AMDGPU: Default to using scalar mov to materialize immediate

This is the conservatively correct way because it's easy to
move or replace a scalar immediate. This was incorrect in the case
when the register class wasn't known from the static instruction
definition, but still needed to be an SGPR. The main example of this
is inlineasm has an SGPR constraint.

Also start verifying the register classes of inlineasm operands.

llvm-svn: 285762

3 files changed, 39 insertions, 11 deletions

diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
index 905ee467319..108995a463f 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.cpp
@@ -1943,6 +1943,28 @@ bool SIInstrInfo::verifyInstruction(const MachineInstr &MI,
     return false;
   }
 
+  if (MI.isInlineAsm()) {
+    // Verify register classes for inlineasm constraints.
+    for (unsigned I = InlineAsm::MIOp_FirstOperand, E = MI.getNumOperands();
+         I != E; ++I) {
+      const TargetRegisterClass *RC = MI.getRegClassConstraint(I, this, &RI);
+      if (!RC)
+        continue;
+
+      const MachineOperand &Op = MI.getOperand(I);
+      if (!Op.isReg())
+        continue;
+
+      unsigned Reg = Op.getReg();
+      if (!TargetRegisterInfo::isVirtualRegister(Reg) && !RC->contains(Reg)) {
+        ErrInfo = "inlineasm operand has incorrect register class.";
+        return false;
+      }
+    }
+
+    return true;
+  }
+
   // Make sure the register classes are correct.
   for (int i = 0, e = Desc.getNumOperands(); i != e; ++i) {
     if (MI.getOperand(i).isFPImm()) {
diff --git a/llvm/lib/Target/AMDGPU/SIInstrInfo.td b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
index a56d909624f..f19e99e7cd1 100644
--- a/llvm/lib/Target/AMDGPU/SIInstrInfo.td
+++ b/llvm/lib/Target/AMDGPU/SIInstrInfo.td
@@ -265,19 +265,25 @@ class InlineFPImm <ValueType vt> : PatLeaf <(vt fpimm), [{
   return isInlineImmediate(N);
 }]>;
 
-class SGPRImm <dag frag> : PatLeaf<frag, [{
+class VGPRImm <dag frag> : PatLeaf<frag, [{
   if (Subtarget->getGeneration() < SISubtarget::SOUTHERN_ISLANDS) {
     return false;
   }
   const SIRegisterInfo *SIRI =
       static_cast<const SIRegisterInfo *>(Subtarget->getRegisterInfo());
+  unsigned Limit = 0;
   for (SDNode::use_iterator U = N->use_begin(), E = SDNode::use_end();
-                                                U != E; ++U) {
+         Limit < 10 && U != E; ++U, ++Limit) {
     const TargetRegisterClass *RC = getOperandRegClass(*U, U.getOperandNo());
-    if (RC && SIRI->isSGPRClass(RC))
-      return true;
+
+    // If the register class is unknown, it could be an unknown
+    // register class that needs to be an SGPR, e.g. an inline asm
+    // constraint
+    if (!RC || SIRI->isSGPRClass(RC))
+      return false;
   }
-  return false;
+
+  return Limit < 10;
 }]>;
 
 //===----------------------------------------------------------------------===//
diff --git a/llvm/lib/Target/AMDGPU/SIInstructions.td b/llvm/lib/Target/AMDGPU/SIInstructions.td
index f3ccee288fd..4122eb915f3 100644
--- a/llvm/lib/Target/AMDGPU/SIInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SIInstructions.td
@@ -615,23 +615,23 @@ def : Pat <
 /********** ================== **********/
 
 def : Pat <
-  (SGPRImm<(i32 imm)>:$imm),
-  (S_MOV_B32 imm:$imm)
+  (VGPRImm<(i32 imm)>:$imm),
+  (V_MOV_B32_e32 imm:$imm)
 >;
 
 def : Pat <
-  (SGPRImm<(f32 fpimm)>:$imm),
-  (S_MOV_B32 (f32 (bitcast_fpimm_to_i32 $imm)))
+  (VGPRImm<(f32 fpimm)>:$imm),
+  (V_MOV_B32_e32 (f32 (bitcast_fpimm_to_i32 $imm)))
 >;
 
 def : Pat <
   (i32 imm:$imm),
-  (V_MOV_B32_e32 imm:$imm)
+  (S_MOV_B32 imm:$imm)
 >;
 
 def : Pat <
   (f32 fpimm:$imm),
-  (V_MOV_B32_e32 (f32 (bitcast_fpimm_to_i32 $imm)))
+  (S_MOV_B32 (f32 (bitcast_fpimm_to_i32 $imm)))
 >;
 
 def : Pat <