[AMDGPU] DPP combiner: recognize identities for more opcodes

[foad]

Summary

This allows the DPP combiner to kick in more often. For example the
exclusive scan generated by the atomic optimizer for a divergent atomic
add used to look like this:

v_mov_b32_e32 v3, v1
v_mov_b32_e32 v5, v1
v_mov_b32_e32 v6, v1
v_mov_b32_dpp v3, v2  wave_shr:1 row_mask:0xf bank_mask:0xf
s_nop 1
v_add_u32_dpp v4, v3, v3  row_shr:1 row_mask:0xf bank_mask:0xf bound_ctrl:0
v_mov_b32_dpp v5, v3  row_shr:2 row_mask:0xf bank_mask:0xf
v_mov_b32_dpp v6, v3  row_shr:3 row_mask:0xf bank_mask:0xf
v_add3_u32 v3, v4, v5, v6
v_mov_b32_e32 v4, v1
s_nop 1
v_mov_b32_dpp v4, v3  row_shr:4 row_mask:0xf bank_mask:0xe
v_add_u32_e32 v3, v3, v4
v_mov_b32_e32 v4, v1
s_nop 1
v_mov_b32_dpp v4, v3  row_shr:8 row_mask:0xf bank_mask:0xc
v_add_u32_e32 v3, v3, v4
v_mov_b32_e32 v4, v1
s_nop 1
v_mov_b32_dpp v4, v3  row_bcast:15 row_mask:0xa bank_mask:0xf
v_add_u32_e32 v3, v3, v4
s_nop 1
v_mov_b32_dpp v1, v3  row_bcast:31 row_mask:0xc bank_mask:0xf
v_add_u32_e32 v1, v3, v1
v_add_u32_e32 v1, v2, v1
v_readlane_b32 s0, v1, 63

But now most of the dpp movs are combined into adds:

v_mov_b32_e32 v3, v1
v_mov_b32_e32 v5, v1
s_nop 0
v_mov_b32_dpp v3, v2  wave_shr:1 row_mask:0xf bank_mask:0xf
s_nop 1
v_add_u32_dpp v4, v3, v3  row_shr:1 row_mask:0xf bank_mask:0xf bound_ctrl:0
v_mov_b32_dpp v5, v3  row_shr:2 row_mask:0xf bank_mask:0xf
v_mov_b32_dpp v1, v3  row_shr:3 row_mask:0xf bank_mask:0xf
v_add3_u32 v1, v4, v5, v1
s_nop 1
v_add_u32_dpp v1, v1, v1  row_shr:4 row_mask:0xf bank_mask:0xe
s_nop 1
v_add_u32_dpp v1, v1, v1  row_shr:8 row_mask:0xf bank_mask:0xc
s_nop 1
v_add_u32_dpp v1, v1, v1  row_bcast:15 row_mask:0xa bank_mask:0xf
s_nop 1
v_add_u32_dpp v1, v1, v1  row_bcast:31 row_mask:0xc bank_mask:0xf
v_add_u32_e32 v1, v2, v1
v_readlane_b32 s0, v1, 63

Also fix some typos in comments and debug output.

[vpykhtin]

I'm not sure if e64 instructions have modifiers that cannot be encoded into DPP version, need to check. Otherwise looks good, though I would split typo corrections into separate patch and submit without review.

[vpykhtin]

I think modifiers are checked correctly by the existing code, but can you add a test for e64 encodings into dpp_combine.mir similar to what is under "check for floating point modifiers" comment?

[foad]

Typo fixes have been committed separately.

[foad]

In D64207#1570456, @vpykhtin wrote:

I think modifiers are checked correctly by the existing code, but can you add a test for e64 encodings into dpp_combine.mir similar to what is under "check for floating point modifiers" comment?

I'm trying but I don't know enough about Machine IR. A typical e64 instruction from my dumps is:

%20:vgpr_32 = V_ADD_U32_e64 %18:vgpr_32, killed %19:vgpr_32, 0, implicit $exec

but I don't know what the third operand (the 0) means or what other values it can take.

[vpykhtin]

Right, this is hard to follow even for me :). 3rd operand is src1_modifiers, you can use a junk value for this to check whether the DPP combiner don't crash and don't combine it.

[foad]

In D64207#1570526, @vpykhtin wrote:

Right, this is hard to follow even for me :). 3rd operand is src1_modifiers, you can use a junk value for this to check whether the DPP combiner don't crash and don't combine it.

I still can't understand the instruction tables but according to my debugger, the third operand is clamp:

(gdb) call OrigMI.dump()
  %10:vgpr_32 = V_ADD_U32_e64 %9:vgpr_32, %1:vgpr_32, 99, implicit $exec
(gdb) p TII->getNamedOperand(OrigMI, AMDGPU::OpName::src0_modifiers)
$5 = (llvm::MachineOperand *) 0x0
(gdb) p TII->getNamedOperand(OrigMI, AMDGPU::OpName::src1_modifiers)
$6 = (llvm::MachineOperand *) 0x0
(gdb) p TII->getNamedOperand(OrigMI, AMDGPU::OpName::clamp)
$7 = (llvm::MachineOperand *) 0x6d1eef0
(gdb) p TII->getNamedOperand(OrigMI, AMDGPU::OpName::omod)
$8 = (llvm::MachineOperand *) 0x0
(gdb) call TII->getNamedOperand(OrigMI, AMDGPU::OpName::clamp)->dump()
99

[foad]

Add a test case, and change the tests to run on gfx9 because it needs
add-no-carry instructions.

[vpykhtin]

LGTM. Thank you!

llvm/trunk/lib/Target/AMDGPU/GCNDPPCombine.cpp

MachineInstr *GCNDPPCombine::createDPPInst(MachineInstr &OrigMI,
   return DPPInst.getInstr();
 }
 
 static bool isIdentityValue(unsigned OrigMIOp, MachineOperand *OldOpnd) {
   assert(OldOpnd->isImm());
   switch (OrigMIOp) {
   default: break;
   case AMDGPU::V_ADD_U32_e32:
+  case AMDGPU::V_ADD_U32_e64:
   case AMDGPU::V_ADD_I32_e32:
+  case AMDGPU::V_ADD_I32_e64:
   case AMDGPU::V_OR_B32_e32:
+  case AMDGPU::V_OR_B32_e64:
   case AMDGPU::V_SUBREV_U32_e32:
+  case AMDGPU::V_SUBREV_U32_e64:
   case AMDGPU::V_SUBREV_I32_e32:
+  case AMDGPU::V_SUBREV_I32_e64:
   case AMDGPU::V_MAX_U32_e32:
+  case AMDGPU::V_MAX_U32_e64:
   case AMDGPU::V_XOR_B32_e32:
+  case AMDGPU::V_XOR_B32_e64:
     if (OldOpnd->getImm() == 0)
       return true;
     break;
   case AMDGPU::V_AND_B32_e32:
+  case AMDGPU::V_AND_B32_e64:
   case AMDGPU::V_MIN_U32_e32:
+  case AMDGPU::V_MIN_U32_e64:
     if (static_cast<uint32_t>(OldOpnd->getImm()) ==
         std::numeric_limits<uint32_t>::max())
       return true;
     break;
   case AMDGPU::V_MIN_I32_e32:
+  case AMDGPU::V_MIN_I32_e64:
     if (static_cast<int32_t>(OldOpnd->getImm()) ==
         std::numeric_limits<int32_t>::max())
       return true;
     break;
   case AMDGPU::V_MAX_I32_e32:
+  case AMDGPU::V_MAX_I32_e64:
     if (static_cast<int32_t>(OldOpnd->getImm()) ==
         std::numeric_limits<int32_t>::min())
       return true;
     break;
   case AMDGPU::V_MUL_I32_I24_e32:
+  case AMDGPU::V_MUL_I32_I24_e64:
   case AMDGPU::V_MUL_U32_U24_e32:
+  case AMDGPU::V_MUL_U32_U24_e64:
     if (OldOpnd->getImm() == 1)
       return true;
     break;
   }
   return false;
 }
 
 MachineInstr *GCNDPPCombine::createDPPInst(MachineInstr &OrigMI,

llvm/trunk/test/CodeGen/AMDGPU/dpp_combine.mir

-# RUN: llc -march=amdgcn -mcpu=tonga  -run-pass=gcn-dpp-combine  -o - %s | FileCheck %s
+# RUN: llc -march=amdgcn -mcpu=gfx900 -run-pass=gcn-dpp-combine -o - %s | FileCheck %s
 
 ---
 # old is undefined: only combine when masks are fully enabled and
 # bound_ctrl:0 is set, otherwise the result of DPP VALU op can be undefined.
 # CHECK-LABEL: name: old_is_undef
 # CHECK: %2:vgpr_32 = IMPLICIT_DEF
 # VOP2:
 # CHECK: %4:vgpr_32 = V_ADD_U32_dpp %2, %0, %1, 1, 15, 15, 1, implicit $exec
bb.0:
     %7:vgpr_32 = V_MOV_B32_dpp undef %2, %1, 1, 15, 15, 1, implicit $exec
     %8:vgpr_32 = V_ADD_F32_e64 1, %7, 2, %0, 0, 0, implicit $exec
 
     ; this shouldn't be combined as modifiers aren't abs|neg
     %9:vgpr_32 = V_MOV_B32_dpp undef %2, %1, 1, 15, 15, 1, implicit $exec
     %10:vgpr_32 = V_ADD_F32_e64 4, %9, 8, %0, 0, 0, implicit $exec
 ...
 
+# check for e64 modifiers
+# CHECK-LABEL: name: add_u32_e64
+# CHECK: %4:vgpr_32 = V_ADD_U32_dpp %2, %0, %1, 1, 15, 15, 1, implicit $exec
+# CHECK: %6:vgpr_32 = V_ADD_U32_e64 %5, %1, 1, implicit $exec
+
+name: add_u32_e64
+tracksRegLiveness: true
+body:             |
+  bb.0:
+    liveins: $vgpr0, $vgpr1
+
+    %0:vgpr_32 = COPY $vgpr0
+    %1:vgpr_32 = COPY $vgpr1
+    %2:vgpr_32 = IMPLICIT_DEF
+
+    ; this should be combined as all modifiers are default
+    %3:vgpr_32 = V_MOV_B32_dpp undef %2, %0, 1, 15, 15, 1, implicit $exec
+    %4:vgpr_32 = V_ADD_U32_e64 %3, %1, 0, implicit $exec
+
+    ; this shouldn't be combined as clamp is set
+    %5:vgpr_32 = V_MOV_B32_dpp undef %2, %0, 1, 15, 15, 1, implicit $exec
+    %6:vgpr_32 = V_ADD_U32_e64 %5, %1, 1, implicit $exec
+...
+
 # tests on sequences of dpp consumers
 # CHECK-LABEL: name: dpp_seq
 # CHECK: %4:vgpr_32 = V_ADD_I32_dpp %1, %0, %1, 1, 14, 15, 0, implicit-def $vcc, implicit $exec
 # CHECK: %5:vgpr_32 = V_SUBREV_I32_dpp %1, %0, %1, 1, 14, 15, 0, implicit-def $vcc, implicit $exec
 # CHECK: %6:vgpr_32 = V_OR_B32_dpp %1, %0, %1, 1, 14, 15, 0, implicit $exec
 # broken sequence:
 # CHECK: %7:vgpr_32 = V_MOV_B32_dpp %2, %0, 1, 14, 15, 0, implicit $exec