Integrate backtrace refactoring and fixes for libunwind

[Geod24]

This cherry-picks commits from the following PRs, in order:

• Refactor some backtrace code to make it possible to use libunwind dlang/druntime#3322
• Minor refactoring / documentation to the backtrace code dlang/druntime#3323
• Add experimental support for libunwind-based stack trace printing dlang/druntime#3321
• backtrace: Look for _d_throwdwarf instead of skipping frames dlang/druntime#3328 (not merged yet)
• backtrace: Get rid of traceHandlerOpApplyImpl2 redundancy dlang/druntime#3329
• core.internal.backtrace.libunwind: Tiny fixup for Android i686 dlang/druntime#3331

[kinke]

Thanks! Had to follow-up with a little fixup:

diff --git a/src/core/internal/backtrace/dwarf.d b/src/core/internal/backtrace/dwarf.d
index 76d5b4d3..4775c61b 100644
--- a/src/core/internal/backtrace/dwarf.d
+++ b/src/core/internal/backtrace/dwarf.d
@@ -52,6 +52,13 @@ module core.internal.backtrace.dwarf;
 import core.internal.execinfo;
 import core.internal.string;

+version (DRuntime_Use_Libunwind)
+    private enum hasLibunwind = true;
+else
+    private enum hasLibunwind = false;
+
+static if (hasExecinfo || hasLibunwind):
+
 version (OSX)
     version = Darwin;
 else version (iOS)
@@ -165,6 +172,9 @@ static if (hasExecinfo)
         import core.stdc.stdio : snprintf;
         import core.sys.posix.stdlib : free;

+        const char** frameList = backtrace_symbols(callstack.ptr, cast(int) callstack.length);
+        scope(exit) free(cast(void*) frameList);
+
         auto image = Image.openSelf();

         // find address -> file, line mapping using dwarf debug_line
diff --git a/src/core/internal/backtrace/libunwind.d b/src/core/internal/backtrace/libunwind.d
index fff25bb1..b0fd6ec2 100644
--- a/src/core/internal/backtrace/libunwind.d
+++ b/src/core/internal/backtrace/libunwind.d
@@ -181,6 +181,7 @@ else version (AArch64)

         version (linux)
         {
+            import core.stdc.config : c_ulong;
             import core.sys.posix.signal : sigset_t, stack_t;

             // libunwind has some special tweaking to reduce
@@ -226,6 +227,8 @@ else version (ARM)
     }
     else
     {
+        import core.stdc.config : c_ulong;
+
         struct unw_tdep_context_t
         {
             c_ulong[16] regs;

There are still compile errors for Android wrt. missing ucontext_t which could presumably be worked around with a version = DRuntime_Use_LLVM_Libunwind for Android (the NDK comes with a libunwind.a); no idea whether that's a good idea.

Should the compiler predefine DRuntime_Use_Libunwind for musl targets etc.? Otherwise the druntime user might not get what he asks for with a prebuilt druntime compiled with a manual version...

Edit: Oh and does the C compiler need an explicit -lunwind or so with DRuntime_Use_Libunwind, or is that one of the automagic libs?

[Geod24]

Should the compiler predefine DRuntime_Use_Libunwind for musl targets etc.? Otherwise the druntime user might not get what he asks for with a prebuilt druntime compiled with a manual version...

I was planning on patching it in the package until it's fully tested and upstreamed, but that works too. For reference, this is what I have for DMD (it's using the old implementation, before I dropped non-LLVM libunwind). But since Alpine is the main (only?) target for Musl at the moment, might as well do it in LDC.

which could presumably be worked around with a version = DRuntime_Use_LLVM_Libunwind

version = DRuntime_Use_Libunwind. I dropped support for the actual libunwind.

Edit: Oh and does the C compiler need an explicit -lunwind or so with DRuntime_Use_Libunwind, or is that one of the automagic libs?

I don't know. But since I'm patching the config file anyway, I can take care of that when I update the packages.

[kinke]

I dropped support for the actual libunwind.

So what's the DRuntime_Use_LLVM_Libunwind in libunwind.d?

[Geod24]

So what's the DRuntime_Use_LLVM_Libunwind in libunwind.d?

Cherry picked from the wrong branch 🤦
Fixed now, and integrated your diff.

EDIT: Actually just found a small bug, hold on.

Fixed: I needed to add the skip code in traceHandlerOpApply2 too.

[kinke]

Okay, I've pushed 2 new commits.

[Geod24]

@kinke : Thanks! Anything else you need from my side ? Are you interested in upstreaming the additional commits ?

[kinke]

Are you interested in upstreaming the additional commits ?

I'll do it at one point unless you're keen on cherry-picking it yourself. ;)

Anything else you need from my side ?

I'm now predefining the version and linking-lunwind for musl targets in ldc-developers/ldc#3641 - could you please test & verify it for a smooth v1.25.0 on Alpine?