[PATCH] D99670: [AMDGPU] Don't rely on SIAddIMGInit for GlobalISel

[Jay Foad via Phabricator via llvm-commits]

foad updated this revision to Diff 334741.
foad added a comment.

Rebase.


Repository:
  rG LLVM Github Monorepo

CHANGES SINCE LAST ACTION
  https://reviews.llvm.org/D99670/new/

https://reviews.llvm.org/D99670

Files:
  llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
  llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp


Index: llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
+++ llvm/lib/Target/AMDGPU/AMDGPUTargetMachine.cpp
@@ -1133,10 +1133,6 @@
 
 bool GCNPassConfig::addGlobalInstructionSelect() {
   addPass(new InstructionSelect(getOptLevel()));
-  // TODO: Fix instruction selection to do the right thing for image
-  // instructions with tfe or lwe in the first place, instead of running a
-  // separate pass to fix them up?
-  addPass(createSIAddIMGInitPass());
   return false;
 }
 
Index: llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -1689,6 +1689,38 @@
   if (BaseOpcode->HasD16)
     MIB.addImm(IsD16 ? -1 : 0);
 
+  if (IsTexFail) {
+    // An image load instruction with TFE/LWE only conditionally writes to its
+    // result registers. Initialize them to zero so that we always get well
+    // defined result values.
+    assert(VDataOut && !VDataIn);
+    Register Tied = MRI->cloneVirtualRegister(VDataOut);
+    Register Zero = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+    BuildMI(*MBB, *MIB, DL, TII.get(AMDGPU::V_MOV_B32_e32), Zero)
+      .addImm(0);
+    auto Parts = TRI.getRegSplitParts(MRI->getRegClass(Tied), 4);
+    if (STI.usePRTStrictNull()) {
+      // With enable-prt-strict-null enabled, initialize all result registers to
+      // zero.
+      auto RegSeq =
+          BuildMI(*MBB, *MIB, DL, TII.get(AMDGPU::REG_SEQUENCE), Tied);
+      for (auto Sub : Parts)
+        RegSeq.addReg(Zero).addImm(Sub);
+    } else {
+      // With enable-prt-strict-null disabled, only initialize the extra TFE/LWE
+      // result register.
+      Register Undef = MRI->createVirtualRegister(&AMDGPU::VGPR_32RegClass);
+      BuildMI(*MBB, *MIB, DL, TII.get(AMDGPU::IMPLICIT_DEF), Undef);
+      auto RegSeq =
+          BuildMI(*MBB, *MIB, DL, TII.get(AMDGPU::REG_SEQUENCE), Tied);
+      for (auto Sub : Parts.drop_back(1))
+        RegSeq.addReg(Undef).addImm(Sub);
+      RegSeq.addReg(Zero).addImm(Parts.back());
+    }
+    MIB.addReg(Tied, RegState::Implicit);
+    MIB->tieOperands(0, MIB->getNumOperands() - 1);
+  }
+
   MI.eraseFromParent();
   return constrainSelectedInstRegOperands(*MIB, TII, TRI, RBI);
 }


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D99670.334741.patch
Type: text/x-patch
Size: 2451 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210401/a2a97e03/attachment-0001.bin>