Lower thumbv4t & thumbv5 lo->lo copies through a push-pop sequence
authorJonathan Roelofs <jonathan@codesourcery.com>
Wed, 20 Aug 2014 23:38:50 +0000 (23:38 +0000)
committerJonathan Roelofs <jonathan@codesourcery.com>
Wed, 20 Aug 2014 23:38:50 +0000 (23:38 +0000)
On pre-v6 hardware, 'MOV lo, lo' gives undefined results, so such copies need to
be avoided. This patch trades simplicity for implementation time at the expense
of performance... As they say: correctness first, then performance.

See http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-August/075998.html for a few
ideas on how to make this better.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@216138 91177308-0d34-0410-b5e6-96231b3b80d8

lib/Target/ARM/Thumb1InstrInfo.cpp
test/CodeGen/ARM/atomic-cmpxchg.ll
test/CodeGen/Thumb/copy_thumb.ll [new file with mode: 0644]

index abde25081b00e6f94b722eb63e500ee87db3f418..8ea912e2703936cfe9d4742d4f887863305bd390 100644 (file)
@@ -11,6 +11,7 @@
 //
 //===----------------------------------------------------------------------===//
 
 //
 //===----------------------------------------------------------------------===//
 
+#include "ARMSubtarget.h"
 #include "Thumb1InstrInfo.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
 #include "Thumb1InstrInfo.h"
 #include "llvm/CodeGen/MachineFrameInfo.h"
 #include "llvm/CodeGen/MachineInstrBuilder.h"
@@ -41,10 +42,30 @@ void Thumb1InstrInfo::copyPhysReg(MachineBasicBlock &MBB,
                                   MachineBasicBlock::iterator I, DebugLoc DL,
                                   unsigned DestReg, unsigned SrcReg,
                                   bool KillSrc) const {
                                   MachineBasicBlock::iterator I, DebugLoc DL,
                                   unsigned DestReg, unsigned SrcReg,
                                   bool KillSrc) const {
-  AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::tMOVr), DestReg)
-    .addReg(SrcReg, getKillRegState(KillSrc)));
+  // Need to check the arch.
+  MachineFunction &MF = *MBB.getParent();
+  const ARMSubtarget &st = MF.getTarget().getSubtarget<ARMSubtarget>();
+
   assert(ARM::GPRRegClass.contains(DestReg, SrcReg) &&
          "Thumb1 can only copy GPR registers");
   assert(ARM::GPRRegClass.contains(DestReg, SrcReg) &&
          "Thumb1 can only copy GPR registers");
+
+  if (st.hasV6Ops() || ARM::hGPRRegClass.contains(SrcReg)
+      || !ARM::tGPRRegClass.contains(DestReg))
+    AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::tMOVr), DestReg)
+      .addReg(SrcReg, getKillRegState(KillSrc)));
+  else {
+    // FIXME: The performance consequences of this are going to be atrocious.
+    // Some things to try that should be better:
+    //   * 'mov hi, $src; mov $dst, hi', with hi as either r10 or r11
+    //   * 'movs $dst, $src' if cpsr isn't live
+    // See: http://lists.cs.uiuc.edu/pipermail/llvmdev/2014-August/075998.html
+
+    // 'MOV lo, lo' is unpredictable on < v6, so use the stack to do it
+    AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::tPUSH)))
+      .addReg(SrcReg, getKillRegState(KillSrc));
+    AddDefaultPred(BuildMI(MBB, I, DL, get(ARM::tPOP)))
+      .addReg(DestReg, getDefRegState(true));
+  }
 }
 
 void Thumb1InstrInfo::
 }
 
 void Thumb1InstrInfo::
index 4b79fa25145bf03d16ed3b36c9f7920838bbd2f4..84790be6d605943c5fa48b1150b51f4fc3158624 100644 (file)
@@ -20,12 +20,15 @@ entry:
 
 ; CHECK-THUMB-LABEL: test_cmpxchg_res_i8
 ; CHECK-THUMB: bl __sync_val_compare_and_swap_1
 
 ; CHECK-THUMB-LABEL: test_cmpxchg_res_i8
 ; CHECK-THUMB: bl __sync_val_compare_and_swap_1
-; CHECK-THUMB: mov [[R1:r[0-9]+]], r0
+; CHECK-THUMB-NOT: mov [[R1:r[0-7]]], r0
+; CHECK-THUMB: push  {r0}
+; CHECK-THUMB: pop {[[R1:r[0-7]]]}
 ; CHECK-THUMB: movs r0, #1
 ; CHECK-THUMB: movs [[R2:r[0-9]+]], #0
 ; CHECK-THUMB: cmp [[R1]], {{r[0-9]+}}
 ; CHECK-THU<B: beq
 ; CHECK-THUMB: movs r0, #1
 ; CHECK-THUMB: movs [[R2:r[0-9]+]], #0
 ; CHECK-THUMB: cmp [[R1]], {{r[0-9]+}}
 ; CHECK-THU<B: beq
-; CHECK-THUMB: mov r0, [[R2]]
+; CHECK-THUMB: push  {[[R2]]}
+; CHECK-THUMB: pop {r0}
 
 ; CHECK-ARMV7-LABEL: test_cmpxchg_res_i8
 ; CHECK-ARMV7: ldrexb [[R3:r[0-9]+]], [r0]
 
 ; CHECK-ARMV7-LABEL: test_cmpxchg_res_i8
 ; CHECK-ARMV7: ldrexb [[R3:r[0-9]+]], [r0]
diff --git a/test/CodeGen/Thumb/copy_thumb.ll b/test/CodeGen/Thumb/copy_thumb.ll
new file mode 100644 (file)
index 0000000..528f54b
--- /dev/null
@@ -0,0 +1,38 @@
+; RUN: llc -mtriple=armv4-none--eabi < %s | FileCheck %s --check-prefix=CHECK-LOLOMOV
+; RUN: llc -mtriple=armv4t-none--eabi < %s | FileCheck %s --check-prefix=CHECK-LOLOMOV
+; RUN: llc -mtriple=armv5-none--eabi < %s | FileCheck %s --check-prefix=CHECK-LOLOMOV
+; RUN: llc -mtriple=armv6-none--eabi < %s | FileCheck %s --check-prefix=CHECK-LOLOMOV
+; RUN: llc -mtriple=armv7-none--eabi < %s | FileCheck %s --check-prefix=CHECK-LOLOMOV
+; RUN: llc -mtriple=thumbv6-none--eabi < %s | FileCheck %s --check-prefix=CHECK-LOLOMOV
+; RUN: llc -mtriple=thumbv7-none--eabi < %s | FileCheck %s --check-prefix=CHECK-LOLOMOV
+; CHECK-LOLOMOV-LABEL:  foo
+; CHECK-LOLOMOV:        mov [[TMP:r[0-7]]], [[SRC1:r[01]]]
+; CHECK-LOLOMOV-NEXT:   mov [[SRC1]], [[SRC2:r[01]]]
+; CHECK-LOLOMOV-NEXT:   mov [[SRC2]], [[TMP]]
+; CHECK-LOLOMOV-LABEL:  bar
+; CHECK-LOLOMOV-LABEL:  fnend
+; 
+; 'MOV lo, lo' in Thumb mode produces undefined results on pre-v6 hardware
+; RUN: llc -mtriple=thumbv4t-none--eabi < %s | FileCheck %s --check-prefix=CHECK-NOLOLOMOV
+; RUN: llc -mtriple=thumbv5-none--eabi < %s | FileCheck %s --check-prefix=CHECK-NOLOLOMOV
+; CHECK-NOLOLOMOV-LABEL: foo
+; CHECK-NOLOLOMOV-NOT:   mov [[TMP:r[0-7]]], [[SRC1:r[01]]]
+; CHECK-NOLOLOMOV:       push  {[[SRC1:r[01]]]}
+; CHECK-NOLOLOMOV-NEXT:  pop {[[TMP:r[0-7]]]}
+; CHECK-NOLOLOMOV-NOT:   mov [[TMP:r[0-7]]], [[SRC1:r[01]]]
+; CHECK-NOLOLOMOV:       push  {[[SRC2:r[01]]]}
+; CHECK-NOLOLOMOV-NEXT:  pop {[[SRC1]]}
+; CHECK-NOLOLOMOV-NOT:   mov [[TMP:r[0-7]]], [[SRC1:r[01]]]
+; CHECK-NOLOLOMOV:       push  {[[TMP]]}
+; CHECK-NOLOLOMOV-NEXT:  pop {[[SRC2]]}
+; CHECK-NOLOLOMOV-LABEL: bar
+; CHECK-NOLOLOMOV-LABEL: fnend
+
+declare void @bar(i32, i32)
+
+define void @foo(i32 %a, i32 %b) {
+entry:
+  call void @bar(i32 %b, i32 %a);
+  ret void
+}
+