AMDGPU: Allow a SGPR for the conditional KILL operand.

[bnieuwenhuizen]

Summary

Just use the _e64 variant if needed. This should be possible as per

def : Pat <

(int_amdgcn_kill (i1 (setcc f32:$src, InlineFPImm<f32>:$imm, cond:$cond))),
(SI_KILL_F32_COND_IMM_PSEUDO $src, (bitcast_fpimm_to_i32 $imm), (cond_as_i32imm $cond))

;

I don't think we can get an immediate for the other operand for which we
need the second 32-bit word.

[nhaehnle]

One minor nitpack, apart from that LGTM.

lib/Target/AMDGPU/SIInsertSkips.cpp

262–263

This could just be an assert. I'm also not entirely convinced that it's required. src0 is a VSrc_b32, same as for V_CMPX_, but I suppose it doesn't hurt.

[bnieuwenhuizen]

Move it to an assert, and removed the redundant check for the same condition
in the following if.

[bnieuwenhuizen]

Could you commit this for me? I don't have commit access yet.

[mareko]

Pushed.

lib/Target/AMDGPU/SIInsertSkips.cpp

void SIInsertSkips::kill(MachineInstr &MI) {
   case AMDGPU::SI_KILL_F32_COND_IMM_TERMINATOR: {
     unsigned Opcode = 0;
 
     // The opcodes are inverted because the inline immediate has to be
     // the first operand, e.g. from "x < imm" to "imm > x"
     switch (MI.getOperand(2).getImm()) {
     case ISD::SETOEQ:
     case ISD::SETEQ:
-      Opcode = AMDGPU::V_CMPX_EQ_F32_e32;
+      Opcode = AMDGPU::V_CMPX_EQ_F32_e64;
       break;
     case ISD::SETOGT:
     case ISD::SETGT:
-      Opcode = AMDGPU::V_CMPX_LT_F32_e32;
+      Opcode = AMDGPU::V_CMPX_LT_F32_e64;
       break;
     case ISD::SETOGE:
     case ISD::SETGE:
-      Opcode = AMDGPU::V_CMPX_LE_F32_e32;
+      Opcode = AMDGPU::V_CMPX_LE_F32_e64;
       break;
     case ISD::SETOLT:
     case ISD::SETLT:
-      Opcode = AMDGPU::V_CMPX_GT_F32_e32;
+      Opcode = AMDGPU::V_CMPX_GT_F32_e64;
       break;
     case ISD::SETOLE:
     case ISD::SETLE:
-      Opcode = AMDGPU::V_CMPX_GE_F32_e32;
+      Opcode = AMDGPU::V_CMPX_GE_F32_e64;
       break;
     case ISD::SETONE:
     case ISD::SETNE:
-      Opcode = AMDGPU::V_CMPX_LG_F32_e32;
+      Opcode = AMDGPU::V_CMPX_LG_F32_e64;
       break;
     case ISD::SETO:
-      Opcode = AMDGPU::V_CMPX_O_F32_e32;
+      Opcode = AMDGPU::V_CMPX_O_F32_e64;
       break;
     case ISD::SETUO:
-      Opcode = AMDGPU::V_CMPX_U_F32_e32;
+      Opcode = AMDGPU::V_CMPX_U_F32_e64;
       break;
     case ISD::SETUEQ:
-      Opcode = AMDGPU::V_CMPX_NLG_F32_e32;
+      Opcode = AMDGPU::V_CMPX_NLG_F32_e64;
       break;
     case ISD::SETUGT:
-      Opcode = AMDGPU::V_CMPX_NGE_F32_e32;
+      Opcode = AMDGPU::V_CMPX_NGE_F32_e64;
       break;
     case ISD::SETUGE:
-      Opcode = AMDGPU::V_CMPX_NGT_F32_e32;
+      Opcode = AMDGPU::V_CMPX_NGT_F32_e64;
       break;
     case ISD::SETULT:
-      Opcode = AMDGPU::V_CMPX_NLE_F32_e32;
+      Opcode = AMDGPU::V_CMPX_NLE_F32_e64;
       break;
     case ISD::SETULE:
-      Opcode = AMDGPU::V_CMPX_NLT_F32_e32;
+      Opcode = AMDGPU::V_CMPX_NLT_F32_e64;
       break;
     case ISD::SETUNE:
-      Opcode = AMDGPU::V_CMPX_NEQ_F32_e32;
+      Opcode = AMDGPU::V_CMPX_NEQ_F32_e64;
       break;
     default:
       llvm_unreachable("invalid ISD:SET cond code");
     }
 
-    // TODO: Allow this:
-    if (!MI.getOperand(0).isReg() ||
-        !TRI->isVGPR(MBB.getParent()->getRegInfo(),
-                     MI.getOperand(0).getReg()))
-      llvm_unreachable("SI_KILL operand should be a VGPR");
+    assert(MI.getOperand(0).isReg());
 
+    if (TRI->isVGPR(MBB.getParent()->getRegInfo(),
+                    MI.getOperand(0).getReg())) {
+      Opcode = AMDGPU::getVOPe32(Opcode);
       BuildMI(MBB, &MI, DL, TII->get(Opcode))
           .add(MI.getOperand(1))
           .add(MI.getOperand(0));
+    } else {
+      BuildMI(MBB, &MI, DL, TII->get(Opcode))
+          .addReg(AMDGPU::VCC, RegState::Define)
+          .addImm(0)  // src0 modifiers
+          .add(MI.getOperand(1))
+          .addImm(0)  // src1 modifiers
+          .add(MI.getOperand(0))
+          .addImm(0);  // omod
+    }
     break;
   }
   case AMDGPU::SI_KILL_I1_TERMINATOR: {
     const MachineOperand &Op = MI.getOperand(0);
     int64_t KillVal = MI.getOperand(1).getImm();
     assert(KillVal == 0 || KillVal == -1);
 
     // Kill all threads if Op0 is an immediate and equal to the Kill value.

test/CodeGen/AMDGPU/llvm.amdgcn.kill.ll

 ; SI: s_and_b64 exec, exec, s[0:1]
 define amdgpu_ps void @wqm(float %a) {
   %c1 = fcmp une float %a, 0.0
   %c2 = call i1 @llvm.amdgcn.wqm.vote(i1 %c1)
   call void @llvm.amdgcn.kill(i1 %c2)
   ret void
 }
 
+; This checks that we use the 64-bit encoding when the operand is a SGPR.
+; SI-LABEL: {{^}}test_sgpr:
+; SI: v_cmpx_ge_f32_e64
+define amdgpu_ps void @test_sgpr(float inreg %a) #0 {
+  %c = fcmp ole float %a, 1.000000e+00
+  call void @llvm.amdgcn.kill(i1 %c) #1
+  ret void
+}
+
+; SI-LABEL: {{^}}test_non_inline_imm_sgpr:
+; SI-NOT: v_cmpx_ge_f32_e64
+define amdgpu_ps void @test_non_inline_imm_sgpr(float inreg %a) #0 {
+  %c = fcmp ole float %a, 1.500000e+00
+  call void @llvm.amdgcn.kill(i1 %c) #1
+  ret void
+}
+
 declare void @llvm.amdgcn.kill(i1) #0
 declare void @llvm.amdgcn.exp.f32(i32, i32, float, float, float, float, i1, i1) #0
 declare i1 @llvm.amdgcn.wqm.vote(i1)
 
 attributes #0 = { nounwind }