[llvm] [AMDGPU] Remove unnecessary add instructions in ctlz.i8 (PR #77615)

Leon Clark via llvm-commits llvm-commits at lists.llvm.org
Mon Jan 15 13:08:39 PST 2024


https://github.com/PeddleSpam updated https://github.com/llvm/llvm-project/pull/77615

>From ca2b9028da065b00f298f1610b40f97d6a8bd9b6 Mon Sep 17 00:00:00 2001
From: Leon Clark <leoclark at amd.com>
Date: Wed, 10 Jan 2024 14:30:26 +0000
Subject: [PATCH 1/3] Remove unnecessary add instructions in ctlz.i8

Add custom lowering for ctlz.i8 to avoid multiple add/sub operations.
---
 llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 25 +++++++++++++++++++
 llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h   |  3 +++
 llvm/test/CodeGen/AMDGPU/ctlz.ll              | 10 +++-----
 llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll   | 10 +++-----
 4 files changed, 35 insertions(+), 13 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 0dbcaf5a1b136c..2c36e6e01e5539 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -446,6 +446,8 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
       {ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF, ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF},
       MVT::i64, Custom);
 
+  setOperationAction({ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF}, MVT::i8, Custom);
+
   static const MVT::SimpleValueType VectorIntTypes[] = {
       MVT::v2i32, MVT::v3i32, MVT::v4i32, MVT::v5i32, MVT::v6i32, MVT::v7i32,
       MVT::v9i32, MVT::v10i32, MVT::v11i32, MVT::v12i32};
@@ -1397,6 +1399,10 @@ void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
     if (SDValue Lowered = lowerFEXP(SDValue(N, 0), DAG))
       Results.push_back(Lowered);
     return;
+  case ISD::CTLZ:
+  case ISD::CTLZ_ZERO_UNDEF:
+    replaceCTLZResults(SDValue(N, 0u), DAG, Results);
+    return;
   default:
     return;
   }
@@ -3062,6 +3068,25 @@ static bool isCttzOpc(unsigned Opc) {
   return Opc == ISD::CTTZ || Opc == ISD::CTTZ_ZERO_UNDEF;
 }
 
+void AMDGPUTargetLowering::replaceCTLZResults(
+    SDValue Op, SelectionDAG &DAG, SmallVectorImpl<SDValue> &Results) const {
+  auto SL = SDLoc(Op);
+  auto Arg = Op.getOperand(0u);
+  auto ResultVT = Op.getValueType();
+
+  if (ResultVT != MVT::i8)
+    return;
+
+  assert(isCtlzOpc(Op.getOpcode()));
+  assert(ResultVT == Arg.getValueType());
+
+  auto SubVal = DAG.getConstant(24u, SL, MVT::i32);
+  auto NewOp = DAG.getNode(ISD::ZERO_EXTEND, SL, MVT::i32, Arg);
+  NewOp = DAG.getNode(Op.getOpcode(), SL, MVT::i32, NewOp);
+  NewOp = DAG.getNode(ISD::SUB, SL, MVT::i32, NewOp, SubVal);
+  Results.push_back(DAG.getNode(ISD::TRUNCATE, SL, ResultVT, NewOp));
+}
+
 SDValue AMDGPUTargetLowering::LowerCTLZ_CTTZ(SDValue Op, SelectionDAG &DAG) const {
   SDLoc SL(Op);
   SDValue Src = Op.getOperand(0);
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index 827fb106b55199..4cbd6c3146d5dc 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -84,6 +84,9 @@ class AMDGPUTargetLowering : public TargetLowering {
                             SDNodeFlags Flags) const;
   SDValue lowerFEXP(SDValue Op, SelectionDAG &DAG) const;
 
+  void replaceCTLZResults(SDValue Op, SelectionDAG &DAG,
+                          SmallVectorImpl<SDValue> &Results) const;
+
   SDValue LowerCTLZ_CTTZ(SDValue Op, SelectionDAG &DAG) const;
 
   SDValue LowerINT_TO_FP32(SDValue Op, SelectionDAG &DAG, bool Signed) const;
diff --git a/llvm/test/CodeGen/AMDGPU/ctlz.ll b/llvm/test/CodeGen/AMDGPU/ctlz.ll
index 3d69655111da63..2267aa6b5496c6 100644
--- a/llvm/test/CodeGen/AMDGPU/ctlz.ll
+++ b/llvm/test/CodeGen/AMDGPU/ctlz.ll
@@ -514,8 +514,7 @@ define amdgpu_kernel void @v_ctlz_i8(ptr addrspace(1) noalias %out, ptr addrspac
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_ffbh_u32_e32 v0, v0
 ; VI-NEXT:    v_min_u32_e32 v0, 32, v0
-; VI-NEXT:    v_add_u32_e32 v0, vcc, -16, v0
-; VI-NEXT:    v_add_u16_e32 v0, -8, v0
+; VI-NEXT:    v_subrev_u32_e32 v0, vcc, 24, v0
 ; VI-NEXT:    buffer_store_byte v0, off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -558,8 +557,7 @@ define amdgpu_kernel void @v_ctlz_i8(ptr addrspace(1) noalias %out, ptr addrspac
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
 ; GFX10-NEXT:    v_ffbh_u32_e32 v1, v1
 ; GFX10-NEXT:    v_min_u32_e32 v1, 32, v1
-; GFX10-NEXT:    v_add_nc_u32_e32 v1, -16, v1
-; GFX10-NEXT:    v_add_nc_u16 v1, v1, -8
+; GFX10-NEXT:    v_subrev_nc_u32_e32 v1, 24, v1
 ; GFX10-NEXT:    global_store_byte v0, v1, s[0:1]
 ; GFX10-NEXT:    s_endpgm
 ;
@@ -586,9 +584,7 @@ define amdgpu_kernel void @v_ctlz_i8(ptr addrspace(1) noalias %out, ptr addrspac
 ; GFX11-NEXT:    v_clz_i32_u32_e32 v1, v1
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
 ; GFX11-NEXT:    v_min_u32_e32 v1, 32, v1
-; GFX11-NEXT:    v_add_nc_u32_e32 v1, -16, v1
-; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1)
-; GFX11-NEXT:    v_add_nc_u16 v1, v1, -8
+; GFX11-NEXT:    v_subrev_nc_u32_e32 v1, 24, v1
 ; GFX11-NEXT:    global_store_b8 v0, v1, s[0:1]
 ; GFX11-NEXT:    s_nop 0
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
diff --git a/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll b/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll
index 03f3d04cf8a68f..1c57a64643b92b 100644
--- a/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll
+++ b/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll
@@ -329,10 +329,10 @@ define amdgpu_kernel void @s_ctlz_zero_undef_i8_with_select(ptr addrspace(1) noa
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
 ; VI-NEXT:    s_and_b32 s2, s2, 0xff
 ; VI-NEXT:    s_flbit_i32_b32 s2, s2
-; VI-NEXT:    s_add_i32 s2, s2, -16
+; VI-NEXT:    s_sub_i32 s2, s2, 24
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
-; VI-NEXT:    v_add_u16_e64 v2, s2, -8
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
+; VI-NEXT:    v_mov_b32_e32 v2, s2
 ; VI-NEXT:    flat_store_byte v[0:1], v2
 ; VI-NEXT:    s_endpgm
 ;
@@ -606,8 +606,7 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i8_with_select(ptr addrspace(1) noa
 ; VI-NEXT:    flat_load_ubyte v0, v[0:1]
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_ffbh_u32_sdwa v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
-; VI-NEXT:    v_add_u32_e32 v1, vcc, -16, v1
-; VI-NEXT:    v_add_u16_e32 v1, -8, v1
+; VI-NEXT:    v_subrev_u32_e32 v1, vcc, 24, v1
 ; VI-NEXT:    v_cmp_ne_u16_e32 vcc, 0, v0
 ; VI-NEXT:    v_cndmask_b32_e32 v2, 32, v1, vcc
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
@@ -1118,8 +1117,7 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i8(ptr addrspace(1) noalias %out, p
 ; VI-NEXT:    flat_load_ubyte v0, v[0:1]
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_ffbh_u32_e32 v0, v0
-; VI-NEXT:    v_add_u32_e32 v0, vcc, -16, v0
-; VI-NEXT:    v_add_u16_e32 v2, -8, v0
+; VI-NEXT:    v_subrev_u32_e32 v2, vcc, 24, v0
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    flat_store_byte v[0:1], v2

>From a5e2ed1c00304dc65b91c1ae48fbb364400e0146 Mon Sep 17 00:00:00 2001
From: Leon Clark <leoclark at amd.com>
Date: Fri, 12 Jan 2024 18:23:50 +0000
Subject: [PATCH 2/3] Address review comments.

---
 llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp | 20 +++++++++++--------
 llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h   |  3 +--
 2 files changed, 13 insertions(+), 10 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index 2c36e6e01e5539..fba60d024c1342 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -446,7 +446,9 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
       {ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF, ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF},
       MVT::i64, Custom);
 
-  setOperationAction({ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF}, MVT::i8, Custom);
+  for (auto VT : {MVT::i8, MVT::i16}) {
+    setOperationAction({ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF}, VT, Custom);
+  }
 
   static const MVT::SimpleValueType VectorIntTypes[] = {
       MVT::v2i32, MVT::v3i32, MVT::v4i32, MVT::v5i32, MVT::v6i32, MVT::v7i32,
@@ -1401,7 +1403,8 @@ void AMDGPUTargetLowering::ReplaceNodeResults(SDNode *N,
     return;
   case ISD::CTLZ:
   case ISD::CTLZ_ZERO_UNDEF:
-    replaceCTLZResults(SDValue(N, 0u), DAG, Results);
+    if (auto Lowered = lowerCTLZResults(SDValue(N, 0u), DAG))
+      Results.push_back(Lowered);
     return;
   default:
     return;
@@ -3068,23 +3071,24 @@ static bool isCttzOpc(unsigned Opc) {
   return Opc == ISD::CTTZ || Opc == ISD::CTTZ_ZERO_UNDEF;
 }
 
-void AMDGPUTargetLowering::replaceCTLZResults(
-    SDValue Op, SelectionDAG &DAG, SmallVectorImpl<SDValue> &Results) const {
+SDValue AMDGPUTargetLowering::lowerCTLZResults(SDValue Op,
+                                               SelectionDAG &DAG) const {
   auto SL = SDLoc(Op);
   auto Arg = Op.getOperand(0u);
   auto ResultVT = Op.getValueType();
 
-  if (ResultVT != MVT::i8)
-    return;
+  if (!(ResultVT == MVT::i8 || ResultVT == MVT::i16))
+    return {};
 
   assert(isCtlzOpc(Op.getOpcode()));
   assert(ResultVT == Arg.getValueType());
 
-  auto SubVal = DAG.getConstant(24u, SL, MVT::i32);
+  auto const LeadingZeroes = 32u - ResultVT.getFixedSizeInBits();
+  auto SubVal = DAG.getConstant(LeadingZeroes, SL, MVT::i32);
   auto NewOp = DAG.getNode(ISD::ZERO_EXTEND, SL, MVT::i32, Arg);
   NewOp = DAG.getNode(Op.getOpcode(), SL, MVT::i32, NewOp);
   NewOp = DAG.getNode(ISD::SUB, SL, MVT::i32, NewOp, SubVal);
-  Results.push_back(DAG.getNode(ISD::TRUNCATE, SL, ResultVT, NewOp));
+  return DAG.getNode(ISD::TRUNCATE, SL, ResultVT, NewOp);
 }
 
 SDValue AMDGPUTargetLowering::LowerCTLZ_CTTZ(SDValue Op, SelectionDAG &DAG) const {
diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
index 4cbd6c3146d5dc..bace6e7cb3f497 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.h
@@ -84,8 +84,7 @@ class AMDGPUTargetLowering : public TargetLowering {
                             SDNodeFlags Flags) const;
   SDValue lowerFEXP(SDValue Op, SelectionDAG &DAG) const;
 
-  void replaceCTLZResults(SDValue Op, SelectionDAG &DAG,
-                          SmallVectorImpl<SDValue> &Results) const;
+  SDValue lowerCTLZResults(SDValue Op, SelectionDAG &DAG) const;
 
   SDValue LowerCTLZ_CTTZ(SDValue Op, SelectionDAG &DAG) const;
 

>From 990dc5fbc3e142aacd98d99d9afe76e1b0d86233 Mon Sep 17 00:00:00 2001
From: Leon Clark <leoclark at amd.com>
Date: Mon, 15 Jan 2024 21:08:12 +0000
Subject: [PATCH 3/3] Address review comments.

---
 llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp |  7 +-
 llvm/test/CodeGen/AMDGPU/ctlz.ll              | 27 +++----
 llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll   | 77 +++++++++----------
 3 files changed, 55 insertions(+), 56 deletions(-)

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
index fba60d024c1342..7bb3dbfacfcc4a 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUISelLowering.cpp
@@ -446,9 +446,8 @@ AMDGPUTargetLowering::AMDGPUTargetLowering(const TargetMachine &TM,
       {ISD::CTTZ, ISD::CTTZ_ZERO_UNDEF, ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF},
       MVT::i64, Custom);
 
-  for (auto VT : {MVT::i8, MVT::i16}) {
+  for (auto VT : {MVT::i8, MVT::i16})
     setOperationAction({ISD::CTLZ, ISD::CTLZ_ZERO_UNDEF}, VT, Custom);
-  }
 
   static const MVT::SimpleValueType VectorIntTypes[] = {
       MVT::v2i32, MVT::v3i32, MVT::v4i32, MVT::v5i32, MVT::v6i32, MVT::v7i32,
@@ -3084,10 +3083,10 @@ SDValue AMDGPUTargetLowering::lowerCTLZResults(SDValue Op,
   assert(ResultVT == Arg.getValueType());
 
   auto const LeadingZeroes = 32u - ResultVT.getFixedSizeInBits();
-  auto SubVal = DAG.getConstant(LeadingZeroes, SL, MVT::i32);
   auto NewOp = DAG.getNode(ISD::ZERO_EXTEND, SL, MVT::i32, Arg);
+  auto ShiftVal = DAG.getConstant(LeadingZeroes, SL, MVT::i32);
+  NewOp = DAG.getNode(ISD::SHL, SL, MVT::i32, NewOp, ShiftVal);
   NewOp = DAG.getNode(Op.getOpcode(), SL, MVT::i32, NewOp);
-  NewOp = DAG.getNode(ISD::SUB, SL, MVT::i32, NewOp, SubVal);
   return DAG.getNode(ISD::TRUNCATE, SL, ResultVT, NewOp);
 }
 
diff --git a/llvm/test/CodeGen/AMDGPU/ctlz.ll b/llvm/test/CodeGen/AMDGPU/ctlz.ll
index 2267aa6b5496c6..387edaf0ec5793 100644
--- a/llvm/test/CodeGen/AMDGPU/ctlz.ll
+++ b/llvm/test/CodeGen/AMDGPU/ctlz.ll
@@ -492,9 +492,9 @@ define amdgpu_kernel void @v_ctlz_i8(ptr addrspace(1) noalias %out, ptr addrspac
 ; SI-NEXT:    s_mov_b32 s4, s0
 ; SI-NEXT:    s_mov_b32 s5, s1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
 ; SI-NEXT:    v_ffbh_u32_e32 v0, v0
 ; SI-NEXT:    v_min_u32_e32 v0, 32, v0
-; SI-NEXT:    v_subrev_i32_e32 v0, vcc, 24, v0
 ; SI-NEXT:    buffer_store_byte v0, off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -512,9 +512,9 @@ define amdgpu_kernel void @v_ctlz_i8(ptr addrspace(1) noalias %out, ptr addrspac
 ; VI-NEXT:    s_mov_b32 s4, s0
 ; VI-NEXT:    s_mov_b32 s5, s1
 ; VI-NEXT:    s_waitcnt vmcnt(0)
+; VI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
 ; VI-NEXT:    v_ffbh_u32_e32 v0, v0
 ; VI-NEXT:    v_min_u32_e32 v0, 32, v0
-; VI-NEXT:    v_subrev_u32_e32 v0, vcc, 24, v0
 ; VI-NEXT:    buffer_store_byte v0, off, s[4:7], 0
 ; VI-NEXT:    s_endpgm
 ;
@@ -522,7 +522,7 @@ define amdgpu_kernel void @v_ctlz_i8(ptr addrspace(1) noalias %out, ptr addrspac
 ; EG:       ; %bb.0:
 ; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    TEX 0 @6
-; EG-NEXT:    ALU 15, @9, KC0[CB0:0-32], KC1[]
+; EG-NEXT:    ALU 16, @9, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    MEM_RAT MSKOR T0.XW, T1.X
 ; EG-NEXT:    CF_END
 ; EG-NEXT:    PAD
@@ -531,14 +531,15 @@ define amdgpu_kernel void @v_ctlz_i8(ptr addrspace(1) noalias %out, ptr addrspac
 ; EG-NEXT:    ALU clause starting at 8:
 ; EG-NEXT:     MOV * T0.X, KC0[2].Z,
 ; EG-NEXT:    ALU clause starting at 9:
-; EG-NEXT:     FFBH_UINT * T0.W, T0.X,
-; EG-NEXT:     CNDE_INT T0.W, T0.X, literal.x, PV.W,
-; EG-NEXT:     AND_INT * T1.W, KC0[2].Y, literal.y,
-; EG-NEXT:    32(4.484155e-44), 3(4.203895e-45)
-; EG-NEXT:     ADD_INT * T0.W, PV.W, literal.x,
-; EG-NEXT:    -24(nan), 0(0.000000e+00)
+; EG-NEXT:     LSHL * T0.W, T0.X, literal.x,
+; EG-NEXT:    24(3.363116e-44), 0(0.000000e+00)
+; EG-NEXT:     FFBH_UINT T1.W, PV.W,
+; EG-NEXT:     AND_INT * T2.W, KC0[2].Y, literal.x,
+; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)
+; EG-NEXT:     CNDE_INT * T0.W, T0.W, literal.x, PV.W,
+; EG-NEXT:    32(4.484155e-44), 0(0.000000e+00)
 ; EG-NEXT:     AND_INT T0.W, PV.W, literal.x,
-; EG-NEXT:     LSHL * T1.W, T1.W, literal.y,
+; EG-NEXT:     LSHL * T1.W, T2.W, literal.y,
 ; EG-NEXT:    255(3.573311e-43), 3(4.203895e-45)
 ; EG-NEXT:     LSHL T0.X, PV.W, PS,
 ; EG-NEXT:     LSHL * T0.W, literal.x, PS,
@@ -555,9 +556,9 @@ define amdgpu_kernel void @v_ctlz_i8(ptr addrspace(1) noalias %out, ptr addrspac
 ; GFX10-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX10-NEXT:    global_load_ubyte v1, v0, s[2:3]
 ; GFX10-NEXT:    s_waitcnt vmcnt(0)
+; GFX10-NEXT:    v_lshlrev_b32_e32 v1, 24, v1
 ; GFX10-NEXT:    v_ffbh_u32_e32 v1, v1
 ; GFX10-NEXT:    v_min_u32_e32 v1, 32, v1
-; GFX10-NEXT:    v_subrev_nc_u32_e32 v1, 24, v1
 ; GFX10-NEXT:    global_store_byte v0, v1, s[0:1]
 ; GFX10-NEXT:    s_endpgm
 ;
@@ -581,10 +582,10 @@ define amdgpu_kernel void @v_ctlz_i8(ptr addrspace(1) noalias %out, ptr addrspac
 ; GFX11-NEXT:    s_waitcnt lgkmcnt(0)
 ; GFX11-NEXT:    global_load_u8 v1, v0, s[2:3]
 ; GFX11-NEXT:    s_waitcnt vmcnt(0)
-; GFX11-NEXT:    v_clz_i32_u32_e32 v1, v1
+; GFX11-NEXT:    v_lshlrev_b32_e32 v1, 24, v1
 ; GFX11-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX11-NEXT:    v_clz_i32_u32_e32 v1, v1
 ; GFX11-NEXT:    v_min_u32_e32 v1, 32, v1
-; GFX11-NEXT:    v_subrev_nc_u32_e32 v1, 24, v1
 ; GFX11-NEXT:    global_store_b8 v0, v1, s[0:1]
 ; GFX11-NEXT:    s_nop 0
 ; GFX11-NEXT:    s_sendmsg sendmsg(MSG_DEALLOC_VGPRS)
diff --git a/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll b/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll
index 1c57a64643b92b..093027d4c08c7b 100644
--- a/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll
+++ b/llvm/test/CodeGen/AMDGPU/ctlz_zero_undef.ll
@@ -314,9 +314,8 @@ define amdgpu_kernel void @s_ctlz_zero_undef_i8_with_select(ptr addrspace(1) noa
 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_and_b32 s2, s2, 0xff
-; SI-NEXT:    s_flbit_i32_b32 s2, s2
-; SI-NEXT:    s_sub_i32 s4, s2, 24
+; SI-NEXT:    s_lshl_b32 s2, s2, 24
+; SI-NEXT:    s_flbit_i32_b32 s4, s2
 ; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
 ; SI-NEXT:    buffer_store_byte v0, off, s[0:3], 0
@@ -327,9 +326,8 @@ define amdgpu_kernel void @s_ctlz_zero_undef_i8_with_select(ptr addrspace(1) noa
 ; VI-NEXT:    s_load_dword s2, s[0:1], 0x2c
 ; VI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x24
 ; VI-NEXT:    s_waitcnt lgkmcnt(0)
-; VI-NEXT:    s_and_b32 s2, s2, 0xff
+; VI-NEXT:    s_lshl_b32 s2, s2, 24
 ; VI-NEXT:    s_flbit_i32_b32 s2, s2
-; VI-NEXT:    s_sub_i32 s2, s2, 24
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    v_mov_b32_e32 v2, s2
@@ -349,13 +347,13 @@ define amdgpu_kernel void @s_ctlz_zero_undef_i8_with_select(ptr addrspace(1) noa
 ; EG-NEXT:    ALU clause starting at 8:
 ; EG-NEXT:     MOV * T0.X, 0.0,
 ; EG-NEXT:    ALU clause starting at 9:
-; EG-NEXT:     FFBH_UINT T0.W, T0.X,
+; EG-NEXT:     LSHL * T0.W, T0.X, literal.x,
+; EG-NEXT:    24(3.363116e-44), 0(0.000000e+00)
+; EG-NEXT:     FFBH_UINT T0.W, PV.W,
 ; EG-NEXT:     AND_INT * T1.W, KC0[2].Y, literal.x,
 ; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)
-; EG-NEXT:     ADD_INT * T0.W, PV.W, literal.x,
-; EG-NEXT:    -24(nan), 0(0.000000e+00)
 ; EG-NEXT:     AND_INT T0.W, PV.W, literal.x,
-; EG-NEXT:     LSHL * T1.W, T1.W, literal.y,
+; EG-NEXT:     LSHL * T1.W, PS, literal.y,
 ; EG-NEXT:    255(3.573311e-43), 3(4.203895e-45)
 ; EG-NEXT:     LSHL T0.X, PV.W, PS,
 ; EG-NEXT:     LSHL * T0.W, literal.x, PS,
@@ -391,9 +389,8 @@ define amdgpu_kernel void @s_ctlz_zero_undef_i16_with_select(ptr addrspace(1) no
 ; SI-NEXT:    s_load_dwordx2 s[0:1], s[0:1], 0x9
 ; SI-NEXT:    s_mov_b32 s3, 0xf000
 ; SI-NEXT:    s_waitcnt lgkmcnt(0)
-; SI-NEXT:    s_and_b32 s2, s2, 0xffff
-; SI-NEXT:    s_flbit_i32_b32 s2, s2
-; SI-NEXT:    s_add_i32 s4, s2, -16
+; SI-NEXT:    s_lshl_b32 s2, s2, 16
+; SI-NEXT:    s_flbit_i32_b32 s4, s2
 ; SI-NEXT:    s_mov_b32 s2, -1
 ; SI-NEXT:    v_mov_b32_e32 v0, s4
 ; SI-NEXT:    buffer_store_short v0, off, s[0:3], 0
@@ -426,13 +423,13 @@ define amdgpu_kernel void @s_ctlz_zero_undef_i16_with_select(ptr addrspace(1) no
 ; EG-NEXT:    ALU clause starting at 8:
 ; EG-NEXT:     MOV * T0.X, 0.0,
 ; EG-NEXT:    ALU clause starting at 9:
-; EG-NEXT:     FFBH_UINT T0.W, T0.X,
+; EG-NEXT:     LSHL * T0.W, T0.X, literal.x,
+; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)
+; EG-NEXT:     FFBH_UINT T0.W, PV.W,
 ; EG-NEXT:     AND_INT * T1.W, KC0[2].Y, literal.x,
 ; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)
-; EG-NEXT:     ADD_INT * T0.W, PV.W, literal.x,
-; EG-NEXT:    -16(nan), 0(0.000000e+00)
 ; EG-NEXT:     AND_INT T0.W, PV.W, literal.x,
-; EG-NEXT:     LSHL * T1.W, T1.W, literal.y,
+; EG-NEXT:     LSHL * T1.W, PS, literal.y,
 ; EG-NEXT:    65535(9.183409e-41), 3(4.203895e-45)
 ; EG-NEXT:     LSHL T0.X, PV.W, PS,
 ; EG-NEXT:     LSHL * T0.W, literal.x, PS,
@@ -590,8 +587,8 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i8_with_select(ptr addrspace(1) noa
 ; SI-NEXT:    s_mov_b32 s4, s0
 ; SI-NEXT:    s_mov_b32 s5, s1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
-; SI-NEXT:    v_ffbh_u32_e32 v1, v0
-; SI-NEXT:    v_subrev_i32_e32 v1, vcc, 24, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 24, v0
+; SI-NEXT:    v_ffbh_u32_e32 v1, v1
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
 ; SI-NEXT:    v_cndmask_b32_e32 v0, 32, v1, vcc
 ; SI-NEXT:    buffer_store_byte v0, off, s[4:7], 0
@@ -605,8 +602,8 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i8_with_select(ptr addrspace(1) noa
 ; VI-NEXT:    v_mov_b32_e32 v1, s3
 ; VI-NEXT:    flat_load_ubyte v0, v[0:1]
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_ffbh_u32_sdwa v1, v0 dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
-; VI-NEXT:    v_subrev_u32_e32 v1, vcc, 24, v1
+; VI-NEXT:    v_lshlrev_b32_e32 v1, 24, v0
+; VI-NEXT:    v_ffbh_u32_e32 v1, v1
 ; VI-NEXT:    v_cmp_ne_u16_e32 vcc, 0, v0
 ; VI-NEXT:    v_cndmask_b32_e32 v2, 32, v1, vcc
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
@@ -618,7 +615,7 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i8_with_select(ptr addrspace(1) noa
 ; EG:       ; %bb.0:
 ; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    TEX 0 @6
-; EG-NEXT:    ALU 15, @9, KC0[CB0:0-32], KC1[]
+; EG-NEXT:    ALU 16, @9, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    MEM_RAT MSKOR T0.XW, T1.X
 ; EG-NEXT:    CF_END
 ; EG-NEXT:    PAD
@@ -627,10 +624,11 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i8_with_select(ptr addrspace(1) noa
 ; EG-NEXT:    ALU clause starting at 8:
 ; EG-NEXT:     MOV * T0.X, KC0[2].Z,
 ; EG-NEXT:    ALU clause starting at 9:
-; EG-NEXT:     FFBH_UINT * T0.W, T0.X,
-; EG-NEXT:     ADD_INT T0.W, PV.W, literal.x,
-; EG-NEXT:     AND_INT * T1.W, KC0[2].Y, literal.y,
-; EG-NEXT:    -24(nan), 3(4.203895e-45)
+; EG-NEXT:     LSHL * T0.W, T0.X, literal.x,
+; EG-NEXT:    24(3.363116e-44), 0(0.000000e+00)
+; EG-NEXT:     FFBH_UINT T0.W, PV.W,
+; EG-NEXT:     AND_INT * T1.W, KC0[2].Y, literal.x,
+; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)
 ; EG-NEXT:     CNDE_INT * T0.W, T0.X, literal.x, PV.W,
 ; EG-NEXT:    32(4.484155e-44), 0(0.000000e+00)
 ; EG-NEXT:     AND_INT T0.W, PV.W, literal.x,
@@ -685,8 +683,8 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i16_with_select(ptr addrspace(1) no
 ; SI-NEXT:    v_lshlrev_b32_e32 v0, 8, v0
 ; SI-NEXT:    s_waitcnt vmcnt(0)
 ; SI-NEXT:    v_or_b32_e32 v0, v0, v1
-; SI-NEXT:    v_ffbh_u32_e32 v1, v0
-; SI-NEXT:    v_add_i32_e32 v1, vcc, -16, v1
+; SI-NEXT:    v_lshlrev_b32_e32 v1, 16, v0
+; SI-NEXT:    v_ffbh_u32_e32 v1, v1
 ; SI-NEXT:    v_cmp_ne_u32_e32 vcc, 0, v0
 ; SI-NEXT:    v_cndmask_b32_e32 v0, 32, v1, vcc
 ; SI-NEXT:    buffer_store_short v0, off, s[4:7], 0
@@ -721,7 +719,7 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i16_with_select(ptr addrspace(1) no
 ; EG:       ; %bb.0:
 ; EG-NEXT:    ALU 0, @8, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    TEX 0 @6
-; EG-NEXT:    ALU 15, @9, KC0[CB0:0-32], KC1[]
+; EG-NEXT:    ALU 16, @9, KC0[CB0:0-32], KC1[]
 ; EG-NEXT:    MEM_RAT MSKOR T0.XW, T1.X
 ; EG-NEXT:    CF_END
 ; EG-NEXT:    PAD
@@ -730,10 +728,11 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i16_with_select(ptr addrspace(1) no
 ; EG-NEXT:    ALU clause starting at 8:
 ; EG-NEXT:     MOV * T0.X, KC0[2].Z,
 ; EG-NEXT:    ALU clause starting at 9:
-; EG-NEXT:     FFBH_UINT * T0.W, T0.X,
-; EG-NEXT:     ADD_INT T0.W, PV.W, literal.x,
-; EG-NEXT:     AND_INT * T1.W, KC0[2].Y, literal.y,
-; EG-NEXT:    -16(nan), 3(4.203895e-45)
+; EG-NEXT:     LSHL * T0.W, T0.X, literal.x,
+; EG-NEXT:    16(2.242078e-44), 0(0.000000e+00)
+; EG-NEXT:     FFBH_UINT T0.W, PV.W,
+; EG-NEXT:     AND_INT * T1.W, KC0[2].Y, literal.x,
+; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)
 ; EG-NEXT:     CNDE_INT * T0.W, T0.X, literal.x, PV.W,
 ; EG-NEXT:    32(4.484155e-44), 0(0.000000e+00)
 ; EG-NEXT:     AND_INT T0.W, PV.W, literal.x,
@@ -1102,8 +1101,8 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i8(ptr addrspace(1) noalias %out, p
 ; SI-NEXT:    s_mov_b32 s4, s0
 ; SI-NEXT:    s_mov_b32 s5, s1
 ; SI-NEXT:    s_waitcnt vmcnt(0)
+; SI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
 ; SI-NEXT:    v_ffbh_u32_e32 v0, v0
-; SI-NEXT:    v_subrev_i32_e32 v0, vcc, 24, v0
 ; SI-NEXT:    buffer_store_byte v0, off, s[4:7], 0
 ; SI-NEXT:    s_endpgm
 ;
@@ -1116,8 +1115,8 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i8(ptr addrspace(1) noalias %out, p
 ; VI-NEXT:    v_addc_u32_e32 v1, vcc, 0, v1, vcc
 ; VI-NEXT:    flat_load_ubyte v0, v[0:1]
 ; VI-NEXT:    s_waitcnt vmcnt(0)
-; VI-NEXT:    v_ffbh_u32_e32 v0, v0
-; VI-NEXT:    v_subrev_u32_e32 v2, vcc, 24, v0
+; VI-NEXT:    v_lshlrev_b32_e32 v0, 24, v0
+; VI-NEXT:    v_ffbh_u32_e32 v2, v0
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
 ; VI-NEXT:    v_mov_b32_e32 v1, s1
 ; VI-NEXT:    flat_store_byte v[0:1], v2
@@ -1136,13 +1135,13 @@ define amdgpu_kernel void @v_ctlz_zero_undef_i8(ptr addrspace(1) noalias %out, p
 ; EG-NEXT:    ALU clause starting at 8:
 ; EG-NEXT:     ADD_INT * T0.X, KC0[2].Z, T0.X,
 ; EG-NEXT:    ALU clause starting at 9:
-; EG-NEXT:     FFBH_UINT T0.W, T0.X,
+; EG-NEXT:     LSHL * T0.W, T0.X, literal.x,
+; EG-NEXT:    24(3.363116e-44), 0(0.000000e+00)
+; EG-NEXT:     FFBH_UINT T0.W, PV.W,
 ; EG-NEXT:     AND_INT * T1.W, KC0[2].Y, literal.x,
 ; EG-NEXT:    3(4.203895e-45), 0(0.000000e+00)
-; EG-NEXT:     ADD_INT * T0.W, PV.W, literal.x,
-; EG-NEXT:    -24(nan), 0(0.000000e+00)
 ; EG-NEXT:     AND_INT T0.W, PV.W, literal.x,
-; EG-NEXT:     LSHL * T1.W, T1.W, literal.y,
+; EG-NEXT:     LSHL * T1.W, PS, literal.y,
 ; EG-NEXT:    255(3.573311e-43), 3(4.203895e-45)
 ; EG-NEXT:     LSHL T0.X, PV.W, PS,
 ; EG-NEXT:     LSHL * T0.W, literal.x, PS,



More information about the llvm-commits mailing list