Add missing i8 max/min/umax/umin support

- Fix PR5145 and turn on test 8-bit atomic ops

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@164358 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86ISelLowering.cpp b/lib/Target/X86/X86ISelLowering.cpp
index 0e6e4a3294b493e134b3c30615271174e46715cb..b2b94794f205ddd05db29da02265d2d08226735c 100644 (file)
--- a/lib/Target/X86/X86ISelLowering.cpp
+++ b/lib/Target/X86/X86ISelLowering.cpp
@@ -11966,15 +11966,19 @@ static unsigned getNonAtomicOpcodeWithExtraOpc(unsigned Opc,
   case X86::ATOMNAND16: ExtraOpc = X86::NOT16r;  return X86::AND16rr;
   case X86::ATOMNAND32: ExtraOpc = X86::NOT32r;  return X86::AND32rr;
   case X86::ATOMNAND64: ExtraOpc = X86::NOT64r;  return X86::AND64rr;
+  case X86::ATOMMAX8:   ExtraOpc = X86::CMP8rr;  return X86::CMOVL32rr;
   case X86::ATOMMAX16:  ExtraOpc = X86::CMP16rr; return X86::CMOVL16rr;
   case X86::ATOMMAX32:  ExtraOpc = X86::CMP32rr; return X86::CMOVL32rr;
   case X86::ATOMMAX64:  ExtraOpc = X86::CMP64rr; return X86::CMOVL64rr;
+  case X86::ATOMMIN8:   ExtraOpc = X86::CMP8rr;  return X86::CMOVG32rr;
   case X86::ATOMMIN16:  ExtraOpc = X86::CMP16rr; return X86::CMOVG16rr;
   case X86::ATOMMIN32:  ExtraOpc = X86::CMP32rr; return X86::CMOVG32rr;
   case X86::ATOMMIN64:  ExtraOpc = X86::CMP64rr; return X86::CMOVG64rr;
+  case X86::ATOMUMAX8:  ExtraOpc = X86::CMP8rr;  return X86::CMOVB32rr;
   case X86::ATOMUMAX16: ExtraOpc = X86::CMP16rr; return X86::CMOVB16rr;
   case X86::ATOMUMAX32: ExtraOpc = X86::CMP32rr; return X86::CMOVB32rr;
   case X86::ATOMUMAX64: ExtraOpc = X86::CMP64rr; return X86::CMOVB64rr;
+  case X86::ATOMUMIN8:  ExtraOpc = X86::CMP8rr;  return X86::CMOVA32rr;
   case X86::ATOMUMIN16: ExtraOpc = X86::CMP16rr; return X86::CMOVA16rr;
   case X86::ATOMUMIN32: ExtraOpc = X86::CMP32rr; return X86::CMOVA32rr;
   case X86::ATOMUMIN64: ExtraOpc = X86::CMP64rr; return X86::CMOVA64rr;
@@ -12013,6 +12017,7 @@ static unsigned getNonAtomic6432OpcodeWithExtraOpc(unsigned Opc,
 // Get pseudo CMOV opcode from the specified data type.
 static unsigned getPseudoCMOVOpc(EVT VT) {
   switch (VT.getSimpleVT().SimpleTy) {
+  case MVT::i8:  return X86::CMOV_GR8;
   case MVT::i16: return X86::CMOV_GR16;
   case MVT::i32: return X86::CMOV_GR32;
   default:
@@ -12154,19 +12159,18 @@ X86TargetLowering::EmitAtomicLoadArith(MachineInstr *MI,
     break;
   }
   case X86::ATOMMAX8:
-  case X86::ATOMMIN8:
-  case X86::ATOMUMAX8:
-  case X86::ATOMUMIN8:
-    llvm_unreachable("Not supported yet!");
   case X86::ATOMMAX16:
   case X86::ATOMMAX32:
   case X86::ATOMMAX64:
+  case X86::ATOMMIN8:
   case X86::ATOMMIN16:
   case X86::ATOMMIN32:
   case X86::ATOMMIN64:
+  case X86::ATOMUMAX8:
   case X86::ATOMUMAX16:
   case X86::ATOMUMAX32:
   case X86::ATOMUMAX64:
+  case X86::ATOMUMIN8:
   case X86::ATOMUMIN16:
   case X86::ATOMUMIN32:
   case X86::ATOMUMIN64: {
@@ -12178,13 +12182,40 @@ X86TargetLowering::EmitAtomicLoadArith(MachineInstr *MI,
       .addReg(AccReg);
 
     if (Subtarget->hasCMov()) {
-      // Native support
-      BuildMI(mainMBB, DL, TII->get(CMOVOpc), t1)
-        .addReg(SrcReg)
-        .addReg(AccReg);
+      if (VT != MVT::i8) {
+        // Native support
+        BuildMI(mainMBB, DL, TII->get(CMOVOpc), t1)
+          .addReg(SrcReg)
+          .addReg(AccReg);
+      } else {
+        // Promote i8 to i32 to use CMOV32
+        const TargetRegisterClass *RC32 = getRegClassFor(MVT::i32);
+        unsigned SrcReg32 = MRI.createVirtualRegister(RC32);
+        unsigned AccReg32 = MRI.createVirtualRegister(RC32);
+        unsigned t2 = MRI.createVirtualRegister(RC32);
+
+        unsigned Undef = MRI.createVirtualRegister(RC32);
+        BuildMI(mainMBB, DL, TII->get(TargetOpcode::IMPLICIT_DEF), Undef);
+
+        BuildMI(mainMBB, DL, TII->get(TargetOpcode::INSERT_SUBREG), SrcReg32)
+          .addReg(Undef)
+          .addReg(SrcReg)
+          .addImm(X86::sub_8bit);
+        BuildMI(mainMBB, DL, TII->get(TargetOpcode::INSERT_SUBREG), AccReg32)
+          .addReg(Undef)
+          .addReg(AccReg)
+          .addImm(X86::sub_8bit);
+
+        BuildMI(mainMBB, DL, TII->get(CMOVOpc), t2)
+          .addReg(SrcReg32)
+          .addReg(AccReg32);
+
+        BuildMI(mainMBB, DL, TII->get(TargetOpcode::COPY), t1)
+          .addReg(t2, 0, X86::sub_8bit);
+      }
     } else {
       // Use pseudo select and lower them.
-      assert((VT == MVT::i16 || VT == MVT::i32) &&
+      assert((VT == MVT::i8 || VT == MVT::i16 || VT == MVT::i32) &&
              "Invalid atomic-load-op transformation!");
       unsigned SelOpc = getPseudoCMOVOpc(VT);
       X86::CondCode CC = X86::getCondFromCMovOpc(CMOVOpc);
@@ -13314,18 +13345,22 @@ X86TargetLowering::EmitInstrWithCustomInserter(MachineInstr *MI,
   case X86::ATOMNAND32:
   case X86::ATOMNAND64:
     // Fall through
+  case X86::ATOMMAX8:
   case X86::ATOMMAX16:
   case X86::ATOMMAX32:
   case X86::ATOMMAX64:
     // Fall through
+  case X86::ATOMMIN8:
   case X86::ATOMMIN16:
   case X86::ATOMMIN32:
   case X86::ATOMMIN64:
     // Fall through
+  case X86::ATOMUMAX8:
   case X86::ATOMUMAX16:
   case X86::ATOMUMAX32:
   case X86::ATOMUMAX64:
     // Fall through
+  case X86::ATOMUMIN8:
   case X86::ATOMUMIN16:
   case X86::ATOMUMIN32:
   case X86::ATOMUMIN64:

diff --git a/test/CodeGen/X86/atomic8.ll b/test/CodeGen/X86/atomic8.ll
index 035a28dbffa3df94692f33a43e78577854a457ad..412428406dcffc69f106caa7cf51e283a99a84bc 100644 (file)
--- a/test/CodeGen/X86/atomic8.ll
+++ b/test/CodeGen/X86/atomic8.ll
@@ -1,6 +1,5 @@
 ; RUN: llc < %s -O0 -march=x86-64 -mcpu=corei7 | FileCheck %s --check-prefix X64
 ; RUN: llc < %s -O0 -march=x86 -mcpu=corei7 | FileCheck %s --check-prefix X32
-; XFAIL: *
 
 @sc8 = external global i8
 

diff --git a/test/CodeGen/X86/pr5145.ll b/test/CodeGen/X86/pr5145.ll
new file mode 100644 (file)
index 0000000..06ecd7a
--- /dev/null
+++ b/test/CodeGen/X86/pr5145.ll
@@ -0,0 +1,35 @@
+; RUN: llc -march=x86-64 < %s | FileCheck %s
+@sc8 = external global i8
+
+define void @atomic_maxmin_i8() {
+; CHECK: atomic_maxmin_i8
+  %1 = atomicrmw max  i8* @sc8, i8 5 acquire
+; CHECK: [[LABEL:.LBB[0-9]+_[0-9]+]]
+; CHECK: cmpb
+; CHECK: cmovl
+; CHECK: lock
+; CHECK-NEXT: cmpxchgb
+; CHECK: jne [[LABEL]]
+  %2 = atomicrmw min  i8* @sc8, i8 6 acquire
+; CHECK: [[LABEL:.LBB[0-9]+_[0-9]+]]
+; CHECK: cmpb
+; CHECK: cmovg
+; CHECK: lock
+; CHECK-NEXT: cmpxchgb
+; CHECK: jne [[LABEL]]
+  %3 = atomicrmw umax i8* @sc8, i8 7 acquire
+; CHECK: [[LABEL:.LBB[0-9]+_[0-9]+]]
+; CHECK: cmpb
+; CHECK: cmovb
+; CHECK: lock
+; CHECK-NEXT: cmpxchgb
+; CHECK: jne [[LABEL]]
+  %4 = atomicrmw umin i8* @sc8, i8 8 acquire
+; CHECK: [[LABEL:.LBB[0-9]+_[0-9]+]]
+; CHECK: cmpb
+; CHECK: cmova
+; CHECK: lock
+; CHECK-NEXT: cmpxchgb
+; CHECK: jne [[LABEL]]
+  ret void
+}