Fix LoongArch support

gen/abi/loongarch64.cpp

@@ -9,7 +9,7 @@
 //===----------------------------------------------------------------------===//
 //
 // ABI spec:
-// https://loongson.github.io/LoongArch-Documentation/LoongArch-ELF-ABI-EN.html
+// https://github.com/loongson/la-abi-specs/blob/release/lapcs.adoc
 //
 //===----------------------------------------------------------------------===//
 
@@ -20,9 +20,101 @@
 #include "gen/llvmhelpers.h"
 #include "gen/tollvm.h"
 
+using namespace dmd;
+
+namespace {
+struct Integer2Rewrite : BaseBitcastABIRewrite {
+  LLType *type(Type *t) override {
+    return LLStructType::get(gIR->context(),
+                             {DtoType(Type::tint64), DtoType(Type::tint64)});
+  }
+};
+
+struct FlattenedFields {
+  Type *fields[2];
+  int length = 0; // use -1 to represent "no need to rewrite" condition
+};
+
+FlattenedFields visitStructFields(Type *ty, unsigned baseOffset) {
+  // recursively visit a POD struct to flatten it
+  FlattenedFields result;
+  if (auto ts = ty->isTypeStruct()) {
+    for (auto fi : ts->sym->fields) {
+      auto sub =
+          visitStructFields(fi->type->toBasetype(), baseOffset + fi->offset);
+      if (sub.length == -1 || result.length + sub.length > 2) {
+        result.length = -1;
+        return result;
+      }
+      for (unsigned i = 0; i < (unsigned)sub.length; ++i) {
+        result.fields[result.length++] = sub.fields[i];
+      }
+    }
+    return result;
+  }
+  switch (ty->ty) {
+  case TY::Tcomplex32: // treat it as {float32, float32}
+    result.fields[0] = pointerTo(Type::tfloat32);
+    result.fields[1] = pointerTo(Type::tfloat32);
+    result.length = 2;
+    break;
+  case TY::Tcomplex64: // treat it as {float64, float64}
+    result.fields[0] = pointerTo(Type::tfloat64);
+    result.fields[1] = pointerTo(Type::tfloat64);
+    result.length = 2;
+    break;
+  default:
+    if (ty->size() > 8) {
+      // field larger than GRLEN and FRLEN
+      result.length = -1;
+      break;
+    }
+    result.fields[0] = ty;
+    result.length = 1;
+    break;
+  }
+  return result;
+}
+
+struct HardfloatRewrite : BaseBitcastABIRewrite {
+  LLType *type(Type *ty, const FlattenedFields &flat) {
+    if (flat.length == 1) {
+      return LLStructType::get(gIR->context(), {DtoType(flat.fields[0])},
+                               false);
+    }
+    assert(flat.length == 2);
+    LLType *t[2];
+    for (unsigned i = 0; i < 2; ++i) {
+      t[i] =
+          flat.fields[i]->isfloating()
+              ? DtoType(flat.fields[i])
+              : LLIntegerType::get(gIR->context(), flat.fields[i]->size() * 8);
+    }
+    return LLStructType::get(gIR->context(), {t[0], t[1]}, false);
+  }
+  LLType *type(Type *ty) override {
+    return type(ty, visitStructFields(ty->toBasetype(), 0));
+  }
+};
+} // anonymous namespace
+
 struct LoongArch64TargetABI : TargetABI {
 private:
+  HardfloatRewrite hardfloatRewrite;
   IndirectByvalRewrite indirectByvalRewrite{};
+  Integer2Rewrite integer2Rewrite;
+  IntegerRewrite integerRewrite;
+
+  bool requireHardfloatRewrite(Type *ty) {
+    if (!isPOD(ty) || !ty->toBasetype()->isTypeStruct())
+      return false;
+    auto result = visitStructFields(ty->toBasetype(), 0);
+    if (result.length <= 0)
+      return false;
+    if (result.length == 1)
+      return result.fields[0]->isfloating();
+    return result.fields[0]->isfloating() || result.fields[1]->isfloating();
+  }
 
 public:
   auto returnInArg(TypeFunction *tf, bool) -> bool override {
@@ -45,27 +137,51 @@ struct LoongArch64TargetABI : TargetABI {
   }
 
   void rewriteFunctionType(IrFuncTy &fty) override {
-    if (!fty.ret->byref) {
-      rewriteArgument(fty, *fty.ret);
+    if (!skipReturnValueRewrite(fty)) {
+      if (requireHardfloatRewrite(fty.ret->type)) {
+        // rewrite here because we should not apply this to variadic arguments
+        hardfloatRewrite.applyTo(*fty.ret);
+      } else {
+        rewriteArgument(fty, *fty.ret);
+      }
     }
 
     for (auto arg : fty.args) {
       if (!arg->byref) {
-        rewriteArgument(fty, *arg);
+        if (requireHardfloatRewrite(arg->type)) {
+          // rewrite here because we should not apply this to variadic arguments
+          hardfloatRewrite.applyTo(*arg);
+        } else {
+          rewriteArgument(fty, *arg);
+        }
       }
     }
   }
 
   void rewriteArgument(IrFuncTy &fty, IrFuncTyArg &arg) override {
-    if (arg.byref) {
-      return;
-    }
-
     if (!isPOD(arg.type)) {
       // non-PODs should be passed in memory
       indirectByvalRewrite.applyTo(arg);
       return;
     }
+
+    Type *ty = arg.type->toBasetype();
+    if (ty->isintegral() && (ty->ty == TY::Tint32 || ty->ty == TY::Tuns32 ||
+                             ty->ty == TY::Tdchar)) {
+      // In the LP64D ABI, both int32 and unsigned int32 are stored in
+      // general-purpose registers as proper sign extensions of their
+      // 32-bit values. So, the native ABI function's int32 arguments and
+      // return values should have the `signext` attribute.
+      // C example: https://godbolt.org/z/vcjErxj76
+      arg.attrs.addAttribute(LLAttribute::SExt);
+    } else if (isAggregate(ty) && ty->size() && ty->size() <= 16) {
+      if (ty->size() > 8 && DtoAlignment(ty) < 16) {
+        // pass the aggregate as {int64, int64} to avoid wrong alignment
+        integer2Rewrite.applyToIfNotObsolete(arg);
+      } else {
+        integerRewrite.applyToIfNotObsolete(arg);
+      }
+    }
   }
 };
 

runtime/druntime/src/core/thread/fiber.d

@@ -229,6 +229,8 @@ private
       extern (C) void fiber_switchContext( void** oldp, void* newp ) nothrow @nogc;
       version (AArch64)
           extern (C) void fiber_trampoline() nothrow;
+      version (LoongArch64)
+          extern (C) void fiber_trampoline() nothrow;
   }
   else version (LDC_Windows)
   {
@@ -1831,7 +1833,6 @@ private:
             // Like others, FP registers and return address ($r1) are kept
             // below the saved stack top (tstack) to hide from GC scanning.
             // fiber_switchContext expects newp sp to look like this:
-            //   10: $r21 (reserved)
             //    9: $r22 (frame pointer)
             //    8: $r23
             //   ...
@@ -1847,8 +1848,8 @@ private:
 
             // Only need to set return address ($r1).  Everything else is fine
             // zero initialized.
-            pstack -= size_t.sizeof * 11;    // skip past space reserved for $r21-$r31
-            push (cast(size_t) &fiber_entryPoint);
+            pstack -= size_t.sizeof * 10;    // skip past space reserved for $r22-$r31
+            push(cast(size_t) &fiber_trampoline); // see threadasm.S for docs
             pstack += size_t.sizeof;         // adjust sp (newp) above lr
         }
         else version (AsmAArch64_Posix)

runtime/druntime/src/core/threadasm.S

@@ -544,9 +544,10 @@ fiber_switchContext:
  */
  .text
 .globl fiber_switchContext
+.type  fiber_switchContext, %function
 fiber_switchContext:
     .cfi_startproc
-    .cfi_undefined $ra
+    .cfi_undefined 1
     # reserve space on stack
     addi.d  $sp, $sp, -19 * 8
 
@@ -610,6 +611,28 @@ fiber_switchContext:
 
     jr $ra
     .cfi_endproc
+
+/**
+ * When generating any kind of backtrace (gdb, exception handling) for
+ * a function called in a Fiber, we need to tell the unwinder to stop
+ * at our Fiber main entry point, i.e. we need to mark the bottom of
+ * the call stack. This can be done by clearing the return address register $ra
+ * prior to calling fiber_entryPoint (i.e. in fiber_switchContext) or
+ * using a .cfi_undefined directive for the return address register in the
+ * Fiber entry point. cfi_undefined seems to yield better results in gdb.
+ * Unfortunately we can't place it into fiber_entryPoint using inline
+ * asm, so we use this trampoline instead.
+ */
+        .text
+        .global CSYM(fiber_trampoline)
+        .p2align  2
+        .type   fiber_trampoline, %function
+CSYM(fiber_trampoline):
+        .cfi_startproc
+        .cfi_undefined 1
+        // fiber_entryPoint never returns
+        bl CSYM(fiber_entryPoint)
+        .cfi_endproc
 #elif defined(__arm__) && (defined(__ARM_EABI__) || defined(__APPLE__))
 /************************************************************************************
  * ARM ASM BITS