[llvm] [AMDGPU] Remove support for ADDC/ADDE/SUBC/SUBE DAG nodes (PR #205082)

via llvm-commits llvm-commits at lists.llvm.org
Mon Jun 22 04:24:45 PDT 2026


llvmorg-github-actions[bot] wrote:


<!--LLVM PR SUMMARY COMMENT-->

@llvm/pr-subscribers-backend-amdgpu

Author: Jay Foad (jayfoad)

<details>
<summary>Changes</summary>

Since #<!-- -->204694 we should be using UADDO/USUBO/UADDO_CARRY/USUBO_CARRY
instead.


---

Patch is 283.95 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/205082.diff


15 Files Affected:

- (modified) llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp (-85) 
- (modified) llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp (+1-2) 
- (modified) llvm/lib/Target/AMDGPU/R600ISelLowering.cpp (-5) 
- (modified) llvm/lib/Target/AMDGPU/SOPInstructions.td (+2-11) 
- (modified) llvm/lib/Target/AMDGPU/VOP2Instructions.td (-3) 
- (modified) llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll (+406-406) 
- (modified) llvm/test/CodeGen/AMDGPU/bypass-div.ll (+258-258) 
- (modified) llvm/test/CodeGen/AMDGPU/carryout-selection.ll (+203-203) 
- (modified) llvm/test/CodeGen/AMDGPU/llvm.mulo.ll (+25-27) 
- (modified) llvm/test/CodeGen/AMDGPU/sdiv64.ll (+196-195) 
- (modified) llvm/test/CodeGen/AMDGPU/srem.ll (+551-551) 
- (modified) llvm/test/CodeGen/AMDGPU/srem64.ll (+243-242) 
- (modified) llvm/test/CodeGen/AMDGPU/udiv64.ll (+103-102) 
- (modified) llvm/test/CodeGen/AMDGPU/urem64.ll (+131-130) 
- (modified) llvm/test/CodeGen/AMDGPU/wave32.ll (+28-28) 


``````````diff
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
index ddd8b9e96d2d6..81053507b1ef5 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelDAGToDAG.cpp
@@ -691,19 +691,6 @@ void AMDGPUDAGToDAGISel::Select(SDNode *N) {
   switch (Opc) {
   default:
     break;
-  // We are selecting i64 ADD here instead of custom lower it during
-  // DAG legalization, so we can fold some i64 ADDs used for address
-  // calculation into the LOAD and STORE instructions.
-  case ISD::ADDC:
-  case ISD::ADDE:
-  case ISD::SUBC:
-  case ISD::SUBE: {
-    if (N->getValueType(0) != MVT::i64)
-      break;
-
-    SelectADD_SUB_I64(N);
-    return;
-  }
   case ISD::UADDO_CARRY:
   case ISD::USUBO_CARRY:
     if (N->getValueType(0) == MVT::i64) {
@@ -1057,78 +1044,6 @@ SDValue AMDGPUDAGToDAGISel::getMaterializedScalarImm32(int64_t Val,
   return SDValue(Mov, 0);
 }
 
-// Keep this as a fallback for i64 ADDC/ADDE/SUBC/SUBE glue nodes. Wide integer
-// add/sub should normally expand through the explicit carry nodes handled in
-// SelectAddcSubbI64.
-void AMDGPUDAGToDAGISel::SelectADD_SUB_I64(SDNode *N) {
-  SDLoc DL(N);
-  SDValue LHS = N->getOperand(0);
-  SDValue RHS = N->getOperand(1);
-
-  unsigned Opcode = N->getOpcode();
-  bool ConsumeCarry = (Opcode == ISD::ADDE || Opcode == ISD::SUBE);
-  bool ProduceCarry =
-      ConsumeCarry || Opcode == ISD::ADDC || Opcode == ISD::SUBC;
-  bool IsAdd = Opcode == ISD::ADD || Opcode == ISD::ADDC || Opcode == ISD::ADDE;
-
-  SDValue Sub0 = CurDAG->getTargetConstant(AMDGPU::sub0, DL, MVT::i32);
-  SDValue Sub1 = CurDAG->getTargetConstant(AMDGPU::sub1, DL, MVT::i32);
-
-  SDNode *Lo0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
-                                       DL, MVT::i32, LHS, Sub0);
-  SDNode *Hi0 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
-                                       DL, MVT::i32, LHS, Sub1);
-
-  SDNode *Lo1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
-                                       DL, MVT::i32, RHS, Sub0);
-  SDNode *Hi1 = CurDAG->getMachineNode(TargetOpcode::EXTRACT_SUBREG,
-                                       DL, MVT::i32, RHS, Sub1);
-
-  SDVTList VTList = CurDAG->getVTList(MVT::i32, MVT::Glue);
-
-  static const unsigned OpcMap[2][2][2] = {
-      {{AMDGPU::S_SUB_U32, AMDGPU::S_ADD_U32},
-       {AMDGPU::V_SUB_CO_U32_e32, AMDGPU::V_ADD_CO_U32_e32}},
-      {{AMDGPU::S_SUBB_U32, AMDGPU::S_ADDC_U32},
-       {AMDGPU::V_SUBB_U32_e32, AMDGPU::V_ADDC_U32_e32}}};
-
-  unsigned Opc = OpcMap[0][N->isDivergent()][IsAdd];
-  unsigned CarryOpc = OpcMap[1][N->isDivergent()][IsAdd];
-
-  SDNode *AddLo;
-  if (!ConsumeCarry) {
-    SDValue Args[] = { SDValue(Lo0, 0), SDValue(Lo1, 0) };
-    AddLo = CurDAG->getMachineNode(Opc, DL, VTList, Args);
-  } else {
-    SDValue Args[] = { SDValue(Lo0, 0), SDValue(Lo1, 0), N->getOperand(2) };
-    AddLo = CurDAG->getMachineNode(CarryOpc, DL, VTList, Args);
-  }
-  SDValue AddHiArgs[] = {
-    SDValue(Hi0, 0),
-    SDValue(Hi1, 0),
-    SDValue(AddLo, 1)
-  };
-  SDNode *AddHi = CurDAG->getMachineNode(CarryOpc, DL, VTList, AddHiArgs);
-
-  SDValue RegSequenceArgs[] = {
-    CurDAG->getTargetConstant(AMDGPU::SReg_64RegClassID, DL, MVT::i32),
-    SDValue(AddLo,0),
-    Sub0,
-    SDValue(AddHi,0),
-    Sub1,
-  };
-  SDNode *RegSequence = CurDAG->getMachineNode(AMDGPU::REG_SEQUENCE, DL,
-                                               MVT::i64, RegSequenceArgs);
-
-  if (ProduceCarry) {
-    // Replace the carry-use
-    ReplaceUses(SDValue(N, 1), SDValue(AddHi, 1));
-  }
-
-  // Replace the remaining uses.
-  ReplaceNode(N, RegSequence);
-}
-
 void AMDGPUDAGToDAGISel::SelectAddcSubb(SDNode *N) {
   SDValue LHS = N->getOperand(0);
   SDValue RHS = N->getOperand(1);
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 7c614670c5dfa..12e6301a2caa1 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -492,8 +492,7 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
 
     setOperationAction({ISD::BSWAP, ISD::CTTZ, ISD::CTLZ}, VT, Expand);
 
-    // AMDGPU uses ADDC/SUBC/ADDE/SUBE
-    setOperationAction({ISD::ADDC, ISD::SUBC, ISD::ADDE, ISD::SUBE}, VT, Legal);
+    setOperationAction({ISD::ADDC, ISD::SUBC, ISD::ADDE, ISD::SUBE}, VT, Expand);
   }
 
   // The hardware supports 32-bit FSHR, but not FSHL.
diff --git a/llvm/lib/Target/AMDGPU/R600ISelLowering.cpp b/llvm/lib/Target/AMDGPU/R600ISelLowering.cpp
index ae8e73eb5f329..a3825ac909881 100644
--- a/llvm/lib/Target/AMDGPU/R600ISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/R600ISelLowering.cpp
@@ -190,11 +190,6 @@ R600TargetLowering::R600TargetLowering(const TargetMachine &TM,
   setOperationAction(ISD::GlobalAddress, MVT::i32, Custom);
   setOperationAction(ISD::ADDRSPACECAST, MVT::i32, Custom);
 
-  const MVT ScalarIntVTs[] = { MVT::i32, MVT::i64 };
-  for (MVT VT : ScalarIntVTs)
-    setOperationAction({ISD::ADDC, ISD::SUBC, ISD::ADDE, ISD::SUBE}, VT,
-                       Expand);
-
   // LLVM will expand these to atomic_cmp_swap(0)
   // and atomic_swap, respectively.
   setOperationAction({ISD::ATOMIC_LOAD, ISD::ATOMIC_STORE}, MVT::i32, Expand);
diff --git a/llvm/lib/Target/AMDGPU/SOPInstructions.td b/llvm/lib/Target/AMDGPU/SOPInstructions.td
index 01dd444c477e2..b28195be51733 100644
--- a/llvm/lib/Target/AMDGPU/SOPInstructions.td
+++ b/llvm/lib/Target/AMDGPU/SOPInstructions.td
@@ -700,12 +700,10 @@ def S_SUB_I32 : SOP2_32 <"s_sub_i32",
 
 let Uses = [SCC] in { // Carry in comes from SCC
 let isCommutable = 1 in {
-def S_ADDC_U32 : SOP2_32 <"s_addc_u32",
-  [(set i32:$sdst, (UniformBinFrag<adde> (i32 SSrc_b32:$src0), (i32 SSrc_b32:$src1)))]>;
+def S_ADDC_U32 : SOP2_32 <"s_addc_u32">;
 } // End isCommutable = 1
 
-def S_SUBB_U32 : SOP2_32 <"s_subb_u32",
-  [(set i32:$sdst, (UniformBinFrag<sube> (i32 SSrc_b32:$src0), (i32 SSrc_b32:$src1)))]>;
+def S_SUBB_U32 : SOP2_32 <"s_subb_u32">;
 } // End Uses = [SCC]
 
 let isCommutable = 1 in {
@@ -2072,13 +2070,6 @@ let AddedComplexity = 20 in {
   >;
 }
 
-// V_ADD_I32_e32/S_ADD_U32 produces carry in VCC/SCC. For the vector
-// case, the sgpr-copies pass will fix this to use the vector version.
-def : GCNPat <
-  (i32 (addc i32:$src0, i32:$src1)),
-  (S_ADD_U32 $src0, $src1)
->;
-
 // FIXME: We need to use COPY_TO_REGCLASS to work-around the fact that
 // REG_SEQUENCE patterns don't support instructions with multiple
 // outputs.
diff --git a/llvm/lib/Target/AMDGPU/VOP2Instructions.td b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
index 74864e10bf4a5..bcf2ca6e8729e 100644
--- a/llvm/lib/Target/AMDGPU/VOP2Instructions.td
+++ b/llvm/lib/Target/AMDGPU/VOP2Instructions.td
@@ -1040,9 +1040,6 @@ def : DivergentClampingBinOp<add, V_ADD_CO_U32_e64>;
 def : DivergentClampingBinOp<sub, V_SUB_CO_U32_e64>;
 }
 
-def : DivergentBinOp<adde, V_ADDC_U32_e32>;
-def : DivergentBinOp<sube, V_SUBB_U32_e32>;
-
 class divergent_i64_BinOp <SDPatternOperator Op, Instruction Inst, ValueType vt = i64> :
   GCNPat<
       (DivergentBinFrag<Op> vt:$src0, vt:$src1),
diff --git a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll
index d83bb728e47f9..4a33008661b8e 100644
--- a/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll
+++ b/llvm/test/CodeGen/AMDGPU/amdgpu-codegenprepare-idiv.ll
@@ -7681,6 +7681,7 @@ define amdgpu_kernel void @sdiv_i64_pow2_shl_denom(ptr addrspace(1) %out, i64 %x
 ; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0xd
 ; GFX6-NEXT:    s_load_dwordx4 s[4:7], s[4:5], 0x9
 ; GFX6-NEXT:    s_mov_b32 s3, 0xf000
+; GFX6-NEXT:    s_mov_b32 s2, -1
 ; GFX6-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX6-NEXT:    s_lshl_b64 s[0:1], 0x1000, s0
 ; GFX6-NEXT:    s_ashr_i32 s8, s1, 31
@@ -7702,60 +7703,60 @@ define amdgpu_kernel void @sdiv_i64_pow2_shl_denom(ptr addrspace(1) %out, i64 %x
 ; GFX6-NEXT:    v_cvt_u32_f32_e32 v1, v1
 ; GFX6-NEXT:    v_mul_hi_u32 v2, s0, v0
 ; GFX6-NEXT:    v_readfirstlane_b32 s12, v1
-; GFX6-NEXT:    v_readfirstlane_b32 s2, v0
-; GFX6-NEXT:    s_mul_i32 s13, s0, s12
-; GFX6-NEXT:    v_readfirstlane_b32 s16, v2
-; GFX6-NEXT:    s_mul_i32 s14, s1, s2
-; GFX6-NEXT:    s_mul_i32 s15, s0, s2
-; GFX6-NEXT:    s_add_i32 s13, s16, s13
-; GFX6-NEXT:    v_mul_hi_u32 v3, v0, s15
-; GFX6-NEXT:    s_add_i32 s13, s13, s14
-; GFX6-NEXT:    v_mul_hi_u32 v0, v0, s13
-; GFX6-NEXT:    v_mul_hi_u32 v4, v1, s15
-; GFX6-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX6-NEXT:    s_mul_i32 s16, s2, s13
-; GFX6-NEXT:    v_mul_hi_u32 v1, v1, s13
-; GFX6-NEXT:    s_add_u32 s14, s14, s16
+; GFX6-NEXT:    v_readfirstlane_b32 s13, v0
+; GFX6-NEXT:    s_mul_i32 s14, s0, s12
+; GFX6-NEXT:    v_readfirstlane_b32 s17, v2
+; GFX6-NEXT:    s_mul_i32 s15, s1, s13
+; GFX6-NEXT:    s_mul_i32 s16, s0, s13
+; GFX6-NEXT:    s_add_i32 s14, s17, s14
+; GFX6-NEXT:    v_mul_hi_u32 v3, v0, s16
+; GFX6-NEXT:    s_add_i32 s14, s14, s15
+; GFX6-NEXT:    v_mul_hi_u32 v0, v0, s14
+; GFX6-NEXT:    v_mul_hi_u32 v4, v1, s16
+; GFX6-NEXT:    v_readfirstlane_b32 s15, v3
+; GFX6-NEXT:    s_mul_i32 s17, s13, s14
+; GFX6-NEXT:    s_add_u32 s15, s15, s17
+; GFX6-NEXT:    v_readfirstlane_b32 s17, v0
+; GFX6-NEXT:    v_mul_hi_u32 v0, v1, s14
+; GFX6-NEXT:    s_mul_i32 s16, s12, s16
+; GFX6-NEXT:    s_addc_u32 s17, 0, s17
+; GFX6-NEXT:    v_readfirstlane_b32 s18, v4
+; GFX6-NEXT:    s_add_u32 s15, s15, s16
+; GFX6-NEXT:    s_addc_u32 s15, s17, s18
 ; GFX6-NEXT:    v_readfirstlane_b32 s16, v0
-; GFX6-NEXT:    s_mul_i32 s15, s12, s15
-; GFX6-NEXT:    s_addc_u32 s16, 0, s16
-; GFX6-NEXT:    v_readfirstlane_b32 s17, v4
-; GFX6-NEXT:    s_add_u32 s14, s14, s15
-; GFX6-NEXT:    s_addc_u32 s14, s16, s17
+; GFX6-NEXT:    s_addc_u32 s16, s16, 0
+; GFX6-NEXT:    s_mul_i32 s14, s12, s14
+; GFX6-NEXT:    s_add_u32 s14, s15, s14
+; GFX6-NEXT:    s_addc_u32 s15, 0, s16
+; GFX6-NEXT:    s_add_u32 s13, s13, s14
+; GFX6-NEXT:    v_mov_b32_e32 v1, s13
+; GFX6-NEXT:    v_mul_hi_u32 v1, s0, v1
+; GFX6-NEXT:    s_addc_u32 s12, s12, s15
+; GFX6-NEXT:    s_mul_i32 s14, s0, s13
+; GFX6-NEXT:    s_mul_i32 s0, s0, s12
 ; GFX6-NEXT:    v_readfirstlane_b32 s15, v1
-; GFX6-NEXT:    s_addc_u32 s15, s15, 0
-; GFX6-NEXT:    s_mul_i32 s13, s12, s13
-; GFX6-NEXT:    s_add_u32 s13, s14, s13
-; GFX6-NEXT:    s_addc_u32 s14, 0, s15
-; GFX6-NEXT:    s_add_u32 s13, s2, s13
-; GFX6-NEXT:    v_mov_b32_e32 v0, s13
-; GFX6-NEXT:    v_mul_hi_u32 v0, s0, v0
-; GFX6-NEXT:    s_addc_u32 s12, s12, s14
-; GFX6-NEXT:    s_mul_i32 s14, s0, s12
+; GFX6-NEXT:    s_add_i32 s0, s15, s0
 ; GFX6-NEXT:    s_mul_i32 s1, s1, s13
-; GFX6-NEXT:    v_readfirstlane_b32 s15, v0
-; GFX6-NEXT:    s_add_i32 s14, s15, s14
-; GFX6-NEXT:    s_mul_i32 s0, s0, s13
-; GFX6-NEXT:    s_add_i32 s1, s14, s1
-; GFX6-NEXT:    v_mov_b32_e32 v2, s0
-; GFX6-NEXT:    v_mov_b32_e32 v0, s1
-; GFX6-NEXT:    v_mul_hi_u32 v3, s12, v2
-; GFX6-NEXT:    v_mul_hi_u32 v2, s13, v2
-; GFX6-NEXT:    v_mul_hi_u32 v1, s12, v0
+; GFX6-NEXT:    v_mov_b32_e32 v0, s14
+; GFX6-NEXT:    s_add_i32 s0, s0, s1
+; GFX6-NEXT:    v_mul_hi_u32 v2, s12, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s0
 ; GFX6-NEXT:    v_mul_hi_u32 v0, s13, v0
-; GFX6-NEXT:    s_mul_i32 s15, s13, s1
-; GFX6-NEXT:    v_readfirstlane_b32 s17, v2
+; GFX6-NEXT:    v_mul_hi_u32 v3, s13, v1
+; GFX6-NEXT:    s_mul_i32 s15, s13, s0
+; GFX6-NEXT:    s_mul_i32 s14, s12, s14
+; GFX6-NEXT:    v_readfirstlane_b32 s17, v0
+; GFX6-NEXT:    v_mul_hi_u32 v0, s12, v1
+; GFX6-NEXT:    v_readfirstlane_b32 s16, v3
 ; GFX6-NEXT:    s_add_u32 s15, s17, s15
-; GFX6-NEXT:    v_readfirstlane_b32 s16, v0
-; GFX6-NEXT:    s_mul_i32 s0, s12, s0
 ; GFX6-NEXT:    s_addc_u32 s16, 0, s16
-; GFX6-NEXT:    v_readfirstlane_b32 s14, v3
-; GFX6-NEXT:    s_add_u32 s0, s15, s0
-; GFX6-NEXT:    s_addc_u32 s0, s16, s14
-; GFX6-NEXT:    v_readfirstlane_b32 s14, v1
+; GFX6-NEXT:    v_readfirstlane_b32 s1, v2
+; GFX6-NEXT:    s_add_u32 s14, s15, s14
+; GFX6-NEXT:    s_addc_u32 s1, s16, s1
+; GFX6-NEXT:    v_readfirstlane_b32 s14, v0
 ; GFX6-NEXT:    s_addc_u32 s14, s14, 0
-; GFX6-NEXT:    s_mul_i32 s1, s12, s1
-; GFX6-NEXT:    s_add_u32 s0, s0, s1
+; GFX6-NEXT:    s_mul_i32 s0, s12, s0
+; GFX6-NEXT:    s_add_u32 s0, s1, s0
 ; GFX6-NEXT:    s_addc_u32 s1, 0, s14
 ; GFX6-NEXT:    s_add_u32 s14, s13, s0
 ; GFX6-NEXT:    s_addc_u32 s15, s12, s1
@@ -7829,7 +7830,6 @@ define amdgpu_kernel void @sdiv_i64_pow2_shl_denom(ptr addrspace(1) %out, i64 %x
 ; GFX6-NEXT:    s_sub_u32 s4, s4, s6
 ; GFX6-NEXT:    s_subb_u32 s5, s5, s7
 ; GFX6-NEXT:    v_mov_b32_e32 v0, s4
-; GFX6-NEXT:    s_mov_b32 s2, -1
 ; GFX6-NEXT:    v_mov_b32_e32 v1, s5
 ; GFX6-NEXT:    buffer_store_dwordx2 v[0:1], off, s[0:3], 0
 ; GFX6-NEXT:    s_endpgm
@@ -7870,37 +7870,37 @@ define amdgpu_kernel void @sdiv_i64_pow2_shl_denom(ptr addrspace(1) %out, i64 %x
 ; GFX9-NEXT:    s_mul_hi_u32 s13, s11, s12
 ; GFX9-NEXT:    s_add_u32 s14, s14, s16
 ; GFX9-NEXT:    s_addc_u32 s13, 0, s13
-; GFX9-NEXT:    s_mul_hi_u32 s17, s10, s15
+; GFX9-NEXT:    s_mul_hi_u32 s16, s10, s15
 ; GFX9-NEXT:    s_mul_i32 s15, s10, s15
 ; GFX9-NEXT:    s_add_u32 s14, s14, s15
-; GFX9-NEXT:    s_mul_hi_u32 s16, s10, s12
-; GFX9-NEXT:    s_addc_u32 s13, s13, s17
-; GFX9-NEXT:    s_addc_u32 s14, s16, 0
+; GFX9-NEXT:    s_addc_u32 s13, s13, s16
+; GFX9-NEXT:    s_mul_hi_u32 s14, s10, s12
+; GFX9-NEXT:    s_addc_u32 s14, s14, 0
 ; GFX9-NEXT:    s_mul_i32 s12, s10, s12
 ; GFX9-NEXT:    s_add_u32 s12, s13, s12
 ; GFX9-NEXT:    s_addc_u32 s13, 0, s14
 ; GFX9-NEXT:    s_add_u32 s11, s11, s12
 ; GFX9-NEXT:    s_addc_u32 s10, s10, s13
-; GFX9-NEXT:    s_mul_i32 s12, s4, s10
-; GFX9-NEXT:    s_mul_hi_u32 s13, s4, s11
-; GFX9-NEXT:    s_add_i32 s12, s13, s12
+; GFX9-NEXT:    s_mul_i32 s12, s4, s11
+; GFX9-NEXT:    s_mul_i32 s15, s4, s10
+; GFX9-NEXT:    s_mul_hi_u32 s4, s4, s11
+; GFX9-NEXT:    s_add_i32 s4, s4, s15
 ; GFX9-NEXT:    s_mul_i32 s5, s5, s11
-; GFX9-NEXT:    s_add_i32 s12, s12, s5
-; GFX9-NEXT:    s_mul_i32 s4, s4, s11
-; GFX9-NEXT:    s_mul_hi_u32 s13, s10, s4
-; GFX9-NEXT:    s_mul_i32 s14, s10, s4
-; GFX9-NEXT:    s_mul_i32 s16, s11, s12
-; GFX9-NEXT:    s_mul_hi_u32 s4, s11, s4
-; GFX9-NEXT:    s_mul_hi_u32 s15, s11, s12
-; GFX9-NEXT:    s_add_u32 s4, s4, s16
-; GFX9-NEXT:    s_addc_u32 s15, 0, s15
-; GFX9-NEXT:    s_add_u32 s4, s4, s14
-; GFX9-NEXT:    s_mul_hi_u32 s5, s10, s12
-; GFX9-NEXT:    s_addc_u32 s4, s15, s13
-; GFX9-NEXT:    s_addc_u32 s5, s5, 0
-; GFX9-NEXT:    s_mul_i32 s12, s10, s12
-; GFX9-NEXT:    s_add_u32 s4, s4, s12
+; GFX9-NEXT:    s_add_i32 s4, s4, s5
+; GFX9-NEXT:    s_mul_hi_u32 s13, s10, s12
+; GFX9-NEXT:    s_mul_i32 s14, s10, s12
+; GFX9-NEXT:    s_mul_i32 s15, s11, s4
+; GFX9-NEXT:    s_mul_hi_u32 s12, s11, s12
+; GFX9-NEXT:    s_mul_hi_u32 s5, s11, s4
+; GFX9-NEXT:    s_add_u32 s12, s12, s15
 ; GFX9-NEXT:    s_addc_u32 s5, 0, s5
+; GFX9-NEXT:    s_add_u32 s12, s12, s14
+; GFX9-NEXT:    s_addc_u32 s5, s5, s13
+; GFX9-NEXT:    s_mul_hi_u32 s12, s10, s4
+; GFX9-NEXT:    s_addc_u32 s12, s12, 0
+; GFX9-NEXT:    s_mul_i32 s4, s10, s4
+; GFX9-NEXT:    s_add_u32 s4, s5, s4
+; GFX9-NEXT:    s_addc_u32 s5, 0, s12
 ; GFX9-NEXT:    s_add_u32 s11, s11, s4
 ; GFX9-NEXT:    s_addc_u32 s10, s10, s5
 ; GFX9-NEXT:    s_waitcnt lgkmcnt(0)
@@ -7914,12 +7914,12 @@ define amdgpu_kernel void @sdiv_i64_pow2_shl_denom(ptr addrspace(1) %out, i64 %x
 ; GFX9-NEXT:    s_mul_hi_u32 s12, s2, s10
 ; GFX9-NEXT:    s_add_u32 s13, s14, s13
 ; GFX9-NEXT:    s_addc_u32 s12, 0, s12
-; GFX9-NEXT:    s_mul_hi_u32 s15, s3, s11
+; GFX9-NEXT:    s_mul_hi_u32 s14, s3, s11
 ; GFX9-NEXT:    s_mul_i32 s11, s3, s11
 ; GFX9-NEXT:    s_add_u32 s11, s13, s11
-; GFX9-NEXT:    s_mul_hi_u32 s14, s3, s10
-; GFX9-NEXT:    s_addc_u32 s11, s12, s15
-; GFX9-NEXT:    s_addc_u32 s12, s14, 0
+; GFX9-NEXT:    s_addc_u32 s11, s12, s14
+; GFX9-NEXT:    s_mul_hi_u32 s12, s3, s10
+; GFX9-NEXT:    s_addc_u32 s12, s12, 0
 ; GFX9-NEXT:    s_mul_i32 s10, s3, s10
 ; GFX9-NEXT:    s_add_u32 s13, s11, s10
 ; GFX9-NEXT:    s_addc_u32 s12, 0, s12
@@ -8218,34 +8218,34 @@ define amdgpu_kernel void @sdiv_v2i64_pow2_shl_denom(ptr addrspace(1) %out, <2 x
 ; GFX6-NEXT:    s_add_u32 s16, s17, s16
 ; GFX6-NEXT:    s_addc_u32 s17, 0, s18
 ; GFX6-NEXT:    s_add_u32 s15, s15, s16
-; GFX6-NEXT:    v_mov_b32_e32 v0, s15
-; GFX6-NEXT:    v_mul_hi_u32 v0, s12, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s15
+; GFX6-NEXT:    v_mul_hi_u32 v1, s12, v1
 ; GFX6-NEXT:    s_addc_u32 s14, s14, s17
-; GFX6-NEXT:    s_mul_i32 s16, s12, s14
+; GFX6-NEXT:    s_mul_i32 s16, s12, s15
+; GFX6-NEXT:    s_mul_i32 s12, s12, s14
+; GFX6-NEXT:    v_readfirstlane_b32 s17, v1
+; GFX6-NEXT:    s_add_i32 s12, s17, s12
 ; GFX6-NEXT:    s_mul_i32 s13, s13, s15
-; GFX6-NEXT:    v_readfirstlane_b32 s17, v0
-; GFX6-NEXT:    s_add_i32 s16, s17, s16
-; GFX6-NEXT:    s_mul_i32 s12, s12, s15
-; GFX6-NEXT:    s_add_i32 s13, s16, s13
-; GFX6-NEXT:    v_mov_b32_e32 v2, s12
-; GFX6-NEXT:    v_mov_b32_e32 v0, s13
-; GFX6-NEXT:    v_mul_hi_u32 v3, s14, v2
-; GFX6-NEXT:    v_mul_hi_u32 v2, s15, v2
-; GFX6-NEXT:    v_mul_hi_u32 v1, s14, v0
+; GFX6-NEXT:    v_mov_b32_e32 v0, s16
+; GFX6-NEXT:    s_add_i32 s12, s12, s13
+; GFX6-NEXT:    v_mul_hi_u32 v2, s14, v0
+; GFX6-NEXT:    v_mov_b32_e32 v1, s12
 ; GFX6-NEXT:    v_mul_hi_u32 v0, s15, v0
-; GFX6-NEXT:    s_mul_i32 s17, s15, s13
-; GFX6-NEXT:    v_readfirstlane_b32 s19, v2
+; GFX6-NEXT:    v_mul_hi_u32 v3, s15, v1
+; GFX6-NEXT:    s_mul_i32 s17, s15, s12
+; GFX6-NEXT:    s_mul_i32 s16, s14, s16
+; GFX6-NEXT:    v_readfirstlane_b32 s19, v0
+; GFX6-NEXT:    v_mul_hi_u32 v0, s14, v1
+; GFX6-NEXT:    v_readfirstlane_b32 s18, v3
 ; GFX6-NEXT:    s_add_u32 s17, s19, s17
-; GFX6-NEXT:    v_readfirstlane_b32 s18, v0
-; GFX6-NEXT:    s_mul_i32 s12, s14, s12
 ; GFX6-NEXT:    s_addc_u32 s18, 0, s18
-; GFX6-NEXT:    v_readfirstlane_b32 s16, v3
-; GFX6-NEXT:    s_add_u32 s12, s17, s12
-; GFX6-NEXT:    s_addc_u32 s12, s18, s16
-; GFX6-NEXT:    v_readfirstlane_b32 s16, v1
+; GFX6-NEXT:    v_readfirstlane_b32 s13, v2
+; GFX6-NEXT:    s_add_u32 s16, s17, s16
+; GFX6-NEXT:    s_addc_u32 s13, s18, s13
+; GFX6-NEXT:    v_readfirstlane_b32 s16, v0
 ; GFX6-NEXT:    s_addc_u32 s16, s16, 0
-; GFX6-NEXT:    s_mul_i32 s13, s14, s13
-; GFX6-NEXT:    s_add_u32 s12, s12, s13
+; GFX6-NEXT:    s_mul_i32 s12, s14, s12
+; GFX6-NEXT:    s_add_u32 s12, s13, s12
 ; GFX6-NEXT:    s_addc_u32 s13, 0, s16
 ; GFX6-NEXT:    s_add_u32 s15, s15, s12
 ; GFX6-NEXT:    s_addc_u32 s14, s14, s13
@@ -8334,65 +8334,65 @@ define amdgpu_kernel void @sdiv_v2i64_pow2_shl_denom(ptr addrspace(1) %out, <2 x
 ; GFX6-NEXT:    v_cvt_u32_f32_e32 v0, v0
 ; GFX6-NEXT:    v_cvt_u32_f32_e32 v1, v1
 ; GFX6-NEXT:    v_mul_hi_u32 v2, s2, v0
-; GFX6-NEXT:    v_readfirstlane_b32 s12, v1
-; GFX6-NEXT:    v_readfirstlane_b32 s0, v0
-; GFX6-NEXT:    s_mul_i32 s13, s2, s12
+; GFX6-NEXT:    v_readfirstlane_b32 s13, v1
+; GFX6-NEXT:    v_readfirstlane_b32 s12, v0
+; GFX6-NEXT:    s_mul_i32 s1, s2, s13
 ; GFX6-NEXT:    v_readfirstlane_b32 s16, v2
-; GFX6-NEXT:    s_mul_i32 s1, s3, s0
-; GFX6-NEXT:    s_add_i32 s13, s16, s13
-; GFX6-NEXT:    s_add_i32 s13, s13, s1
-; GFX6-NEXT:    s_mul_i32 s1, s2, s0
-; GFX6-NEXT:    v_mul_hi_u32 v2, v0, s13
-; GFX6-NEXT:    v_mul_hi_u32 v0, v0, s1
-; GFX6-NEXT:    s_mul_i32 s16, s0, s13
-; GFX6-NEXT:    v_readfirstlane_b32 s17, v2
+; GFX6-NEXT:    s_mul_i32 s0, s3, s12
+; GFX6-NEXT:    s_add_i32 s1, s16, s1
+; GFX6-NEXT:    s_add_i32 s16, s1, s0
+; GFX6-NEXT:    s_mul_i32 s17, s2, s12
+; GFX6-NEXT:    v_mul_hi_u32 v2, v0, s16
+; GFX6-NEXT:    v_mul_hi_u32 v0, v0, s17
+; GFX6-NEXT:    s_load_dwordx2 s[0:1], s[4:5], 0x9
+; GFX6-NEXT:    s_mul_i32 s4, s12, s16
+; GFX6-NEXT:    v_readfirstlane_b32 s5, v2
 ; GFX6-NEXT:    v_readfirstlane_b32 s18, v0
-; GFX6-NEXT:    v_mul_hi_u32 v0, v1, s1
-; GFX6-NEXT:    v_mul_hi_u32 v1, v1, s13
-; GFX6-NEXT:    s_add_u32 s16, s18, s16
-; GFX6-NEXT:    s_addc_u32 s17, 0, s17
-; GFX6-NEXT:    s_mul_i32 s1, s12, s1
+; GFX6-NEXT:    v_mul_hi_u32 v0, v1, s17
+; GFX6-NEXT:    v_mul_hi_u32 v1, v1, s16
+; GFX6-NEXT:    s_add_u32 s4, s18, s4
+; GFX6-NEXT:    s_addc_u32 s5, 0, s5
+; GFX6-NEXT:    s_mul_i32 s17, s13, s17
 ; GFX6-NEXT:    v_readfirstlane_b32 s18, v0
-; GFX6-NEXT:    s_add_u32 s1, s16, s1
-; GFX6-NEXT:    s_addc_u32 s1, s17, s18
-; GFX6-NEXT:    v_readfirstlane_b32 s16, v1
-; GFX6-NEXT:    s_addc_u32 s16, s16, 0
-; GFX6-NEXT:    s_mul_i32 s13, s12, s13
-; GFX6-NEXT:  ...
[truncated]

``````````

</details>


https://github.com/llvm/llvm-project/pull/205082


More information about the llvm-commits mailing list