[lld][LoongArch] Handle extreme code model relocs according to psABI v2.30 (#73387)

psABI v2.30 requires the extreme code model instructions sequence
(pcalau12i+addi.d+lu32i.d+lu52i.d) to be adjacent.

See llvm/llvm-project#71907 and
loongson-community/discussions#17 for details.

(cherry picked from commit 38394a3)

Author: SixWeining

lld/ELF/Arch/LoongArch.cpp

@@ -85,89 +85,33 @@ static uint64_t getLoongArchPage(uint64_t p) {
 static uint32_t lo12(uint32_t val) { return val & 0xfff; }
 
 // Calculate the adjusted page delta between dest and PC.
-uint64_t elf::getLoongArchPageDelta(uint64_t dest, uint64_t pc) {
-  // Consider the large code model access pattern, of which the smaller code
-  // models' access patterns are a subset:
-  //
-  //     pcalau12i       U, %foo_hi20(sym)        ; b in [-0x80000, 0x7ffff]
-  //     addi.d          T, zero, %foo_lo12(sym)  ; a in [-0x800, 0x7ff]
-  //     lu32i.d         T, %foo64_lo20(sym)      ; c in [-0x80000, 0x7ffff]
-  //     lu52i.d         T, T, %foo64_hi12(sym)   ; d in [-0x800, 0x7ff]
-  //     {ldx,stx,add}.* dest, U, T
-  //
-  // Let page(pc) = 0xRRR'QQQQQ'PPPPP'000 and dest = 0xZZZ'YYYYY'XXXXX'AAA,
-  // with RQ, P, ZY, X and A representing the respective bitfields as unsigned
-  // integers. We have:
-  //
-  //     page(dest) = 0xZZZ'YYYYY'XXXXX'000
-  //     - page(pc) = 0xRRR'QQQQQ'PPPPP'000
-  //     ----------------------------------
-  //                  0xddd'ccccc'bbbbb'000
-  //
-  // Now consider the above pattern's actual effects:
-  //
-  //     page(pc)                     0xRRR'QQQQQ'PPPPP'000
-  //     pcalau12i                  + 0xiii'iiiii'bbbbb'000
-  //     addi                       + 0xjjj'jjjjj'kkkkk'AAA
-  //     lu32i.d & lu52i.d          + 0xddd'ccccc'00000'000
-  //     --------------------------------------------------
-  //     dest = U + T
-  //          = ((RQ<<32) + (P<<12) + i + (b<<12)) + (j + k + A + (cd<<32))
-  //          = (((RQ+cd)<<32) + i + j) + (((P+b)<<12) + k) + A
-  //          = (ZY<<32)                + (X<<12)           + A
-  //
-  //     ZY<<32 = (RQ<<32)+(cd<<32)+i+j, X<<12 = (P<<12)+(b<<12)+k
-  //     cd<<32 = (ZY<<32)-(RQ<<32)-i-j, b<<12 = (X<<12)-(P<<12)-k
-  //
-  // where i and k are terms representing the effect of b's and A's sign
-  // extension respectively.
-  //
-  //     i = signed b < 0 ? -0x10000'0000 : 0
-  //     k = signed A < 0 ? -0x1000 : 0
-  //
-  // The j term is a bit complex: it represents the higher half of
-  // sign-extended bits from A that are effectively lost if i == 0 but k != 0,
-  // due to overwriting by lu32i.d & lu52i.d.
-  //
-  //     j = signed A < 0 && signed b >= 0 ? 0x10000'0000 : 0
-  //
-  // The actual effect of the instruction sequence before the final addition,
-  // i.e. our desired result value, is thus:
-  //
-  //     result = (cd<<32) + (b<<12)
-  //            = (ZY<<32)-(RQ<<32)-i-j + (X<<12)-(P<<12)-k
-  //            = ((ZY<<32)+(X<<12)) - ((RQ<<32)+(P<<12)) - i - j - k
-  //            = page(dest) - page(pc) - i - j - k
-  //
-  // when signed A >= 0 && signed b >= 0:
-  //
-  //     i = j = k = 0
-  //     result = page(dest) - page(pc)
-  //
-  // when signed A >= 0 && signed b < 0:
-  //
-  //     i = -0x10000'0000, j = k = 0
-  //     result = page(dest) - page(pc) + 0x10000'0000
-  //
-  // when signed A < 0 && signed b >= 0:
-  //
-  //     i = 0, j = 0x10000'0000, k = -0x1000
-  //     result = page(dest) - page(pc) - 0x10000'0000 + 0x1000
-  //
-  // when signed A < 0 && signed b < 0:
-  //
-  //     i = -0x10000'0000, j = 0, k = -0x1000
-  //     result = page(dest) - page(pc) + 0x1000
-  uint64_t result = getLoongArchPage(dest) - getLoongArchPage(pc);
-  bool negativeA = lo12(dest) > 0x7ff;
-  bool negativeB = (result & 0x8000'0000) != 0;
-
-  if (negativeA)
-    result += 0x1000;
-  if (negativeA && !negativeB)
-    result -= 0x10000'0000;
-  else if (!negativeA && negativeB)
-    result += 0x10000'0000;
+uint64_t elf::getLoongArchPageDelta(uint64_t dest, uint64_t pc, RelType type) {
+  // Note that if the sequence being relocated is `pcalau12i + addi.d + lu32i.d
+  // + lu52i.d`, they must be adjancent so that we can infer the PC of
+  // `pcalau12i` when calculating the page delta for the other two instructions
+  // (lu32i.d and lu52i.d). Compensate all the sign-extensions is a bit
+  // complicated. Just use psABI recommended algorithm.
+  uint64_t pcalau12i_pc;
+  switch (type) {
+  case R_LARCH_PCALA64_LO20:
+  case R_LARCH_GOT64_PC_LO20:
+  case R_LARCH_TLS_IE64_PC_LO20:
+    pcalau12i_pc = pc - 8;
+    break;
+  case R_LARCH_PCALA64_HI12:
+  case R_LARCH_GOT64_PC_HI12:
+  case R_LARCH_TLS_IE64_PC_HI12:
+    pcalau12i_pc = pc - 12;
+    break;
+  default:
+    pcalau12i_pc = pc;
+    break;
+  }
+  uint64_t result = getLoongArchPage(dest) - getLoongArchPage(pcalau12i_pc);
+  if (dest & 0x800)
+    result += 0x1000 - 0x1'0000'0000;
+  if (result & 0x8000'0000)
+    result += 0x1'0000'0000;
   return result;
 }
 

lld/ELF/InputSection.cpp

@@ -712,8 +712,8 @@ uint64_t InputSectionBase::getRelocTargetVA(const InputFile *file, RelType type,
     return sym.getGotVA() + a - p;
   case R_LOONGARCH_GOT_PAGE_PC:
     if (sym.hasFlag(NEEDS_TLSGD))
-      return getLoongArchPageDelta(in.got->getGlobalDynAddr(sym) + a, p);
-    return getLoongArchPageDelta(sym.getGotVA() + a, p);
+      return getLoongArchPageDelta(in.got->getGlobalDynAddr(sym) + a, p, type);
+    return getLoongArchPageDelta(sym.getGotVA() + a, p, type);
   case R_MIPS_GOTREL:
     return sym.getVA(a) - in.mipsGot->getGp(file);
   case R_MIPS_GOT_GP:
@@ -763,7 +763,7 @@ uint64_t InputSectionBase::getRelocTargetVA(const InputFile *file, RelType type,
     return 0;
   }
   case R_LOONGARCH_PAGE_PC:
-    return getLoongArchPageDelta(sym.getVA(a), p);
+    return getLoongArchPageDelta(sym.getVA(a), p, type);
   case R_PC:
   case R_ARM_PCA: {
     uint64_t dest;
@@ -798,7 +798,7 @@ uint64_t InputSectionBase::getRelocTargetVA(const InputFile *file, RelType type,
   case R_PPC64_CALL_PLT:
     return sym.getPltVA() + a - p;
   case R_LOONGARCH_PLT_PAGE_PC:
-    return getLoongArchPageDelta(sym.getPltVA() + a, p);
+    return getLoongArchPageDelta(sym.getPltVA() + a, p, type);
   case R_PLT_GOTPLT:
     return sym.getPltVA() + a - in.gotPlt->getVA();
   case R_PPC32_PLTREL:
@@ -860,7 +860,7 @@ uint64_t InputSectionBase::getRelocTargetVA(const InputFile *file, RelType type,
   case R_TLSGD_PC:
     return in.got->getGlobalDynAddr(sym) + a - p;
   case R_LOONGARCH_TLSGD_PAGE_PC:
-    return getLoongArchPageDelta(in.got->getGlobalDynAddr(sym) + a, p);
+    return getLoongArchPageDelta(in.got->getGlobalDynAddr(sym) + a, p, type);
   case R_TLSLD_GOTPLT:
     return in.got->getVA() + in.got->getTlsIndexOff() + a - in.gotPlt->getVA();
   case R_TLSLD_GOT:

lld/ELF/Target.h

@@ -229,7 +229,7 @@ void addPPC64SaveRestore();
 uint64_t getPPC64TocBase();
 uint64_t getAArch64Page(uint64_t expr);
 template <typename ELFT> void writeARMCmseImportLib();
-uint64_t getLoongArchPageDelta(uint64_t dest, uint64_t pc);
+uint64_t getLoongArchPageDelta(uint64_t dest, uint64_t pc, RelType type);
 void riscvFinalizeRelax(int passes);
 void mergeRISCVAttributesSections();
 void addArmInputSectionMappingSymbols();