[AArch64] Place the first ldp at the end when ReverseCSRRestoreSeq is true

Put the first ldp at the end, so that the load-store optimizer can run
and merge the ldp and the add into a post-index ldp.

This didn't work in case no frame was needed and resulted in code size
regressions.

llvm-svn: 331044

Author: francisvm

llvm/lib/Target/AArch64/AArch64FrameLowering.cpp

@@ -514,6 +514,38 @@ static void fixupCalleeSaveRestoreStackOffset(MachineInstr &MI,
   OffsetOpnd.setImm(OffsetOpnd.getImm() + LocalStackSize / 8);
 }
 
+static void adaptForLdStOpt(MachineBasicBlock &MBB,
+                            MachineBasicBlock::iterator FirstSPPopI,
+                            MachineBasicBlock::iterator LastPopI) {
+  // Sometimes (when we restore in the same order as we save), we can end up
+  // with code like this:
+  //
+  // ldp      x26, x25, [sp]
+  // ldp      x24, x23, [sp, #16]
+  // ldp      x22, x21, [sp, #32]
+  // ldp      x20, x19, [sp, #48]
+  // add      sp, sp, #64
+  //
+  // In this case, it is always better to put the first ldp at the end, so
+  // that the load-store optimizer can run and merge the ldp and the add into
+  // a post-index ldp.
+  // If we managed to grab the first pop instruction, move it to the end.
+  if (ReverseCSRRestoreSeq)
+    MBB.splice(FirstSPPopI, &MBB, LastPopI);
+  // We should end up with something like this now:
+  //
+  // ldp      x24, x23, [sp, #16]
+  // ldp      x22, x21, [sp, #32]
+  // ldp      x20, x19, [sp, #48]
+  // ldp      x26, x25, [sp]
+  // add      sp, sp, #64
+  //
+  // and the load-store optimizer can merge the last two instructions into:
+  //
+  // ldp      x26, x25, [sp], #64
+  //
+}
+
 void AArch64FrameLowering::emitPrologue(MachineFunction &MF,
                                         MachineBasicBlock &MBB) const {
   MachineBasicBlock::iterator MBBI = MBB.begin();
@@ -930,12 +962,20 @@ void AArch64FrameLowering::emitEpilogue(MachineFunction &MF,
     int StackRestoreBytes = RedZone ? 0 : NumBytes;
     if (NoCalleeSaveRestore)
       StackRestoreBytes += AfterCSRPopSize;
-    emitFrameOffset(MBB, LastPopI, DL, AArch64::SP, AArch64::SP,
-                    StackRestoreBytes, TII, MachineInstr::FrameDestroy);
+
     // If we were able to combine the local stack pop with the argument pop,
     // then we're done.
-    if (NoCalleeSaveRestore || AfterCSRPopSize == 0)
+    bool Done = NoCalleeSaveRestore || AfterCSRPopSize == 0;
+
+    // If we're done after this, make sure to help the load store optimizer.
+    if (Done)
+      adaptForLdStOpt(MBB, MBB.getFirstTerminator(), LastPopI);
+
+    emitFrameOffset(MBB, LastPopI, DL, AArch64::SP, AArch64::SP,
+                    StackRestoreBytes, TII, MachineInstr::FrameDestroy);
+    if (Done)
       return;
+
     NumBytes = 0;
   }
 
@@ -967,33 +1007,8 @@ void AArch64FrameLowering::emitEpilogue(MachineFunction &MF,
       FirstSPPopI = Prev;
     }
 
-    // Sometimes (when we restore in the same order as we save), we can end up
-    // with code like this:
-    //
-    // ldp      x26, x25, [sp]
-    // ldp      x24, x23, [sp, #16]
-    // ldp      x22, x21, [sp, #32]
-    // ldp      x20, x19, [sp, #48]
-    // add      sp, sp, #64
-    //
-    // In this case, it is always better to put the first ldp at the end, so
-    // that the load-store optimizer can run and merge the ldp and the add into
-    // a post-index ldp.
-    // If we managed to grab the first pop instruction, move it to the end.
-    if (LastPopI != Begin)
-      MBB.splice(FirstSPPopI, &MBB, LastPopI);
-    // We should end up with something like this now:
-    //
-    // ldp      x24, x23, [sp, #16]
-    // ldp      x22, x21, [sp, #32]
-    // ldp      x20, x19, [sp, #48]
-    // ldp      x26, x25, [sp]
-    // add      sp, sp, #64
-    //
-    // and the load-store optimizer can merge the last two instructions into:
-    //
-    // ldp      x26, x25, [sp], #64
-    //
+    adaptForLdStOpt(MBB, FirstSPPopI, LastPopI);
+
     emitFrameOffset(MBB, FirstSPPopI, DL, AArch64::SP, AArch64::SP,
                     AfterCSRPopSize, TII, MachineInstr::FrameDestroy);
   }

llvm/test/CodeGen/AArch64/reverse-csr-restore-seq.mir

@@ -7,6 +7,8 @@
 
   define void @bar() nounwind { entry: unreachable }
 
+  define void @baz() nounwind { entry: unreachable }
+
 ...
 ---
 name:            foo
@@ -71,3 +73,34 @@ body:             |
   ; CHECK-NEXT: $sp = frame-destroy ADDXri $sp, 80, 0
     RET_ReallyLR
 ...
+---
+# Check that the load from the offset 0 is moved at the end even when hasFP is
+# false.
+name:            baz
+# CHECK-LABEL: name: baz
+alignment:       2
+tracksRegLiveness: true
+frameInfo:
+  adjustsStack:    true
+  hasCalls:        true
+body:             |
+  bb.0:
+    successors: %bb.1
+
+    $x0 = IMPLICIT_DEF
+    $x20 = IMPLICIT_DEF
+    $x21 = IMPLICIT_DEF
+
+    ADJCALLSTACKDOWN 0, 0, implicit-def dead $sp, implicit $sp
+    BL @foo, csr_aarch64_aapcs, implicit-def dead $lr, implicit $sp, implicit $x0, implicit-def $sp, implicit-def $x0
+    ADJCALLSTACKUP 0, 0, implicit-def dead $sp, implicit $sp
+    B %bb.1
+
+  bb.1:
+   ; CHECK: $x20, $lr = frame-destroy LDPXi $sp, 2
+   ; BEFORELDSTOPT-NEXT: $x21 = frame-destroy LDRXui $sp, 0
+   ; BEFORELDSTOPT-NEXT: $sp = frame-destroy ADDXri $sp, 32, 0
+
+   ; AFTERLDSTOPT-NEXT: early-clobber $sp, $x21 = frame-destroy LDRXpost $sp, 32
+    RET_ReallyLR
+...