ARM: Thumb2 LDRD/STRD supports independent input/output regs

The existing code would unnecessarily break LDRD/STRD apart with
non-adjacent registers, on thumb2 this is not necessary.

Ideally on thumb2 we shouldn't match for ldrd/strd pre-regalloc anymore
as there is not reason to set register hints anymore, changing that is
something for a future patch however.

Differential Revision: http://reviews.llvm.org/D9694

Recommiting after the revert in r238821, the buildbot still failed with
the patch removed so there seems to be another reason for the breakage.

llvm-svn: 238935

diff --git a/llvm/lib/Target/ARM/ARMLoadStoreOptimizer.cpp b/llvm/lib/Target/ARM/ARMLoadStoreOptimizer.cpp
index 10787b1..8570ecd 100644 (file)
--- a/llvm/lib/Target/ARM/ARMLoadStoreOptimizer.cpp
+++ b/llvm/lib/Target/ARM/ARMLoadStoreOptimizer.cpp
@@ -1470,8 +1470,7 @@ bool ARMLoadStoreOpt::FixInvalidRegPairOp(MachineBasicBlock &MBB,
                                           MachineBasicBlock::iterator &MBBI) {
   MachineInstr *MI = &*MBBI;
   unsigned Opcode = MI->getOpcode();
-  if (Opcode == ARM::LDRD || Opcode == ARM::STRD ||
-      Opcode == ARM::t2LDRDi8 || Opcode == ARM::t2STRDi8) {
+  if (Opcode == ARM::LDRD || Opcode == ARM::STRD) {
     const MachineOperand &BaseOp = MI->getOperand(2);
     unsigned BaseReg = BaseOp.getReg();
     unsigned EvenReg = MI->getOperand(0).getReg();
@@ -1949,10 +1948,11 @@ static void concatenateMemOperands(MachineInstr *MI, MachineInstr *Op0,
 
 bool
 ARMPreAllocLoadStoreOpt::CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1,
-                                          DebugLoc &dl,
-                                          unsigned &NewOpc, unsigned &EvenReg,
-                                          unsigned &OddReg, unsigned &BaseReg,
-                                          int &Offset, unsigned &PredReg,
+                                          DebugLoc &dl, unsigned &NewOpc,
+                                          unsigned &FirstReg,
+                                          unsigned &SecondReg,
+                                          unsigned &BaseReg, int &Offset,
+                                          unsigned &PredReg,
                                           ARMCC::CondCodes &Pred,
                                           bool &isT2) {
   // Make sure we're allowed to generate LDRD/STRD.
@@ -2011,9 +2011,9 @@ ARMPreAllocLoadStoreOpt::CanFormLdStDWord(MachineInstr *Op0, MachineInstr *Op1,
       return false;
     Offset = ARM_AM::getAM3Opc(AddSub, OffImm);
   }
-  EvenReg = Op0->getOperand(0).getReg();
-  OddReg  = Op1->getOperand(0).getReg();
-  if (EvenReg == OddReg)
+  FirstReg = Op0->getOperand(0).getReg();
+  SecondReg = Op1->getOperand(0).getReg();
+  if (FirstReg == SecondReg)
     return false;
   BaseReg = Op0->getOperand(1).getReg();
   Pred = getInstrPredicate(Op0, PredReg);
@@ -2109,7 +2109,7 @@ bool ARMPreAllocLoadStoreOpt::RescheduleOps(MachineBasicBlock *MBB,
         // to try to allocate a pair of registers that can form register pairs.
         MachineInstr *Op0 = Ops.back();
         MachineInstr *Op1 = Ops[Ops.size()-2];
-        unsigned EvenReg = 0, OddReg = 0;
+        unsigned FirstReg = 0, SecondReg = 0;
         unsigned BaseReg = 0, PredReg = 0;
         ARMCC::CondCodes Pred = ARMCC::AL;
         bool isT2 = false;
@@ -2117,21 +2117,21 @@ bool ARMPreAllocLoadStoreOpt::RescheduleOps(MachineBasicBlock *MBB,
         int Offset = 0;
         DebugLoc dl;
         if (NumMove == 2 && CanFormLdStDWord(Op0, Op1, dl, NewOpc,
-                                             EvenReg, OddReg, BaseReg,
+                                             FirstReg, SecondReg, BaseReg,
                                              Offset, PredReg, Pred, isT2)) {
           Ops.pop_back();
           Ops.pop_back();
 
           const MCInstrDesc &MCID = TII->get(NewOpc);
           const TargetRegisterClass *TRC = TII->getRegClass(MCID, 0, TRI, *MF);
-          MRI->constrainRegClass(EvenReg, TRC);
-          MRI->constrainRegClass(OddReg, TRC);
+          MRI->constrainRegClass(FirstReg, TRC);
+          MRI->constrainRegClass(SecondReg, TRC);
 
           // Form the pair instruction.
           if (isLd) {
             MachineInstrBuilder MIB = BuildMI(*MBB, InsertPos, dl, MCID)
-              .addReg(EvenReg, RegState::Define)
-              .addReg(OddReg, RegState::Define)
+              .addReg(FirstReg, RegState::Define)
+              .addReg(SecondReg, RegState::Define)
               .addReg(BaseReg);
             // FIXME: We're converting from LDRi12 to an insn that still
             // uses addrmode2, so we need an explicit offset reg. It should
@@ -2144,8 +2144,8 @@ bool ARMPreAllocLoadStoreOpt::RescheduleOps(MachineBasicBlock *MBB,
             ++NumLDRDFormed;
           } else {
             MachineInstrBuilder MIB = BuildMI(*MBB, InsertPos, dl, MCID)
-              .addReg(EvenReg)
-              .addReg(OddReg)
+              .addReg(FirstReg)
+              .addReg(SecondReg)
               .addReg(BaseReg);
             // FIXME: We're converting from LDRi12 to an insn that still
             // uses addrmode2, so we need an explicit offset reg. It should
@@ -2160,9 +2160,11 @@ bool ARMPreAllocLoadStoreOpt::RescheduleOps(MachineBasicBlock *MBB,
           MBB->erase(Op0);
           MBB->erase(Op1);
 
-          // Add register allocation hints to form register pairs.
-          MRI->setRegAllocationHint(EvenReg, ARMRI::RegPairEven, OddReg);
-          MRI->setRegAllocationHint(OddReg,  ARMRI::RegPairOdd, EvenReg);
+          if (!isT2) {
+            // Add register allocation hints to form register pairs.
+            MRI->setRegAllocationHint(FirstReg, ARMRI::RegPairEven, SecondReg);
+            MRI->setRegAllocationHint(SecondReg,  ARMRI::RegPairOdd, FirstReg);
+          }
         } else {
           for (unsigned i = 0; i != NumMove; ++i) {
             MachineInstr *Op = Ops.back();

diff --git a/llvm/test/CodeGen/ARM/atomic-ops-v8.ll b/llvm/test/CodeGen/ARM/atomic-ops-v8.ll
index db5007b..86287c1 100644 (file)
--- a/llvm/test/CodeGen/ARM/atomic-ops-v8.ll
+++ b/llvm/test/CodeGen/ARM/atomic-ops-v8.ll
@@ -664,7 +664,7 @@ define void @test_atomic_load_min_i64(i64 %offset) nounwind {
 ; CHECK: movt r[[ADDR]], :upper16:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK: ldaexd [[OLD1:r[0-9]+]], [[OLD2:r[0-9]+]], [r[[ADDR]]]
+; CHECK: ldaexd [[OLD1:r[0-9]+|lr]], [[OLD2:r[0-9]+|lr]], [r[[ADDR]]]
   ; r0, r1 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-ARM: mov [[LOCARRY:r[0-9]+|lr]], #0
@@ -782,7 +782,7 @@ define void @test_atomic_load_max_i64(i64 %offset) nounwind {
 ; CHECK: movt r[[ADDR]], :upper16:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK: ldrexd [[OLD1:r[0-9]+]], [[OLD2:r[0-9]+]], [r[[ADDR]]]
+; CHECK: ldrexd [[OLD1:r[0-9]+]], [[OLD2:r[0-9]+|lr]], [r[[ADDR]]]
   ; r0, r1 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-ARM: mov [[LOCARRY:r[0-9]+|lr]], #0
@@ -900,7 +900,7 @@ define void @test_atomic_load_umin_i64(i64 %offset) nounwind {
 ; CHECK: movt r[[ADDR]], :upper16:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK: ldaexd [[OLD1:r[0-9]+]], [[OLD2:r[0-9]+]], [r[[ADDR]]]
+; CHECK: ldaexd [[OLD1:r[0-9]+|lr]], [[OLD2:r[0-9]+|lr]], [r[[ADDR]]]
   ; r0, r1 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-ARM: mov [[LOCARRY:r[0-9]+|lr]], #0
@@ -1018,7 +1018,7 @@ define void @test_atomic_load_umax_i64(i64 %offset) nounwind {
 ; CHECK: movt r[[ADDR]], :upper16:var64
 
 ; CHECK: .LBB{{[0-9]+}}_1:
-; CHECK: ldaexd [[OLD1:r[0-9]+]], [[OLD2:r[0-9]+]], [r[[ADDR]]]
+; CHECK: ldaexd [[OLD1:r[0-9]+|lr]], [[OLD2:r[0-9]+|lr]], [r[[ADDR]]]
   ; r0, r1 below is a reasonable guess but could change: it certainly comes into the
   ; function there.
 ; CHECK-ARM: mov [[LOCARRY:r[0-9]+|lr]], #0
@@ -1146,10 +1146,12 @@ define void @test_atomic_cmpxchg_i64(i64 %wanted, i64 %new) nounwind {
   ; function there.
 ; CHECK-LE-DAG: eor{{(\.w)?}} [[MISMATCH_LO:r[0-9]+|lr]], [[OLD1]], r0
 ; CHECK-LE-DAG: eor{{(\.w)?}} [[MISMATCH_HI:r[0-9]+|lr]], [[OLD2]], r1
-; CHECK-LE: orrs{{(\.w)?}} {{r[0-9]+}}, [[MISMATCH_LO]], [[MISMATCH_HI]]
+; CHECK-ARM-LE: orrs{{(\.w)?}} {{r[0-9]+}}, [[MISMATCH_LO]], [[MISMATCH_HI]]
+; CHECK-THUMB-LE: orrs{{(\.w)?}} {{(r[0-9]+, )?}}[[MISMATCH_HI]], [[MISMATCH_LO]]
 ; CHECK-BE-DAG: eor{{(\.w)?}} [[MISMATCH_HI:r[0-9]+|lr]], [[OLD2]], r1
 ; CHECK-BE-DAG: eor{{(\.w)?}} [[MISMATCH_LO:r[0-9]+|lr]], [[OLD1]], r0
-; CHECK-BE: orrs{{(\.w)?}} {{r[0-9]+}}, [[MISMATCH_HI]], [[MISMATCH_LO]]
+; CHECK-ARM-BE: orrs{{(\.w)?}} {{r[0-9]+}}, [[MISMATCH_HI]], [[MISMATCH_LO]]
+; CHECK-THUMB-BE: orrs{{(\.w)?}} {{(r[0-9]+, )?}}[[MISMATCH_LO]], [[MISMATCH_HI]]
 ; CHECK-NEXT: bne .LBB{{[0-9]+}}_3
 ; CHECK-NEXT: BB#2:
   ; As above, r2, r3 is a reasonable guess.

diff --git a/llvm/test/CodeGen/ARM/ldrd.ll b/llvm/test/CodeGen/ARM/ldrd.ll
index 7ce8468..a8070ea 100644 (file)
--- a/llvm/test/CodeGen/ARM/ldrd.ll
+++ b/llvm/test/CodeGen/ARM/ldrd.ll
@@ -92,6 +92,22 @@ entry:
   ret void
 }
 
+declare void @extfunc(i32, i32, i32, i32)
+
+; CHECK-LABEL: Func2:
+; A8: ldrd
+; A8: blx
+; A8: pop
+define void @Func2(i32* %p) {
+entry:
+  %addr0 = getelementptr i32, i32* %p, i32 0
+  %addr1 = getelementptr i32, i32* %p, i32 1
+  %v0 = load i32, i32* %addr0
+  %v1 = load i32, i32* %addr1
+  ; try to force %v0/%v1 into non-adjacent registers
+  call void @extfunc(i32 %v0, i32 0, i32 0, i32 %v1)
+  ret void
+}
 
 declare void @llvm.lifetime.start(i64, i8* nocapture) nounwind
 declare void @llvm.lifetime.end(i64, i8* nocapture) nounwind

diff --git a/llvm/test/CodeGen/Thumb2/float-ops.ll b/llvm/test/CodeGen/Thumb2/float-ops.ll
index 7ec08f8..4c42908 100644 (file)
--- a/llvm/test/CodeGen/Thumb2/float-ops.ll
+++ b/llvm/test/CodeGen/Thumb2/float-ops.ll
@@ -109,7 +109,7 @@ entry:
 define double @load_d(double* %a) {
 entry:
 ; CHECK-LABEL: load_d:
-; NONE: ldm r0, {r0, r1}
+; NONE: ldrd r0, r1, [r0]
 ; HARD: vldr d0, [r0]
   %0 = load double, double* %a, align 8
   ret double %0
@@ -127,9 +127,7 @@ entry:
 define void @store_d(double* %a, double %b) {
 entry:
 ; CHECK-LABEL: store_d:
-; NONE: mov r1, r3
-; NONE: str r2, [r0]
-; NONE: str r1, [r0, #4]
+; NONE: strd r2, r3, [r0]
 ; HARD: vstr d0, [r0]
   store double %b, double* %a, align 8
   ret void