[llvm-branch-commits] [llvm] [X86] Avoid generating nested CALLSEQ for TLS pointer function arguments (PR #106965)

[via llvm-branch-commits]

llvmbot wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-x86

Author: Fabian Ritter (ritter-x2a)

<details>
<summary>Changes</summary>

When a pointer to thread-local storage is passed in a function call,
ISel first lowers the call and wraps the resulting code in CALLSEQ
markers. Afterwards, to compute the pointer to TLS, a call to retrieve
the TLS base address is generated and then wrapped in a set of CALLSEQ
markers. If the latter call is inserted into the call sequence of the
former call, this leads to nested call frames, which are illegal and
lead to errors in the machine verifier.

This patch avoids surrounding the call to compute the TLS base address
in CALLSEQ markers if it is already surrounded by such markers. It
relies on zero-sized call frames being represented in the call frame
size info stored in the MachineBBs.

Fixes #<!-- -->45574 and #<!-- -->98042.

---
Full diff: https://github.com/llvm/llvm-project/pull/106965.diff


2 Files Affected:

- (modified) llvm/lib/Target/X86/X86ISelLowering.cpp (+7) 
- (added) llvm/test/CodeGen/X86/tls-function-argument.ll (+17) 


``````````diff
diff --git a/llvm/lib/Target/X86/X86ISelLowering.cpp b/llvm/lib/Target/X86/X86ISelLowering.cpp
index bbee0af109c74b..bf9777888df831 100644
--- a/llvm/lib/Target/X86/X86ISelLowering.cpp
+++ b/llvm/lib/Target/X86/X86ISelLowering.cpp
@@ -35593,6 +35593,13 @@ X86TargetLowering::EmitLoweredTLSAddr(MachineInstr &MI,
   // inside MC, therefore without the two markers shrink-wrapping
   // may push the prologue/epilogue pass them.
   const TargetInstrInfo &TII = *Subtarget.getInstrInfo();
+
+  // Do not introduce CALLSEQ markers if we are already in a call sequence.
+  // Nested call sequences are not allowed and cause errors in the machine
+  // verifier.
+  if (TII.getCallFrameSizeAt(MI).has_value())
+    return BB;
+
   const MIMetadata MIMD(MI);
   MachineFunction &MF = *BB->getParent();
 
diff --git a/llvm/test/CodeGen/X86/tls-function-argument.ll b/llvm/test/CodeGen/X86/tls-function-argument.ll
new file mode 100644
index 00000000000000..ec2d664fc6b96f
--- /dev/null
+++ b/llvm/test/CodeGen/X86/tls-function-argument.ll
@@ -0,0 +1,17 @@
+; RUN: llc -verify-machineinstrs < %s -relocation-model=pic
+
+target datalayout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-unknown-linux-gnu"
+
+; Passing a pointer to thread-local storage to a function can be problematic
+; since computing such addresses requires a function call that is introduced
+; very late in instruction selection. We need to ensure that we don't introduce
+; nested call sequence markers if this function call happens in a call sequence.
+
+ at TLS = internal thread_local global i64 zeroinitializer, align 8
+declare void @bar(ptr)
+define internal void @foo() {
+call void @bar(ptr @TLS)
+call void @bar(ptr @TLS)
+ret void
+}
\ No newline at end of file

``````````

</details>


https://github.com/llvm/llvm-project/pull/106965