[llvm] [llvm][X86] Fix merging of large sp updates (PR #125007)

[via llvm-commits]

================
@@ -415,27 +415,38 @@ int X86FrameLowering::mergeSPUpdates(MachineBasicBlock &MBB,
   if (doMergeWithPrevious && PI != MBB.begin() && PI->isCFIInstruction())
     PI = std::prev(PI);
 
-  unsigned Opc = PI->getOpcode();
-  int Offset = 0;
-
-  if ((Opc == X86::ADD64ri32 || Opc == X86::ADD32ri) &&
-      PI->getOperand(0).getReg() == StackPtr) {
-    assert(PI->getOperand(1).getReg() == StackPtr);
-    Offset = PI->getOperand(2).getImm();
-  } else if ((Opc == X86::LEA32r || Opc == X86::LEA64_32r) &&
-             PI->getOperand(0).getReg() == StackPtr &&
-             PI->getOperand(1).getReg() == StackPtr &&
-             PI->getOperand(2).getImm() == 1 &&
-             PI->getOperand(3).getReg() == X86::NoRegister &&
-             PI->getOperand(5).getReg() == X86::NoRegister) {
-    // For LEAs we have: def = lea SP, FI, noreg, Offset, noreg.
-    Offset = PI->getOperand(4).getImm();
-  } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB32ri) &&
-             PI->getOperand(0).getReg() == StackPtr) {
-    assert(PI->getOperand(1).getReg() == StackPtr);
-    Offset = -PI->getOperand(2).getImm();
-  } else
-    return 0;
+  int64_t Offset = 0;
+  for (;;) {
+    unsigned Opc = PI->getOpcode();
+
+    if ((Opc == X86::ADD64ri32 || Opc == X86::ADD32ri) &&
+        PI->getOperand(0).getReg() == StackPtr) {
+      assert(PI->getOperand(1).getReg() == StackPtr);
+      Offset = PI->getOperand(2).getImm();
+    } else if ((Opc == X86::LEA32r || Opc == X86::LEA64_32r) &&
+               PI->getOperand(0).getReg() == StackPtr &&
+               PI->getOperand(1).getReg() == StackPtr &&
+               PI->getOperand(2).getImm() == 1 &&
+               PI->getOperand(3).getReg() == X86::NoRegister &&
+               PI->getOperand(5).getReg() == X86::NoRegister) {
+      // For LEAs we have: def = lea SP, FI, noreg, Offset, noreg.
+      Offset = PI->getOperand(4).getImm();
+    } else if ((Opc == X86::SUB64ri32 || Opc == X86::SUB32ri) &&
+               PI->getOperand(0).getReg() == StackPtr) {
+      assert(PI->getOperand(1).getReg() == StackPtr);
+      Offset = -PI->getOperand(2).getImm();
+    } else
+      return 0;
+
+    constexpr int64_t Chunk = (1LL << 31) - 1;
+    if (Offset < Chunk)
----------------
macurtis-amd wrote:

Just realized that this needs to take into account the offset that the caller wants to add.
Working on an updated patch.

https://github.com/llvm/llvm-project/pull/125007