[RuntimeDyld][LoongArch] Support large code model (llvm#148584)

Co-authored-by: Weining Lu <[email protected]>
(cherry picked from commit 4e35ae1)

Author: 2 people

llvm/lib/ExecutionEngine/RuntimeDyld/RuntimeDyldELF.cpp

@@ -728,6 +728,32 @@ static inline uint32_t extractBits(uint64_t Val, uint32_t Hi, uint32_t Lo) {
   return Hi == 63 ? Val >> Lo : (Val & (((1ULL << (Hi + 1)) - 1))) >> Lo;
 }
 
+// Calculate the adjusted page delta between dest and PC. The code is copied
+// from lld and see comments there for more details.
+static uint64_t getLoongArchPageDelta(uint64_t dest, uint64_t pc,
+                                      uint32_t type) {
+  uint64_t pcalau12i_pc;
+  switch (type) {
+  case ELF::R_LARCH_PCALA64_LO20:
+  case ELF::R_LARCH_GOT64_PC_LO20:
+    pcalau12i_pc = pc - 8;
+    break;
+  case ELF::R_LARCH_PCALA64_HI12:
+  case ELF::R_LARCH_GOT64_PC_HI12:
+    pcalau12i_pc = pc - 12;
+    break;
+  default:
+    pcalau12i_pc = pc;
+    break;
+  }
+  uint64_t result = (dest & ~0xfffULL) - (pcalau12i_pc & ~0xfffULL);
+  if (dest & 0x800)
+    result += 0x1000 - 0x1'0000'0000;
+  if (result & 0x8000'0000)
+    result += 0x1'0000'0000;
+  return result;
+}
+
 void RuntimeDyldELF::resolveLoongArch64Relocation(const SectionEntry &Section,
                                                   uint64_t Offset,
                                                   uint64_t Value, uint32_t Type,
@@ -779,10 +805,7 @@ void RuntimeDyldELF::resolveLoongArch64Relocation(const SectionEntry &Section,
   case ELF::R_LARCH_GOT_PC_HI20:
   case ELF::R_LARCH_PCALA_HI20: {
     uint64_t Target = Value + Addend;
-    uint64_t TargetPage =
-        (Target + (Target & 0x800)) & ~static_cast<uint64_t>(0xfff);
-    uint64_t PCPage = FinalAddress & ~static_cast<uint64_t>(0xfff);
-    int64_t PageDelta = TargetPage - PCPage;
+    int64_t PageDelta = getLoongArchPageDelta(Target, FinalAddress, Type);
     auto Instr = support::ulittle32_t::ref(TargetPtr);
     uint32_t Imm31_12 = extractBits(PageDelta, /*Hi=*/31, /*Lo=*/12) << 5;
     Instr = (Instr & 0xfe00001f) | Imm31_12;
@@ -796,6 +819,24 @@ void RuntimeDyldELF::resolveLoongArch64Relocation(const SectionEntry &Section,
     Instr = (Instr & 0xffc003ff) | Imm11_0;
     break;
   }
+  case ELF::R_LARCH_GOT64_PC_LO20:
+  case ELF::R_LARCH_PCALA64_LO20: {
+    uint64_t Target = Value + Addend;
+    int64_t PageDelta = getLoongArchPageDelta(Target, FinalAddress, Type);
+    auto Instr = support::ulittle32_t::ref(TargetPtr);
+    uint32_t Imm51_32 = extractBits(PageDelta, /*Hi=*/51, /*Lo=*/32) << 5;
+    Instr = (Instr & 0xfe00001f) | Imm51_32;
+    break;
+  }
+  case ELF::R_LARCH_GOT64_PC_HI12:
+  case ELF::R_LARCH_PCALA64_HI12: {
+    uint64_t Target = Value + Addend;
+    int64_t PageDelta = getLoongArchPageDelta(Target, FinalAddress, Type);
+    auto Instr = support::ulittle32_t::ref(TargetPtr);
+    uint32_t Imm63_52 = extractBits(PageDelta, /*Hi=*/63, /*Lo=*/52) << 10;
+    Instr = (Instr & 0xffc003ff) | Imm63_52;
+    break;
+  }
   case ELF::R_LARCH_ABS_HI20: {
     uint64_t Target = Value + Addend;
     auto Instr = support::ulittle32_t::ref(TargetPtr);
@@ -1581,7 +1622,9 @@ RuntimeDyldELF::processRelocationRef(
     if (RelType == ELF::R_LARCH_B26 && MemMgr.allowStubAllocation()) {
       resolveLoongArch64Branch(SectionID, Value, RelI, Stubs);
     } else if (RelType == ELF::R_LARCH_GOT_PC_HI20 ||
-               RelType == ELF::R_LARCH_GOT_PC_LO12) {
+               RelType == ELF::R_LARCH_GOT_PC_LO12 ||
+               RelType == ELF::R_LARCH_GOT64_PC_HI12 ||
+               RelType == ELF::R_LARCH_GOT64_PC_LO20) {
       uint64_t GOTOffset = findOrAllocGOTEntry(Value, ELF::R_LARCH_64);
       resolveGOTOffsetRelocation(SectionID, Offset, GOTOffset + Addend,
                                  RelType);
@@ -2748,7 +2791,9 @@ bool RuntimeDyldELF::relocationNeedsGot(const RelocationRef &R) const {
 
   if (Arch == Triple::loongarch64)
     return RelTy == ELF::R_LARCH_GOT_PC_HI20 ||
-           RelTy == ELF::R_LARCH_GOT_PC_LO12;
+           RelTy == ELF::R_LARCH_GOT_PC_LO12 ||
+           RelTy == ELF::R_LARCH_GOT64_PC_HI12 ||
+           RelTy == ELF::R_LARCH_GOT64_PC_LO20;
 
   if (Arch == Triple::x86_64)
     return RelTy == ELF::R_X86_64_GOTPCREL ||

llvm/test/ExecutionEngine/RuntimeDyld/LoongArch/ELF_LoongArch_relocations.s

@@ -2,6 +2,9 @@
 # RUN: llvm-mc --triple=loongarch64 --filetype=obj -o %t/reloc.o %s
 # RUN: llvm-rtdyld --triple=loongarch64 --verify --check=%s %t/reloc.o \
 # RUN:     --map-section reloc.o,.got=0x21f00 \
+# RUN:     --map-section reloc.o,.sec.large.pc=0x0000000012345000 \
+# RUN:     --map-section reloc.o,.sec.large.got=0x44433333abcde000 \
+# RUN:     --map-section reloc.o,.sec.dummy=0x4443333334567111 \
 # RUN:     --dummy-extern abs=0x0123456789abcdef \
 # RUN:     --dummy-extern external_data=0x1234
 
@@ -100,3 +103,42 @@ named_data:
     .quad 0x2222222222222222
     .quad 0x3333333333333333
     .size named_data, .-named_data
+
+    .section .sec.large.pc,"ax"
+    .globl test_large_pc
+test_large_pc:
+## Code after link should be:
+## 1a44444d pcalau12i $t1, 139810
+## 02c4440c addi.d    $t0, $zero, 273
+## 1666666c lu32i.d   $t0, 209715
+## 0311118c lu52i.d   $t0, $t0, 1092
+
+# rtdyld-check: *{4}(test_large_pc) = 0x1a44444d
+    pcalau12i $t1, %pc_hi20(.sec.dummy)
+# rtdyld-check: *{4}(test_large_pc + 4) = 0x02c4440c
+    addi.d $t0, $zero, %pc_lo12(.sec.dummy)
+# rtdyld-check: *{4}(test_large_pc + 8) = 0x1666666c
+    lu32i.d $t0, %pc64_lo20(.sec.dummy)
+# rtdyld-check: *{4}(test_large_pc + 12) = 0x0311118c
+    lu52i.d $t0, $t0, %pc64_hi12(.sec.dummy)
+
+    .section .sec.large.got,"ax"
+    .globl test_large_got
+test_large_got:
+## Code after link should be:
+## 1aa8688d pcalau12i $t1, 344900
+## 02fc000c addi.d    $t0, $zero, -256
+## 1799996c lu32i.d   $t0, -209717
+## 032eed8c lu52i.d   $t0, $t0, -1093
+
+# rtdyld-check: *{4}(test_large_got) = 0x1aa8688d
+    pcalau12i $t1, %got_pc_hi20(external_data)
+# rtdyld-check: *{4}(test_large_got + 4) = 0x02fc000c
+    addi.d $t0, $zero, %got_pc_lo12(external_data)
+# rtdyld-check: *{4}(test_large_got + 8) = 0x1799996c
+    lu32i.d $t0, %got64_pc_lo20(external_data)
+# rtdyld-check: *{4}(test_large_got + 12) = 0x032eed8c
+    lu52i.d $t0, $t0, %got64_pc_hi12(external_data)
+
+    .section .sec.dummy,"a"
+    .word 0