[AMDGPU] Disable rematerialization of cndmask

[rampitec]

Summary

VOP2e promoted to VOP3 is marked as rematerializable because it does
not use implicit vcc, but that shall exclude cndmask. A result of
the cndmask instruction uses exec as a mask, so I believe it may not
move across basic blocks.

[foad]

A result of the cndmask instruction uses exec as a mask

As noted on D117814, I don't understand this.

[rampitec]

In D117821#3260869, @foad wrote:

A result of the cndmask instruction uses exec as a mask

As noted on D117814, I don't understand this.

Scalar result is different when you move it across basic blocks.

llvm/lib/Target/AMDGPU/VOP2Instructions.td

def VOP_WRITELANE : VOPProfile<[i32, i32, i32, i32]> {
   let HasExtSDWA = 0;
   let HasExtSDWA9 = 0;
 }
 
 //===----------------------------------------------------------------------===//
 // VOP2 Instructions
 //===----------------------------------------------------------------------===//
 
+let isReMaterializable = 0 in
 defm V_CNDMASK_B32 : VOP2eInst <"v_cndmask_b32", VOP2e_I32_I32_I32_I1>;
 let SubtargetPredicate = HasMadMacF32Insts, isReMaterializable = 1 in
 def V_MADMK_F32 : VOP2_Pseudo <"v_madmk_f32", VOP_MADMK_F32, []>;
 
 let isCommutable = 1 in {
 let isReMaterializable = 1 in {
 defm V_ADD_F32 : VOP2Inst <"v_add_f32", VOP_F32_F32_F32, any_fadd>;
 defm V_SUB_F32 : VOP2Inst <"v_sub_f32", VOP_F32_F32_F32, any_fsub>;

llvm/test/CodeGen/AMDGPU/remat-vop.mir

bb.0:
     %2:vgpr_32 = V_CNDMASK_B32_dpp undef %2:vgpr_32, 0, undef %0:vgpr_32, 0, undef %0:vgpr_32, 1, 15, 15, 10, implicit $exec, implicit undef $vcc
     %3:vgpr_32 = V_CNDMASK_B32_dpp undef %3:vgpr_32, 0, undef %0:vgpr_32, 0, undef %0:vgpr_32, 1, 15, 15, 10, implicit $exec, implicit undef $vcc
     S_NOP 0, implicit %1
     S_NOP 0, implicit %2
     S_NOP 0, implicit %3
     S_ENDPGM 0
 ...
 ---
-name:            test_remat_v_cndmask_b32_e64
+name:            test_no_remat_v_cndmask_b32_e64
 tracksRegLiveness: true
+machineFunctionInfo:
+  stackPtrOffsetReg:  $sgpr32
 body:             |
   bb.0:
-    ; GCN-LABEL: name: test_remat_v_cndmask_b32_e64
+    ; GCN-LABEL: name: test_no_remat_v_cndmask_b32_e64
     ; GCN: renamable $vgpr0 = V_CNDMASK_B32_e64 0, 1, 0, 2, undef $sgpr0_sgpr1, implicit $exec
+    ; GCN-NEXT: SI_SPILL_V32_SAVE killed $vgpr0, %stack.1, $sgpr32, 0, implicit $exec :: (store (s32) into %stack.1, addrspace 5)
     ; GCN-NEXT: renamable $vgpr1 = V_CNDMASK_B32_e64 0, 1, 0, 2, undef $sgpr0_sgpr1, implicit $exec
+    ; GCN-NEXT: renamable $vgpr0 = V_CNDMASK_B32_e64 0, 1, 0, 2, undef $sgpr0_sgpr1, implicit $exec
+    ; GCN-NEXT: SI_SPILL_V32_SAVE killed $vgpr0, %stack.0, $sgpr32, 0, implicit $exec :: (store (s32) into %stack.0, addrspace 5)
+    ; GCN-NEXT: renamable $vgpr0 = SI_SPILL_V32_RESTORE %stack.1, $sgpr32, 0, implicit $exec :: (load (s32) from %stack.1, addrspace 5)
     ; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr0
     ; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr1
-    ; GCN-NEXT: renamable $vgpr0 = V_CNDMASK_B32_e64 0, 1, 0, 2, undef $sgpr0_sgpr1, implicit $exec
+    ; GCN-NEXT: renamable $vgpr0 = SI_SPILL_V32_RESTORE %stack.0, $sgpr32, 0, implicit $exec :: (load (s32) from %stack.0, addrspace 5)
     ; GCN-NEXT: S_NOP 0, implicit killed renamable $vgpr0
     ; GCN-NEXT: S_ENDPGM 0
     %1:vgpr_32 = V_CNDMASK_B32_e64 0, 1, 0, 2, undef %0:sreg_64_xexec, implicit $exec
     %2:vgpr_32 = V_CNDMASK_B32_e64 0, 1, 0, 2, undef %0:sreg_64_xexec, implicit $exec
     %3:vgpr_32 = V_CNDMASK_B32_e64 0, 1, 0, 2, undef %0:sreg_64_xexec, implicit $exec
     S_NOP 0, implicit %1
     S_NOP 0, implicit %2
     S_NOP 0, implicit %3