[PATCH] D105709: [AMDGPU][GlobalISel] Insert an and with exec before s_cbranch_vccnz if necessary

[Mirko Brkusanin via Phabricator via llvm-commits]

mbrkusanin created this revision.
mbrkusanin added reviewers: foad, arsenm.
mbrkusanin added a project: LLVM.
Herald added subscribers: kerbowa, hiraditya, t-tye, tpr, dstuttard, rovka, yaxunl, nhaehnle, jvesely, kzhuravl.
mbrkusanin requested review of this revision.
Herald added a subscriber: wdng.

While v_cmp will AND inactive lanes with 0, that is not the case for logical
operations.

This fixes a Vulkan CTS test that would hang otherwise.


Repository:
  rG LLVM Github Monorepo

https://reviews.llvm.org/D105709

Files:
  llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
  llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-brcond.mir


Index: llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-brcond.mir
===================================================================
--- llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-brcond.mir
+++ llvm/test/CodeGen/AMDGPU/GlobalISel/inst-select-brcond.mir
@@ -174,3 +174,40 @@
   bb.1:
 
 ...
+
+---
+
+name:            brcond_vcc_not_cmp
+legalized:       true
+regBankSelected: true
+
+body: |
+  ; GCN-LABEL: name: brcond_vcc_not_cmp
+  ; GCN: bb.0:
+  ; GCN:   successors: %bb.1(0x80000000)
+  ; GCN:   [[COPY:%[0-9]+]]:vgpr_32 = COPY $vgpr0
+  ; GCN:   [[COPY1:%[0-9]+]]:vgpr_32 = COPY $vgpr1
+  ; GCN:   [[COPY2:%[0-9]+]]:vgpr_32 = COPY $vgpr2
+  ; GCN:   [[COPY3:%[0-9]+]]:vgpr_32 = COPY $vgpr3
+  ; GCN:   [[V_CMP_EQ_U32_e64_:%[0-9]+]]:sreg_64 = V_CMP_EQ_U32_e64 [[COPY]], [[COPY1]], implicit $exec
+  ; GCN:   [[V_CMP_EQ_U32_e64_1:%[0-9]+]]:sreg_64 = V_CMP_EQ_U32_e64 [[COPY2]], [[COPY3]], implicit $exec
+  ; GCN:   [[S_AND_B64_:%[0-9]+]]:sreg_64 = S_AND_B64 [[V_CMP_EQ_U32_e64_]], [[V_CMP_EQ_U32_e64_1]], implicit-def dead $scc
+  ; GCN:   [[S_AND_B64_1:%[0-9]+]]:sreg_64 = S_AND_B64 [[S_AND_B64_]], $exec, implicit-def $scc
+  ; GCN:   $vcc = COPY [[S_AND_B64_1]]
+  ; GCN:   S_CBRANCH_VCCNZ %bb.1, implicit $vcc
+  ; GCN: bb.1:
+  bb.0:
+    liveins: $vgpr0, $vgpr1, $vgpr2, $vgpr3
+
+    %0:vgpr(s32) = COPY $vgpr0
+    %1:vgpr(s32) = COPY $vgpr1
+    %2:vgpr(s32) = COPY $vgpr2
+    %3:vgpr(s32) = COPY $vgpr3
+    %4:vcc(s1) = G_ICMP intpred(eq), %0, %1
+    %5:vcc(s1) = G_ICMP intpred(eq), %2, %3
+    %6:vcc(s1) = G_AND %4, %5
+    G_BRCOND %6(s1), %bb.1
+
+  bb.1:
+
+...
Index: llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
===================================================================
--- llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -2488,11 +2488,25 @@
     BrOpcode = AMDGPU::S_CBRANCH_SCC1;
     ConstrainRC = &AMDGPU::SReg_32RegClass;
   } else {
-    // FIXME: Do we have to insert an and with exec here, like in SelectionDAG?
-    // We sort of know that a VCC producer based on the register bank, that ands
-    // inactive lanes with 0. What if there was a logical operation with vcc
-    // producers in different blocks/with different exec masks?
     // FIXME: Should scc->vcc copies and with exec?
+
+    // If there was an instruction other than V_CMP then we need to insert an
+    // and with exec.
+    const unsigned CondDefOpc = MRI->getUniqueVRegDef(CondReg)->getOpcode();
+    if (CondDefOpc != AMDGPU::G_ICMP && CondDefOpc != AMDGPU::G_FCMP) {
+      const bool Is64 = STI.isWave64();
+      const TargetRegisterClass *RC =
+          Is64 ? &AMDGPU::SReg_64RegClass : &AMDGPU::SReg_32RegClass;
+      const unsigned Opcode = Is64 ? AMDGPU::S_AND_B64 : AMDGPU::S_AND_B32;
+      const Register Exec = Is64 ? AMDGPU::EXEC : AMDGPU::EXEC_LO;
+
+      Register TmpReg = MRI->createVirtualRegister(RC);
+      BuildMI(*BB, &I, DL, TII.get(Opcode), TmpReg)
+          .addReg(CondReg)
+          .addReg(Exec);
+      CondReg = TmpReg;
+    }
+
     CondPhysReg = TRI.getVCC();
     BrOpcode = AMDGPU::S_CBRANCH_VCCNZ;
     ConstrainRC = TRI.getBoolRC();


-------------- next part --------------
A non-text attachment was scrubbed...
Name: D105709.357541.patch
Type: text/x-patch
Size: 3179 bytes
Desc: not available
URL: <http://lists.llvm.org/pipermail/llvm-commits/attachments/20210709/fca6acf5/attachment.bin>