Skip to content

[SelectionDAG] Widen cttz to cttz_zero_undef #92514

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
merged 1 commit into from
May 17, 2024

Conversation

jayfoad
Copy link
Contributor

@jayfoad jayfoad commented May 17, 2024

Instead of widening e.g. i8 cttz(x) to i16 cttz(x | 0x100), use the more
optimizable form cttz_zero_undef(x | 0x100) since the widened operand is
definitely not zero.

Instead of widening e.g. i8 cttz(x) to i16 cttz(x | 0x100), use the more
optimizable form cttz_zero_undef(x | 0x100) since the widened operand is
definitely not zero.
@llvmbot llvmbot added backend:AMDGPU llvm:SelectionDAG SelectionDAGISel as well labels May 17, 2024
@llvmbot
Copy link
Member

llvmbot commented May 17, 2024

@llvm/pr-subscribers-llvm-selectiondag

@llvm/pr-subscribers-backend-amdgpu

Author: Jay Foad (jayfoad)

Changes

Instead of widening e.g. i8 cttz(x) to i16 cttz(x | 0x100), use the more
optimizable form cttz_zero_undef(x | 0x100) since the widened operand is
definitely not zero.


Full diff: https://github.com/llvm/llvm-project/pull/92514.diff

5 Files Affected:

  • (modified) llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp (+7-5)
  • (modified) llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp (+9-6)
  • (modified) llvm/test/CodeGen/AMDGPU/cttz.ll (-2)
  • (modified) llvm/test/CodeGen/AMDGPU/cttz_zero_undef.ll (-1)
  • (modified) llvm/test/CodeGen/RISCV/rvv/cttz-vp.ll (+8-22)
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
index fd97a1283b65a..0543c211c4971 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeDAG.cpp
@@ -5032,7 +5032,7 @@ void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
   case ISD::CTTZ_ZERO_UNDEF:
   case ISD::CTLZ:
   case ISD::CTLZ_ZERO_UNDEF:
-  case ISD::CTPOP:
+  case ISD::CTPOP: {
     // Zero extend the argument unless its cttz, then use any_extend.
     if (Node->getOpcode() == ISD::CTTZ ||
         Node->getOpcode() == ISD::CTTZ_ZERO_UNDEF)
@@ -5040,7 +5040,8 @@ void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
     else
       Tmp1 = DAG.getNode(ISD::ZERO_EXTEND, dl, NVT, Node->getOperand(0));
 
-    if (Node->getOpcode() == ISD::CTTZ) {
+    unsigned NewOpc = Node->getOpcode();
+    if (NewOpc == ISD::CTTZ) {
       // The count is the same in the promoted type except if the original
       // value was zero.  This can be handled by setting the bit just off
       // the top of the original type.
@@ -5048,12 +5049,12 @@ void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
                                         OVT.getSizeInBits());
       Tmp1 = DAG.getNode(ISD::OR, dl, NVT, Tmp1,
                          DAG.getConstant(TopBit, dl, NVT));
+      NewOpc = ISD::CTTZ_ZERO_UNDEF;
     }
     // Perform the larger operation. For CTPOP and CTTZ_ZERO_UNDEF, this is
     // already the correct result.
-    Tmp1 = DAG.getNode(Node->getOpcode(), dl, NVT, Tmp1);
-    if (Node->getOpcode() == ISD::CTLZ ||
-        Node->getOpcode() == ISD::CTLZ_ZERO_UNDEF) {
+    Tmp1 = DAG.getNode(NewOpc, dl, NVT, Tmp1);
+    if (NewOpc == ISD::CTLZ || NewOpc == ISD::CTLZ_ZERO_UNDEF) {
       // Tmp1 = Tmp1 - (sizeinbits(NVT) - sizeinbits(Old VT))
       Tmp1 = DAG.getNode(ISD::SUB, dl, NVT, Tmp1,
                           DAG.getConstant(NVT.getSizeInBits() -
@@ -5061,6 +5062,7 @@ void SelectionDAGLegalize::PromoteNode(SDNode *Node) {
     }
     Results.push_back(DAG.getNode(ISD::TRUNCATE, dl, OVT, Tmp1));
     break;
+  }
   case ISD::BITREVERSE:
   case ISD::BSWAP: {
     unsigned DiffBits = NVT.getSizeInBits() - OVT.getSizeInBits();
diff --git a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
index 0aa36deda79dc..98f64947bcabc 100644
--- a/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/LegalizeIntegerTypes.cpp
@@ -709,23 +709,26 @@ SDValue DAGTypeLegalizer::PromoteIntRes_CTTZ(SDNode *N) {
     }
   }
 
-  if (N->getOpcode() == ISD::CTTZ || N->getOpcode() == ISD::VP_CTTZ) {
+  unsigned NewOpc = N->getOpcode();
+  if (NewOpc == ISD::CTTZ || NewOpc == ISD::VP_CTTZ) {
     // The count is the same in the promoted type except if the original
     // value was zero.  This can be handled by setting the bit just off
     // the top of the original type.
     auto TopBit = APInt::getOneBitSet(NVT.getScalarSizeInBits(),
                                       OVT.getScalarSizeInBits());
-    if (N->getOpcode() == ISD::CTTZ)
+    if (NewOpc == ISD::CTTZ) {
       Op = DAG.getNode(ISD::OR, dl, NVT, Op, DAG.getConstant(TopBit, dl, NVT));
-    else
+      NewOpc = ISD::CTTZ_ZERO_UNDEF;
+    } else {
       Op =
           DAG.getNode(ISD::VP_OR, dl, NVT, Op, DAG.getConstant(TopBit, dl, NVT),
                       N->getOperand(1), N->getOperand(2));
+      NewOpc = ISD::VP_CTTZ_ZERO_UNDEF;
+    }
   }
   if (!N->isVPOpcode())
-    return DAG.getNode(N->getOpcode(), dl, NVT, Op);
-  return DAG.getNode(N->getOpcode(), dl, NVT, Op, N->getOperand(1),
-                     N->getOperand(2));
+    return DAG.getNode(NewOpc, dl, NVT, Op);
+  return DAG.getNode(NewOpc, dl, NVT, Op, N->getOperand(1), N->getOperand(2));
 }
 
 SDValue DAGTypeLegalizer::PromoteIntRes_VP_CttzElements(SDNode *N) {
diff --git a/llvm/test/CodeGen/AMDGPU/cttz.ll b/llvm/test/CodeGen/AMDGPU/cttz.ll
index 118d6c123046b..ee2894a66fbfc 100644
--- a/llvm/test/CodeGen/AMDGPU/cttz.ll
+++ b/llvm/test/CodeGen/AMDGPU/cttz.ll
@@ -1408,7 +1408,6 @@ define amdgpu_kernel void @v_cttz_i32_sel_ne_bitwidth(ptr addrspace(1) noalias %
 ; VI-NEXT:    s_waitcnt vmcnt(0)
 ; VI-NEXT:    v_or_b32_e32 v2, 0x10000, v0
 ; VI-NEXT:    v_ffbl_b32_e32 v2, v2
-; VI-NEXT:    v_min_u32_e32 v2, 32, v2
 ; VI-NEXT:    v_cmp_ne_u16_e32 vcc, 0, v0
 ; VI-NEXT:    v_cndmask_b32_e32 v0, v1, v2, vcc
 ; VI-NEXT:    buffer_store_short v0, off, s[4:7], 0
@@ -1451,7 +1450,6 @@ define amdgpu_kernel void @v_cttz_i32_sel_ne_bitwidth(ptr addrspace(1) noalias %
 ; GFX10-NEXT:    v_or_b32_e32 v2, 0x10000, v1
 ; GFX10-NEXT:    v_cmp_ne_u16_e32 vcc_lo, 0, v1
 ; GFX10-NEXT:    v_ffbl_b32_e32 v2, v2
-; GFX10-NEXT:    v_min_u32_e32 v2, 32, v2
 ; GFX10-NEXT:    v_cndmask_b32_e32 v1, 0xffff, v2, vcc_lo
 ; GFX10-NEXT:    global_store_short v0, v1, s[0:1]
 ; GFX10-NEXT:    s_endpgm
diff --git a/llvm/test/CodeGen/AMDGPU/cttz_zero_undef.ll b/llvm/test/CodeGen/AMDGPU/cttz_zero_undef.ll
index 71f1cd54d705c..392a44318b0a5 100644
--- a/llvm/test/CodeGen/AMDGPU/cttz_zero_undef.ll
+++ b/llvm/test/CodeGen/AMDGPU/cttz_zero_undef.ll
@@ -1561,7 +1561,6 @@ define amdgpu_kernel void @v_cttz_i32_sel_ne_bitwidth(ptr addrspace(1) noalias %
 ; VI-NEXT:    v_or_b32_e32 v0, v2, v0
 ; VI-NEXT:    v_or_b32_e32 v2, 0x10000, v0
 ; VI-NEXT:    v_ffbl_b32_e32 v2, v2
-; VI-NEXT:    v_min_u32_e32 v2, 32, v2
 ; VI-NEXT:    v_cmp_ne_u16_e32 vcc, 0, v0
 ; VI-NEXT:    v_cndmask_b32_e32 v2, v1, v2, vcc
 ; VI-NEXT:    v_mov_b32_e32 v0, s0
diff --git a/llvm/test/CodeGen/RISCV/rvv/cttz-vp.ll b/llvm/test/CodeGen/RISCV/rvv/cttz-vp.ll
index ef8a6c704a44b..4a001662ce2ca 100644
--- a/llvm/test/CodeGen/RISCV/rvv/cttz-vp.ll
+++ b/llvm/test/CodeGen/RISCV/rvv/cttz-vp.ll
@@ -4145,29 +4145,15 @@ define <vscale x 1 x i9> @vp_cttz_nxv1i9(<vscale x 1 x i9> %va, <vscale x 1 x i1
 ; CHECK-NEXT:    li a1, 512
 ; CHECK-NEXT:    vsetvli zero, a0, e16, mf4, ta, ma
 ; CHECK-NEXT:    vor.vx v8, v8, a1, v0.t
-; CHECK-NEXT:    li a0, 1
-; CHECK-NEXT:    vsub.vx v9, v8, a0, v0.t
-; CHECK-NEXT:    vnot.v v8, v8, v0.t
+; CHECK-NEXT:    vrsub.vi v9, v8, 0, v0.t
 ; CHECK-NEXT:    vand.vv v8, v8, v9, v0.t
-; CHECK-NEXT:    vsrl.vi v9, v8, 1, v0.t
-; CHECK-NEXT:    lui a0, 5
-; CHECK-NEXT:    addi a0, a0, 1365
-; CHECK-NEXT:    vand.vx v9, v9, a0, v0.t
-; CHECK-NEXT:    vsub.vv v8, v8, v9, v0.t
-; CHECK-NEXT:    lui a0, 3
-; CHECK-NEXT:    addi a0, a0, 819
-; CHECK-NEXT:    vand.vx v9, v8, a0, v0.t
-; CHECK-NEXT:    vsrl.vi v8, v8, 2, v0.t
-; CHECK-NEXT:    vand.vx v8, v8, a0, v0.t
-; CHECK-NEXT:    vadd.vv v8, v9, v8, v0.t
-; CHECK-NEXT:    vsrl.vi v9, v8, 4, v0.t
-; CHECK-NEXT:    vadd.vv v8, v8, v9, v0.t
-; CHECK-NEXT:    lui a0, 1
-; CHECK-NEXT:    addi a0, a0, -241
-; CHECK-NEXT:    vand.vx v8, v8, a0, v0.t
-; CHECK-NEXT:    li a0, 257
-; CHECK-NEXT:    vmul.vx v8, v8, a0, v0.t
-; CHECK-NEXT:    vsrl.vi v8, v8, 8, v0.t
+; CHECK-NEXT:    vfwcvt.f.xu.v v9, v8, v0.t
+; CHECK-NEXT:    vsetvli zero, zero, e32, mf2, ta, ma
+; CHECK-NEXT:    vsrl.vi v8, v9, 23, v0.t
+; CHECK-NEXT:    vsetvli zero, zero, e16, mf4, ta, ma
+; CHECK-NEXT:    vnsrl.wi v8, v8, 0, v0.t
+; CHECK-NEXT:    li a0, 127
+; CHECK-NEXT:    vsub.vx v8, v8, a0, v0.t
 ; CHECK-NEXT:    ret
 ;
 ; CHECK-ZVBB-LABEL: vp_cttz_nxv1i9:

Copy link
Contributor

@arsenm arsenm left a comment

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Curiously GlobalISel seems to already have this:

// Now we know the operand is non-zero, use the more relaxed opcode.

@jayfoad jayfoad merged commit ac09292 into llvm:main May 17, 2024
6 of 7 checks passed
@jayfoad jayfoad deleted the widen-cttz-to-zero-undef branch May 17, 2024 11:39
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
backend:AMDGPU llvm:SelectionDAG SelectionDAGISel as well
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants