Preserve BasePtr for LEA64_32r

[hjl.tools]

Summary

When replacing FrameIndex with BasePtr, we must preserve BasePtr for
LEA64_32r since BasePtr is used later for stack adjustment if it is
the same as StackPtr.

[mkuper]

lib/Target/X86/X86RegisterInfo.cpp

638

Maybe something like:

unsigned MachineBasePtr = BasePtr;
if (Opc == X86::LEA64_32r && X86::GR32RegClass.contains(BasePtr))
  MachineBasePtr = getX86SubSuperRegister(BasePtr, 64);

?

645

As long as you're touching this, can you fix this comment to make sense?

test/CodeGen/X86/x32-movtopush64.ll

1

You shouldn't need the -O2, I think.

[hjl.tools]

Updated with comments from Michael.

[mkuper]

LGTM

[Eugene.Zelenko]

Looks like patch was not committed.

lib/Target/X86/X86RegisterInfo.cpp

if (Opc == TargetOpcode::LOCAL_ESCAPE) {
     Offset = TFI->getFrameIndexReference(MF, FrameIndex, IgnoredFrameReg);
     FI.ChangeToImmediate(Offset);
     return;
   }
 
   // For LEA64_32r when BasePtr is 32-bits (X32) we can use full-size 64-bit
   // register as source operand, semantic is the same and destination is
   // 32-bits. It saves one byte per lea in code since 0x67 prefix is avoided.
+  // Don't change BasePtr since it is used later for stack adjustment.
+  unsigned MachineBasePtr;
   if (Opc == X86::LEA64_32r && X86::GR32RegClass.contains(BasePtr))
-    BasePtr = getX86SubSuperRegister(BasePtr, 64);
+    MachineBasePtr = getX86SubSuperRegister(BasePtr, 64);
+  else
+    MachineBasePtr = BasePtr;
 
   // This must be part of a four operand memory reference.  Replace the
   // FrameIndex with base register with EBP.  Add an offset to the offset.
-  MI.getOperand(FIOperandNum).ChangeToRegister(BasePtr, false);
+  MI.getOperand(FIOperandNum).ChangeToRegister(MachineBasePtr, false);
 
   // Now add the frame object offset to the offset from EBP.
   int FIOffset;
   if (AfterFPPop) {
     // Tail call jmp happens after FP is popped.
     const MachineFrameInfo &MFI = MF.getFrameInfo();
     FIOffset = MFI.getObjectOffset(FrameIndex) - TFI->getOffsetOfLocalArea();
   } else

test/CodeGen/X86/x32-movtopush64.ll

+; RUN: llc < %s -mtriple=x86_64-linux-gnux32 -O2 | FileCheck %s -check-prefix=CHECK
+
+declare void @bar(i32*, i32*, i32*, i32*, i32*, i64*, i32, i32, i32)
+
+; Function Attrs: nounwind uwtable
+define void @foo() {
+entry:
+  %i1 = alloca i32, align 4
+  %i2 = alloca i32, align 4
+  %i3 = alloca i32, align 4
+  %i4 = alloca i32, align 4
+  %i5 = alloca i32, align 4
+  %i6 = alloca i64, align 8
+  %0 = bitcast i32* %i1 to i8*
+  store i32 1, i32* %i1, align 4
+; CHECK: movl $1, 28(%esp)
+  %1 = bitcast i32* %i2 to i8*
+  store i32 2, i32* %i2, align 4
+; CHECK-NEXT: movl $2, 24(%esp)
+  %2 = bitcast i32* %i3 to i8*
+  store i32 3, i32* %i3, align 4
+; CHECK-NEXT: movl $3, 20(%esp)
+  %3 = bitcast i32* %i4 to i8*
+  store i32 4, i32* %i4, align 4
+; CHECK-NEXT: movl $4, 16(%esp)
+  %4 = bitcast i32* %i5 to i8*
+  store i32 5, i32* %i5, align 4
+; CHECK-NEXT: movl $5, 12(%esp)
+  %5 = bitcast i64* %i6 to i8*
+  store i64 6, i64* %i6, align 8
+; CHECK-NEXT: movq $6, 32(%esp)
+; CHECK-NEXT: subl $8, %esp
+; CHECK: leal 36(%rsp), %edi
+; CHECK-NEXT: leal 32(%rsp), %esi
+; CHECK-NEXT: leal 28(%rsp), %edx
+; CHECK-NEXT: leal 24(%rsp), %ecx
+; CHECK-NEXT: leal 20(%rsp), %r8d
+; CHECK-NEXT: leal 40(%rsp), %r9d
+; CHECK: pushq $0
+; CHECK: pushq $0
+; CHECK: pushq $0
+  call void @bar(i32* nonnull %i1, i32* nonnull %i2, i32* nonnull %i3, i32* nonnull %i4, i32* nonnull %i5, i64* nonnull %i6, i32 0, i32 0, i32 0)
+  ret void
+}