[AMDGPU] Save/Restore SCC bit across waterfall loop. (#68363)

Waterfall loop is overwriting SCC bit of status register. Make sure SCC
bit is saved and restored across.
We need to save/restore only in cases where SCC is live across waterfall
loop.

Co-authored-by: Sirish Pande <[email protected]>

Author: srpande

llvm/lib/Target/AMDGPU/SIInstrInfo.cpp

@@ -6079,6 +6079,17 @@ loadMBUFScalarOperandsFromVGPR(const SIInstrInfo &TII, MachineInstr &MI,
   unsigned MovExecOpc = ST.isWave32() ? AMDGPU::S_MOV_B32 : AMDGPU::S_MOV_B64;
   const auto *BoolXExecRC = TRI->getRegClass(AMDGPU::SReg_1_XEXECRegClassID);
 
+  // Save SCC. Waterfall Loop may overwrite SCC.
+  Register SaveSCCReg;
+  bool SCCNotDead = (MBB.computeRegisterLiveness(TRI, AMDGPU::SCC, MI, 30) !=
+                     MachineBasicBlock::LQR_Dead);
+  if (SCCNotDead) {
+    SaveSCCReg = MRI.createVirtualRegister(&AMDGPU::SReg_32RegClass);
+    BuildMI(MBB, Begin, DL, TII.get(AMDGPU::S_CSELECT_B32), SaveSCCReg)
+        .addImm(1)
+        .addImm(0);
+  }
+
   Register SaveExec = MRI.createVirtualRegister(BoolXExecRC);
 
   // Save the EXEC mask
@@ -6134,8 +6145,15 @@ loadMBUFScalarOperandsFromVGPR(const SIInstrInfo &TII, MachineInstr &MI,
 
   emitLoadScalarOpsFromVGPRLoop(TII, MRI, MBB, *LoopBB, *BodyBB, DL, ScalarOps);
 
-  // Restore the EXEC mask
   MachineBasicBlock::iterator First = RemainderBB->begin();
+  // Restore SCC
+  if (SCCNotDead) {
+    BuildMI(*RemainderBB, First, DL, TII.get(AMDGPU::S_CMP_LG_U32))
+        .addReg(SaveSCCReg, RegState::Kill)
+        .addImm(0);
+  }
+
+  // Restore the EXEC mask
   BuildMI(*RemainderBB, First, DL, TII.get(MovExecOpc), Exec).addReg(SaveExec);
   return BodyBB;
 }

llvm/test/CodeGen/AMDGPU/waterfall_kills_scc.ll

@@ -0,0 +1,83 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN:  llc -march=amdgcn -mcpu=gfx906 -verify-machineinstrs < %s | FileCheck -check-prefix=GFX906 %s
+declare float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32>, i32, i32, i32 immarg) #0
+declare void @llvm.amdgcn.raw.buffer.store.f32(float, <4 x i32>, i32, i32, i32 immarg) #1
+
+; Check that the compiler doesn't crash with a "undefined physical register" error;
+; bb.0 sets SCC bit in s_cmp_eq_u32 s0, 1
+; bb.1 overrides it
+; bb.2 uses the value from bb.0
+; Preserve SCC across bb.1 with s_cselect_b32 s5, 1, 0 -> s_cmp_lg_u32 s5, 0
+; Otherwise, we will see the following error.
+;*** Bad machine code: Using an undefined physical register ***
+;- function:    foo
+;- basic block: %bb.3  (0x53198c0)
+;- instruction: %33.sub1:sgpr_128 = S_CSELECT_B32 1072693248, 0, implicit $scc
+;- operand 3:   implicit $scc
+
+define amdgpu_kernel void  @foo(i1 %cmp1) {
+; GFX906-LABEL: foo:
+; GFX906:       ; %bb.0: ; %entry
+; GFX906-NEXT:    s_mov_b32 s8, SCRATCH_RSRC_DWORD0
+; GFX906-NEXT:    s_mov_b32 s9, SCRATCH_RSRC_DWORD1
+; GFX906-NEXT:    s_mov_b32 s10, -1
+; GFX906-NEXT:    s_mov_b32 s11, 0xe00000
+; GFX906-NEXT:    s_add_u32 s8, s8, s3
+; GFX906-NEXT:    s_addc_u32 s9, s9, 0
+; GFX906-NEXT:    buffer_load_dword v3, off, s[8:11], 0
+; GFX906-NEXT:    buffer_load_dword v4, off, s[8:11], 0 offset:4
+; GFX906-NEXT:    buffer_load_dword v5, off, s[8:11], 0 offset:8
+; GFX906-NEXT:    buffer_load_dword v6, off, s[8:11], 0 offset:12
+; GFX906-NEXT:    s_load_dword s4, s[0:1], 0x24
+; GFX906-NEXT:    s_load_dwordx2 s[2:3], s[0:1], 0x1c
+; GFX906-NEXT:    s_waitcnt lgkmcnt(0)
+; GFX906-NEXT:    s_bitcmp1_b32 s4, 0
+; GFX906-NEXT:    s_mul_i32 s0, s2, s3
+; GFX906-NEXT:    v_mul_u32_u24_e32 v1, s3, v1
+; GFX906-NEXT:    v_mad_u32_u24 v0, s0, v0, v1
+; GFX906-NEXT:    v_add_lshl_u32 v2, v0, v2, 4
+; GFX906-NEXT:    v_mov_b32_e32 v0, 0
+; GFX906-NEXT:    s_mov_b32 s4, 0
+; GFX906-NEXT:    v_mov_b32_e32 v1, v0
+; GFX906-NEXT:    s_cselect_b32 s5, 1, 0
+; GFX906-NEXT:    s_mov_b64 s[2:3], exec
+; GFX906-NEXT:    ds_write_b64 v2, v[0:1]
+; GFX906-NEXT:  .LBB0_1: ; =>This Inner Loop Header: Depth=1
+; GFX906-NEXT:    s_waitcnt vmcnt(3)
+; GFX906-NEXT:    v_readfirstlane_b32 s0, v3
+; GFX906-NEXT:    s_waitcnt vmcnt(2)
+; GFX906-NEXT:    v_readfirstlane_b32 s1, v4
+; GFX906-NEXT:    v_cmp_eq_u64_e32 vcc, s[0:1], v[3:4]
+; GFX906-NEXT:    s_waitcnt vmcnt(1)
+; GFX906-NEXT:    v_readfirstlane_b32 s0, v5
+; GFX906-NEXT:    s_waitcnt vmcnt(0)
+; GFX906-NEXT:    v_readfirstlane_b32 s1, v6
+; GFX906-NEXT:    v_cmp_eq_u64_e64 s[0:1], s[0:1], v[5:6]
+; GFX906-NEXT:    s_and_b64 s[0:1], vcc, s[0:1]
+; GFX906-NEXT:    s_and_saveexec_b64 s[0:1], s[0:1]
+; GFX906-NEXT:    ; implicit-def: $vgpr3_vgpr4_vgpr5_vgpr6
+; GFX906-NEXT:    s_xor_b64 exec, exec, s[0:1]
+; GFX906-NEXT:    s_cbranch_execnz .LBB0_1
+; GFX906-NEXT:  ; %bb.2:
+; GFX906-NEXT:    s_cmp_lg_u32 s5, 0
+; GFX906-NEXT:    s_mov_b64 exec, s[2:3]
+; GFX906-NEXT:    s_cselect_b32 s5, 0x3ff00000, 0
+; GFX906-NEXT:    v_cvt_f32_f64_e32 v0, s[4:5]
+; GFX906-NEXT:    s_mov_b32 s5, s4
+; GFX906-NEXT:    s_mov_b32 s6, s4
+; GFX906-NEXT:    s_mov_b32 s7, s4
+; GFX906-NEXT:    buffer_store_dword v0, off, s[4:7], 0
+; GFX906-NEXT:    s_endpgm
+entry:
+  %wbr = alloca <4 x i32>, align 16, addrspace(5)
+  store ptr null, ptr addrspace(5) %wbr, align 16
+  %wbr_1 = load <4 x i32>, ptr addrspace(5) null, align 16
+  %call1 = tail call float @llvm.amdgcn.raw.buffer.load.f32(<4 x i32> %wbr_1, i32 0, i32 0, i32 0)
+  %0 = fpext float %call1 to double
+  %sel1 = select i1 %cmp1, double 1.000000e+00, double 0.000000e+00
+  %sel2 = select i1 %cmp1, double %0, double 0.000000e+00
+  %mul = fmul double %sel2, 0.000000e+00
+  %fptruncate = fptrunc double %sel1 to float
+  tail call void @llvm.amdgcn.raw.buffer.store.f32(float %fptruncate, <4 x i32> zeroinitializer, i32 0, i32 0, i32 0)
+  ret void
+}