[PATCH] D125876: [MC] [Win64EH] Don't produce packed ARM64 unwind info with homed parameters

[Martin Storsjö via Phabricator via llvm-commits]

mstorsjo created this revision.
mstorsjo added reviewers: efriedma, zzheng.
Herald added subscribers: hiraditya, kristof.beyls.
Herald added a project: All.
mstorsjo requested review of this revision.
Herald added a project: LLVM.

There's an inconsistency regarding the epilogs of packed ARM64
unwind info with homed parameters; according to the documentation
(and according to common sense), the epilog wouldn't have a series
of nop instructions matching the stp x0-x7 in the prolog - however
in practice, RtlVirtualUnwind still seems to behave as if the epilog
does have the mirrored nops from the prolog.

In practice, MSVC doesn't seem to produce packed unwind info with
homed parameters, which might be why this inconsistency hasn't
been noticed.

Thus, to play it safe, avoid creating such packed unwind info with
homed parameters. (LLVM's current behaviour matches the current
runtime behaviour of RtlVirtualUnwind, but if it later is bug fixed
to match the documentation, such unwind information would be
incorrect.)

See https://github.com/llvm/llvm-project/issues/54879 for further
discussion on the matter.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D125876

Files:
  llvm/lib/MC/MCWin64EH.cpp
  llvm/test/MC/AArch64/seh-packed-unwind.s


Index: llvm/test/MC/AArch64/seh-packed-unwind.s
===================================================================
--- llvm/test/MC/AArch64/seh-packed-unwind.s
+++ llvm/test/MC/AArch64/seh-packed-unwind.s
@@ -81,25 +81,10 @@
 // CHECK-NEXT:     ]
 // CHECK-NEXT:   }
 // CHECK-NEXT:   RuntimeFunction {
-// CHECK-NEXT:     Function: func5
-// CHECK-NEXT:     Fragment: No
-// CHECK-NEXT:     FunctionLength: 56
-// CHECK-NEXT:     RegF: 0
-// CHECK-NEXT:     RegI: 1
-// CHECK-NEXT:     HomedParameters: Yes
-// CHECK-NEXT:     CR: 0
-// CHECK-NEXT:     FrameSize: 112
-// CHECK-NEXT:     Prologue [
-// CHECK-NEXT:       sub sp, sp, #32
-// CHECK-NEXT:       stp x6, x7, [sp, #64]
-// CHECK-NEXT:       stp x4, x5, [sp, #48]
-// CHECK-NEXT:       stp x2, x3, [sp, #32]
-// CHECK-NEXT:       stp x0, x1, [sp, #16]
-// CHECK-NEXT:       str x19, [sp, #-80]!
-// CHECK-NEXT:       end
-// CHECK-NEXT:     ]
-// CHECK-NEXT:   }
-// CHECK-NEXT:   RuntimeFunction {
+// CHECK-NEXT:     Function: notpacked_func5
+// CHECK-NEXT:     ExceptionRecord:
+// CHECK-NEXT:     ExceptionData {
+// CHECK:        RuntimeFunction {
 // CHECK-NEXT:     Function: func6
 // CHECK-NEXT:     Fragment: No
 // CHECK-NEXT:     FunctionLength: 24
@@ -441,8 +426,8 @@
     ret
     .seh_endproc
 
-func5:
-    .seh_proc func5
+notpacked_func5:
+    .seh_proc notpacked_func5
     str x19, [sp, #-80]!
     .seh_save_reg_x x19, 80
     stp x0,  x1,  [sp, #16]
Index: llvm/lib/MC/MCWin64EH.cpp
===================================================================
--- llvm/lib/MC/MCWin64EH.cpp
+++ llvm/lib/MC/MCWin64EH.cpp
@@ -860,6 +860,16 @@
   if (Nops != 0 && Nops != 4)
     return false;
   int H = Nops == 4;
+  // There's an inconsistency regarding packed unwind info with homed
+  // parameters; according to the documentation, the epilog shouldn't have
+  // the same corresponding nops (and thus, to set the H bit, we should
+  // require an epilog which isn't exactly symmetrical - we shouldn't accept
+  // an exact mirrored epilog for those cases), but in practice,
+  // RtlVirtualUnwind behaves as if it does expect the epilogue to contain
+  // the same nops. See https://github.com/llvm/llvm-project/issues/54879.
+  // To play it safe, don't produce packed unwind info with homed parameters.
+  if (H)
+    return false;
   int IntSZ = 8 * RegI;
   if (StandaloneLR)
     IntSZ += 8;


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D125876.430332.patch
Type: text/x-patch
Size: 2377 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20220518/c2617d6e/attachment-0001.bin>