[PATCH] D88541: [AArch64] Don't merge sp decrement into later stores when using WinCFI

[Martin Storsjö via Phabricator via llvm-commits]

mstorsjo created this revision.
mstorsjo added reviewers: efriedma, dnsampaio, rnk, ssijaric, TomTan.
Herald added subscribers: danielkiss, hiraditya, kristof.beyls.
Herald added a project: LLVM.
mstorsjo requested review of this revision.

This matches the corresponding existing case in AArch64LoadStoreOpt::findMatchingUpdateInsnForward. (The fixme comment is a copy of that one, from D75755 <https://reviews.llvm.org/D75755>.)

Both cases could also be modified to check MBBI->getFlag(FrameSetup/FrameDestroy) instead of forbidding any optimization involving SP, but the effect is probably pretty much the same.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D88541

Files:
  llvm/lib/Target/AArch64/AArch64LoadStoreOptimizer.cpp
  llvm/test/CodeGen/AArch64/arm64-windows-calls.ll


Index: llvm/test/CodeGen/AArch64/arm64-windows-calls.ll
===================================================================
--- llvm/test/CodeGen/AArch64/arm64-windows-calls.ll
+++ llvm/test/CodeGen/AArch64/arm64-windows-calls.ll
@@ -26,7 +26,10 @@
 entry:
 ; FIXME: Missed optimization, the entire SP push/pop could be removed
 ; CHECK-LABEL: f2
-; CHECK:         stp     xzr, xzr, [sp, #-16]!
+; CHECK:         sub     sp, sp, #16
+; CHECK-NEXT:    .seh_stackalloc 16
+; CHECK-NEXT:    .seh_endprologue
+; CHECK-NEXT:    stp     xzr, xzr, [sp]
 ; CHECK-NEXT:    mov     x0, xzr
 ; CHECK-NEXT:    mov     x1, xzr
 ; CHECK-NEXT:    .seh_startepilogue
Index: llvm/lib/Target/AArch64/AArch64LoadStoreOptimizer.cpp
===================================================================
--- llvm/lib/Target/AArch64/AArch64LoadStoreOptimizer.cpp
+++ llvm/lib/Target/AArch64/AArch64LoadStoreOptimizer.cpp
@@ -1857,6 +1857,17 @@
     }
   }
 
+  const bool BaseRegSP = BaseReg == AArch64::SP;
+  if (BaseRegSP) {
+    // FIXME: For now, we always block the optimization over SP in windows
+    // targets as it requires to adjust the unwind/debug info, messing up
+    // the unwind info can actually cause a miscompile.
+    const MCAsmInfo *MAI = I->getMF()->getTarget().getMCAsmInfo();
+    if (MAI->usesWindowsCFI() &&
+        I->getMF()->getFunction().needsUnwindTableEntry())
+      return E;
+  }
+
   // Track which register units have been modified and used between the first
   // insn (inclusive) and the second insn.
   ModifiedRegUnits.clear();


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D88541.295200.patch
Type: text/x-patch
Size: 1549 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20200930/cbb6a3a6/attachment.bin>