[llvm] [X86] Correct 32-bit immediate assertion and fix 64-bit lowering for huge frame offsets (PR #123872)

[Wesley Wiser via llvm-commits]

================
@@ -958,11 +959,37 @@ X86RegisterInfo::eliminateFrameIndex(MachineBasicBlock::iterator II,
   }
 
   if (MI.getOperand(FIOperandNum+3).isImm()) {
-    // Offset is a 32-bit integer.
-    int Imm = (int)(MI.getOperand(FIOperandNum + 3).getImm());
-    int Offset = FIOffset + Imm;
-    assert((!Is64Bit || isInt<32>((long long)FIOffset + Imm)) &&
-           "Requesting 64-bit offset in 32-bit immediate!");
+    int64_t Imm = MI.getOperand(FIOperandNum + 3).getImm();
+    int64_t Offset = FIOffset + Imm;
+    bool FitsIn32Bits = isInt<32>(Offset);
+    // If the offset will not fit in a 32-bit displacement, then for 64-bit
+    // targets, scavenge a register to hold it. Otherwise...
+    if (Is64Bit && !FitsIn32Bits) {
+      assert(RS && "RegisterScavenger was NULL");
+      const X86InstrInfo *TII = MF.getSubtarget<X86Subtarget>().getInstrInfo();
+      const DebugLoc &DL = MI.getDebugLoc();
+
+      RS->enterBasicBlockEnd(MBB);
+      RS->backward(std::next(II));
+
+      Register ScratchReg = RS->scavengeRegisterBackwards(
+          X86::GR64RegClass, II, /*RestoreAfter=*/false, /*SPAdj=*/0,
+          /*AllowSpill=*/true);
+      assert(ScratchReg != 0 && "scratch reg was 0");
+      RS->setRegUsed(ScratchReg);
+
+      BuildMI(MBB, II, DL, TII->get(X86::MOV64ri), ScratchReg).addImm(Offset);
+
+      MI.getOperand(FIOperandNum + 3).setImm(0);
+      MI.getOperand(FIOperandNum + 2).setReg(ScratchReg);
+
+      return false;
+    }
+
+    // ... for 32-bit targets, this is a bug!
+    if (!Is64Bit && !FitsIn32Bits)
+      MI.emitGenericError(("64-bit offset calculated but target is 32-bit"));
+
----------------
wesleywiser wrote:

I think the issue is that while we report an error here, we leave the operand in the instruction which the verifier then reports. What's the right approach to resolve that? 

I guess we could replace the overside operand with a saturated value or `0` or something. Or perhaps replace the entire instruction with a trap, etc. 

https://github.com/llvm/llvm-project/pull/123872