs.

Full diff: https://github.com/llvm/llvm-project/pull/123730.diff

2 Files Affected:

(modified) lld/ELF/Arch/LoongArch.cpp (+29-3)
(modified) lld/test/ELF/loongarch-relax-tlsdesc.s (+16-29)

diff --git a/lld/ELF/Arch/LoongArch.cpp b/lld/ELF/Arch/LoongArch.cpp
index 37614c3e9615d6..5f49b23e8ffb1a 100644
--- a/lld/ELF/Arch/LoongArch.cpp
+++ b/lld/ELF/Arch/LoongArch.cpp
@@ -965,10 +965,16 @@ static bool relax(Ctx &ctx, InputSection &sec) {
     case R_LARCH_GOT_PC_HI20:
     case R_LARCH_TLS_GD_PC_HI20:
     case R_LARCH_TLS_LD_PC_HI20:
-    case R_LARCH_TLS_DESC_PC_HI20:
       // The overflow check for i+2 will be carried out in isPairRelaxable.
-      if (r.expr != RE_LOONGARCH_RELAX_TLS_GD_TO_IE_PAGE_PC &&
-          r.expr != R_RELAX_TLS_GD_TO_LE && isPairRelaxable(relocs, i))
+      if (isPairRelaxable(relocs, i))
+        relaxPCHi20Lo12(ctx, sec, i, loc, r, relocs[i + 2], remove);
+      break;
+    case R_LARCH_TLS_DESC_PC_HI20:
+      if (r.expr == RE_LOONGARCH_RELAX_TLS_GD_TO_IE_PAGE_PC ||
+          r.expr == R_RELAX_TLS_GD_TO_LE) {
+        if (relaxable(relocs, i))
+          remove = 4;
+      } else if (isPairRelaxable(relocs, i))
         relaxPCHi20Lo12(ctx, sec, i, loc, r, relocs[i + 2], remove);
       break;
     case R_LARCH_CALL36:
@@ -986,6 +992,17 @@ static bool relax(Ctx &ctx, InputSection &sec) {
           isUInt<12>(r.sym->getVA(ctx, r.addend)))
         remove = 4;
       break;
+    case R_LARCH_TLS_DESC_PC_LO12:
+      if (relaxable(relocs, i) &&
+          (r.expr == RE_LOONGARCH_RELAX_TLS_GD_TO_IE_PAGE_PC ||
+           r.expr == R_RELAX_TLS_GD_TO_LE))
+        remove = 4;
+      break;
+    case R_LARCH_TLS_DESC_LD:
+      if (relaxable(relocs, i) && r.expr == R_RELAX_TLS_GD_TO_LE &&
+          isUInt<12>(r.sym->getVA(ctx, r.addend)))
+        remove = 4;
+      break;
     }
 
     // For all anchors whose offsets are <= r.offset, they are preceded by
@@ -1214,6 +1231,10 @@ void LoongArch::relocateAlloc(InputSectionBase &sec, uint8_t *buf) const {
                            bits);
         relocateNoSym(loc, rel.type, val);
       } else {
+        isRelax = relaxable(relocs, i);
+        if (isRelax && (rel.type == R_LARCH_TLS_DESC_PC_HI20 ||
+                        rel.type == R_LARCH_TLS_DESC_PC_LO12))
+          continue;
         tlsdescToIe(loc, rel, val);
       }
       continue;
@@ -1230,6 +1251,11 @@ void LoongArch::relocateAlloc(InputSectionBase &sec, uint8_t *buf) const {
                            bits);
         relocateNoSym(loc, rel.type, val);
       } else {
+        isRelax = relaxable(relocs, i);
+        if (isRelax && (rel.type == R_LARCH_TLS_DESC_PC_HI20 ||
+                        rel.type == R_LARCH_TLS_DESC_PC_LO12 ||
+                        (rel.type == R_LARCH_TLS_DESC_LD && isUInt<12>(val))))
+          continue;
         tlsdescToLe(loc, rel, val);
       }
       continue;
diff --git a/lld/test/ELF/loongarch-relax-tlsdesc.s b/lld/test/ELF/loongarch-relax-tlsdesc.s
index 5e538985d1402c..666ca6bd1e7243 100644
--- a/lld/test/ELF/loongarch-relax-tlsdesc.s
+++ b/lld/test/ELF/loongarch-relax-tlsdesc.s
@@ -9,7 +9,6 @@
 # RUN: llvm-readobj -r -x .got a.64.so | FileCheck --check-prefix=GD64-RELA %s
 # RUN: llvm-objdump --no-show-raw-insn -dr -h a.64.so | FileCheck %s --check-prefix=GD64
 
-## FIXME: IE/LE relaxation have not yet been implemented, --relax/--no-relax obtain the same results.
 ## Transition from TLSDESC to IE/LE. Also check --emit-relocs.
 # RUN: ld.lld --relax -e 0 -z now --emit-relocs a.64.o c.64.o -o a.64.le
 # RUN: llvm-readobj -r -x .got a.64.le 2>&1 | FileCheck --check-prefix=LE64-RELA %s
@@ -73,25 +72,21 @@
 # LE64-RELA: could not find section '.got'
 
 ## a@tprel = 0x8
-# LE64:        20158: nop
+# LE64:        20158: ori     $a0, $zero, 8
 # LE64-NEXT:            R_LARCH_TLS_DESC_PC_HI20 a
 # LE64-NEXT:            R_LARCH_RELAX *ABS*
-# LE64-NEXT:          nop
 # LE64-NEXT:            R_LARCH_TLS_DESC_PC_LO12 a
 # LE64-NEXT:            R_LARCH_RELAX *ABS*
-# LE64-NEXT:          nop
 # LE64-NEXT:            R_LARCH_TLS_DESC_LD a
 # LE64-NEXT:            R_LARCH_RELAX *ABS*
-# LE64-NEXT:          ori     $a0, $zero, 8
 # LE64-NEXT:            R_LARCH_TLS_DESC_CALL a
 # LE64-NEXT:            R_LARCH_RELAX *ABS*
 # LE64-NEXT:          add.d   $a1, $a0, $tp
 
 ## b@tprel = 0x7ff
-# LE64:        2016c: nop
+# LE64:        20160: nop
 # LE64-NEXT:            R_LARCH_TLS_DESC_PC_HI20 b
 # LE64-NEXT:            R_LARCH_RELAX *ABS*
-# LE64-NEXT:          nop
 # LE64-NEXT:            R_LARCH_TLS_DESC_PC_LO12 b
 # LE64-NEXT:          nop
 # LE64-NEXT:            R_LARCH_TLS_DESC_LD b
@@ -101,7 +96,7 @@
 
 ## c@tprel = 0x800
 ## Without R_LARCH_RELAX relocation. No relaxation.
-# LE64:        20180: nop
+# LE64:        20170: nop
 # LE64-NEXT:            R_LARCH_TLS_DESC_PC_HI20 c
 # LE64-NEXT:          addi.d  $t0, $zero, 0
 # LE64-NEXT:          nop
@@ -115,13 +110,11 @@
 # LE64-NEXT:          add.d   $a3, $a0, $tp
 
 ## d@tprel = 0x1000
-# LE64:        201a0: nop
+# LE64:        20190: lu12i.w $a0, 1
 # LE64-NEXT:            R_LARCH_TLS_DESC_PC_HI20 d
 # LE64-NEXT:            R_LARCH_RELAX *ABS*
-# LE64-NEXT:          nop
 # LE64-NEXT:            R_LARCH_TLS_DESC_PC_LO12 d
 # LE64-NEXT:            R_LARCH_RELAX *ABS*
-# LE64-NEXT:          lu12i.w $a0, 1
 # LE64-NEXT:            R_LARCH_TLS_DESC_LD d
 # LE64-NEXT:          ori     $a0, $a0, 0
 # LE64-NEXT:            R_LARCH_TLS_DESC_CALL d
@@ -160,35 +153,31 @@
 # LE64-NORELAX-NEXT:          add.d   $a4, $a0, $tp
 
 # IE64-RELA:      .rela.dyn {
-# IE64-RELA-NEXT:   0x30408 R_LARCH_TLS_TPREL64 c 0x0
-# IE64-RELA-NEXT:   0x30410 R_LARCH_TLS_TPREL64 d 0x0
+# IE64-RELA-NEXT:   0x303F0 R_LARCH_TLS_TPREL64 c 0x0
+# IE64-RELA-NEXT:   0x303F8 R_LARCH_TLS_TPREL64 d 0x0
 # IE64-RELA-NEXT: }
 # IE64-RELA:      Hex dump of section '.got':
-# IE64-RELA-NEXT: 0x00030408 00000000 00000000 00000000 00000000 .
+# IE64-RELA-NEXT: 0x000303f0 00000000 00000000 00000000 00000000 .
 
-# IE64:   .got           00000010 0000000000030408
+# IE64:   .got           00000010 00000000000303f0
 
 ## a and b are optimized to use LE. c and d are optimized to IE.
 ## a@tprel = 0x8
-# IE64:        202c8: nop
+# IE64:        202c8: ori     $a0, $zero, 8
 # IE64-NEXT:            R_LARCH_TLS_DESC_PC_HI20 a
 # IE64-NEXT:            R_LARCH_RELAX *ABS*
-# IE64-NEXT:          nop
 # IE64-NEXT:            R_LARCH_TLS_DESC_PC_LO12 a
 # IE64-NEXT:            R_LARCH_RELAX *ABS*
-# IE64-NEXT:          nop
 # IE64-NEXT:            R_LARCH_TLS_DESC_LD a
 # IE64-NEXT:            R_LARCH_RELAX *ABS*
-# IE64-NEXT:          ori     $a0, $zero, 8
 # IE64-NEXT:            R_LARCH_TLS_DESC_CALL a
 # IE64-NEXT:            R_LARCH_RELAX *ABS*
 # IE64-NEXT:          add.d   $a1, $a0, $tp
 
 ## b@tprel = 0x7ff
-# IE64:        202dc: nop
+# IE64:        202d0: nop
 # IE64-NEXT:            R_LARCH_TLS_DESC_PC_HI20 b
 # IE64-NEXT:            R_LARCH_RELAX *ABS*
-# IE64-NEXT:          nop
 # IE64-NEXT:            R_LARCH_TLS_DESC_PC_LO12 b
 # IE64-NEXT:          nop
 # IE64-NEXT:            R_LARCH_TLS_DESC_LD b
@@ -196,9 +185,9 @@
 # IE64-NEXT:            R_LARCH_TLS_DESC_CALL b
 # IE64-NEXT:          add.d   $a2, $a0, $tp
 
-## &.got[c]-. = 0x30408 - 0x20300: 0x10 pages, page offset 0x408
+## &.got[c]-. = 0x303f0 - 0x202f0: 0x10 pages, page offset 0x3f0
 ## Without R_LARCH_RELAX relocation. No relaxation.
-# IE64:        202f0: nop
+# IE64:        202e0: nop
 # IE64-NEXT:            R_LARCH_TLS_DESC_PC_HI20 c
 # IE64-NEXT:          addi.d  $t0, $zero, 0
 # IE64-NEXT:          nop
@@ -207,20 +196,18 @@
 # IE64-NEXT:          pcalau12i $a0, 16
 # IE64-NEXT:            R_LARCH_TLS_DESC_LD c
 # IE64-NEXT:          addi.d  $t0, $t0, 1
-# IE64-NEXT:          ld.d    $a0, $a0, 1032
+# IE64-NEXT:          ld.d    $a0, $a0, 1008
 # IE64-NEXT:            R_LARCH_TLS_DESC_CALL c
 # IE64-NEXT:          add.d   $a3, $a0, $tp
 
-## &.got[d]-. = 0x30408+8 - 0x20318: 0x10 pages, page offset 0x410
-# IE64:        20310: nop
+## &.got[d]-. = 0x303f0+8 - 0x20300: 0x10 pages, page offset 0x3f8
+# IE64:        20300: pcalau12i $a0, 16
 # IE64-NEXT:            R_LARCH_TLS_DESC_PC_HI20 d
 # IE64-NEXT:            R_LARCH_RELAX *ABS*
-# IE64-NEXT:          nop
 # IE64-NEXT:            R_LARCH_TLS_DESC_PC_LO12 d
 # IE64-NEXT:            R_LARCH_RELAX *ABS*
-# IE64-NEXT:          pcalau12i $a0, 16
 # IE64-NEXT:            R_LARCH_TLS_DESC_LD d
-# IE64-NEXT:          ld.d    $a0, $a0, 1040
+# IE64-NEXT:          ld.d    $a0, $a0, 1016
 # IE64-NEXT:            R_LARCH_TLS_DESC_CALL d
 # IE64-NEXT:          add.d   $a4, $a0, $tp

This reverts commit 653281c.

Complement https://. When relaxation enable, remove redundant NOPs.

SixWeining · 2025-02-19T10:20:44Z

cc @xen0n

ylzsx added 30 commits January 20, 2025 10:08

la.pcrel relax test modify.

95f4540

Add test for PCHi20Lo12

7b133c2

Add test for got symbols relaxation.

abc1a45

Similar to aarch64-adrp-ldr-got-symbols.s.

Modify test. NFC

1b1804e

Dependency on #123017

Add check for register.

30cb382

Relax call36/tail36.

f1f995b

Instructions with relocation `R_LARCH_CALL36` may be relax as follows: ``` From: pcaddu18i $dest, %call36(foo) R_LARCH_CALL36, R_LARCH_RELAX jirl $r, $dest, 0 To: b/bl foo # bl if r=$ra, b if r=$zero R_LARCH_B26 ```

modify test for call36/tail36.

f227ae5

Modify test. Add the option --relax.

f2aae15

Add test for TLSLD/TLSGD when relax enabled.

1e9aa52

Modify test for TLSLE when relax enabled.

91da25e

Add test for loongarch-relax-tls-le.s and modify loongarch-relax-emit…

2066c5f

…-relocs.s

Modify test. Add --relax option.

b57c40e

Add test loongarch-relax-tlsdesc.s

880523c

Relax desc_pc_hi20+desc_pc_lo12 ==> desc_pcrel_20

a48a2b9

Modify test loongarch-relax-tlsdesc.s

576773c

Remove --emit-relocs in loongarch-relax-tlsdesc.s

b2f11e3

llvm-objdump -dr cannot print the relocations, however objudmp works well. It also occurs for RISCV.

Modify test. Add --relax option.

d5c9d64

Test loongarch-tls-ie.s pass.

78a16b0

Add test loongarch-relax-tls-ie.s

4f2dca4

modify RE_LOONGARCH_GOT.

81d44f2

Fix bug.

31a9dab

Support relaxation during IE to LE conversion.

83d8b7e

Complement https://. When relaxation enable, remove redundant NOPs.

Modify loongarch-relax-tls-ie.s when relaxation is enabled.

6b79432

Modify test. Add --relax option.

9d99de8

Modify loongarch-tlsdesc.s and loongarch-tlsdesc-pcrel20-s2.s

459e25d

ylzsx added 3 commits January 21, 2025 14:05

Modify loongarch-relax-tlsdesc.s init.

894a97a

Modify loongarch-relax-tlsdesc.s

5262935

Add comments.

653281c

llvmbot added lld lld:ELF labels Jan 21, 2025

ylzsx added the backend:loongarch label Jan 21, 2025

Revert "Add comments."

2a30b6c

This reverts commit 653281c.

ylzsx force-pushed the users/ylzsx/r-tlsdesc-to-iele-relax branch from ad372af to e024b7c Compare January 21, 2025 12:21

ylzsx mentioned this pull request Jan 21, 2025

[lld][LoongArch] GOT indirection to PC relative optimization #123743

Open

ylzsx added 3 commits January 22, 2025 08:47

Delete the wrong flag NEEDS_TLSGD_TO_IE in the extreme code model.

0c6ba1d

Support relaxation during TLSDESC GD/LD to IE/LE conversion.

63a756b

Complement https://. When relaxation enable, remove redundant NOPs.

Modify loongarch-relax-tlsdesc.s.

99a1e07

ylzsx changed the title ~~[lld][LoongArch] Support relaxation during TLSDESC GD/LD to IE/LE conversion.~~ [lld][LoongArch] Support relaxation during TLSDESC GD/LD to IE/LE conversion Jan 22, 2025

ylzsx force-pushed the users/ylzsx/r-tlsdesc-to-iele-relax branch from e024b7c to 99a1e07 Compare January 22, 2025 06:15

ylzsx requested review from MaskRay, wangleiat and SixWeining January 23, 2025 02:06

ylzsx force-pushed the users/ylzsx/r-tlsdesc-to-iele-norelax branch from 0c6ba1d to baee006 Compare April 11, 2025 09:50

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

[lld][LoongArch] Support relaxation during TLSDESC GD/LD to IE/LE conversion #123730

[lld][LoongArch] Support relaxation during TLSDESC GD/LD to IE/LE conversion #123730

ylzsx commented Jan 21, 2025

llvmbot commented Jan 21, 2025 •

edited

Loading

SixWeining commented Feb 19, 2025

[lld][LoongArch] Support relaxation during TLSDESC GD/LD to IE/LE conversion #123730

Are you sure you want to change the base?

[lld][LoongArch] Support relaxation during TLSDESC GD/LD to IE/LE conversion #123730

Conversation

ylzsx commented Jan 21, 2025

llvmbot commented Jan 21, 2025 • edited Loading

SixWeining commented Feb 19, 2025

llvmbot commented Jan 21, 2025 •

edited

Loading