[llvm] [RISCV][GISEL] instruction-select vmclr (PR #110782)

Michael Maitland via llvm-commits llvm-commits at lists.llvm.org
Tue Oct 1 21:07:43 PDT 2024


https://github.com/michaelmaitland updated https://github.com/llvm/llvm-project/pull/110782

>From 8de800a2fb129efefd90ee2c993a7933acb6b6ee Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Tue, 1 Oct 2024 19:14:34 -0700
Subject: [PATCH 1/5] [RISCV][GISEL] instruction-select vmclr

---
 .../RISCV/GISel/RISCVInstructionSelector.cpp  |  22 ++++
 llvm/lib/Target/RISCV/RISCVGISel.td           |   5 +
 .../instruction-select/rvv/vmclr-rv32.mir     | 124 ++++++++++++++++++
 .../instruction-select/rvv/vmclr-rv64.mir     | 124 ++++++++++++++++++
 4 files changed, 275 insertions(+)
 create mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/vmclr-rv32.mir
 create mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/vmclr-rv64.mir

diff --git a/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp b/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
index 92d00c26bd219c..cb28f40d64b176 100644
--- a/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
+++ b/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
@@ -100,6 +100,8 @@ class RISCVInstructionSelector : public InstructionSelector {
     return selectSHXADD_UWOp(Root, ShAmt);
   }
 
+  ComplexRendererFns selectVLOp(MachineOperand &Root) const;
+
   // Custom renderers for tablegen
   void renderNegImm(MachineInstrBuilder &MIB, const MachineInstr &MI,
                     int OpIdx) const;
@@ -379,6 +381,26 @@ RISCVInstructionSelector::selectSHXADD_UWOp(MachineOperand &Root,
   return std::nullopt;
 }
 
+InstructionSelector::ComplexRendererFns
+RISCVInstructionSelector::selectVLOp(MachineOperand &Root) const {
+  MachineRegisterInfo &MRI =
+      Root.getParent()->getParent()->getParent()->getRegInfo();
+  assert(Root.isReg() && "Expected operand to be a Register");
+  MachineInstr *RootDef = MRI.getVRegDef(Root.getReg());
+
+  if (RootDef->getOpcode() == TargetOpcode::G_CONSTANT &&
+      RootDef->getOperand(1).getCImm()->getSExtValue() == RISCV::VLMaxSentinel)
+    // If the operand is a G_CONSTANT with value VLMaxSentinel, convert it
+    // to an immediate with value VLMaxSentinel. This is recognized specially by
+    // the vsetvli insertion pass.
+    return {
+        {[=](MachineInstrBuilder &MIB) { MIB.addImm(RISCV::VLMaxSentinel); }}};
+
+  // FIXME: Implement non-VLMAX case. ISEL will fail gracefully by returning
+  // like this for now.
+  return std::nullopt;
+}
+
 InstructionSelector::ComplexRendererFns
 RISCVInstructionSelector::selectAddrRegImm(MachineOperand &Root) const {
   MachineFunction &MF = *Root.getParent()->getParent()->getParent();
diff --git a/llvm/lib/Target/RISCV/RISCVGISel.td b/llvm/lib/Target/RISCV/RISCVGISel.td
index 319611111cf470..a7d46e6edd39f1 100644
--- a/llvm/lib/Target/RISCV/RISCVGISel.td
+++ b/llvm/lib/Target/RISCV/RISCVGISel.td
@@ -50,6 +50,11 @@ def GIAddrRegImm :
   GIComplexOperandMatcher<s32, "selectAddrRegImm">,
   GIComplexPatternEquiv<AddrRegImm>;
 
+def GIVLOpS32 : GIComplexOperandMatcher<s32, "selectVLOp">,
+                GIComplexPatternEquiv<VLOp>;
+def GIVLOpS64 : GIComplexOperandMatcher<s64, "selectVLOp">,
+                GIComplexPatternEquiv<VLOp>;
+
 // Convert from i32 immediate to i64 target immediate to make SelectionDAG type
 // checking happy so we can use ADDIW which expects an XLen immediate.
 def as_i64imm : SDNodeXForm<imm, [{
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/vmclr-rv32.mir b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/vmclr-rv32.mir
new file mode 100644
index 00000000000000..1ef1312cc17c0e
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/vmclr-rv32.mir
@@ -0,0 +1,124 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=riscv32 -mattr=+v,+m -run-pass=instruction-select \
+# RUN:   -simplify-mir -verify-machineinstrs %s -o - | FileCheck %s
+
+---
+name:            splat_zero_nxv1i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv1i1
+    ; CHECK: [[PseudoVMCLR_M_B1_:%[0-9]+]]:vr = PseudoVMCLR_M_B1 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B1_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s32) = G_CONSTANT i32 -1
+    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 1 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv2i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv2i1
+    ; CHECK: [[PseudoVMCLR_M_B2_:%[0-9]+]]:vr = PseudoVMCLR_M_B2 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B2_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s32) = G_CONSTANT i32 -1
+    %1:vrb(<vscale x 2 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 2 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv4i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv4i1
+    ; CHECK: [[PseudoVMCLR_M_B4_:%[0-9]+]]:vr = PseudoVMCLR_M_B4 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B4_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s32) = G_CONSTANT i32 -1
+    %1:vrb(<vscale x 4 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 4 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv8i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv8i1
+    ; CHECK: [[PseudoVMCLR_M_B8_:%[0-9]+]]:vr = PseudoVMCLR_M_B8 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B8_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s32) = G_CONSTANT i32 -1
+    %1:vrb(<vscale x 8 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 8 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv16i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv16i1
+    ; CHECK: [[PseudoVMCLR_M_B16_:%[0-9]+]]:vr = PseudoVMCLR_M_B16 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B16_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s32) = G_CONSTANT i32 -1
+    %1:vrb(<vscale x 16 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 16 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv32i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv32i1
+    ; CHECK: [[PseudoVMCLR_M_B32_:%[0-9]+]]:vr = PseudoVMCLR_M_B32 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B32_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s32) = G_CONSTANT i32 -1
+    %1:vrb(<vscale x 32 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 32 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv64i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv64i1
+    ; CHECK: [[PseudoVMCLR_M_B64_:%[0-9]+]]:vr = PseudoVMCLR_M_B64 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B64_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s32) = G_CONSTANT i32 -1
+    %1:vrb(<vscale x 64 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 64 x s1>)
+    PseudoRET implicit $v0
+
+...
+
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/vmclr-rv64.mir b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/vmclr-rv64.mir
new file mode 100644
index 00000000000000..b7541cd4e96fb4
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/vmclr-rv64.mir
@@ -0,0 +1,124 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
+# RUN: llc -mtriple=riscv64 -mattr=+v,+m -run-pass=instruction-select \
+# RUN:   -simplify-mir -verify-machineinstrs %s -o - | FileCheck %s
+
+---
+name:            splat_zero_nxv1i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv1i1
+    ; CHECK: [[PseudoVMCLR_M_B1_:%[0-9]+]]:vr = PseudoVMCLR_M_B1 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B1_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s64) = G_CONSTANT i64 -1
+    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s64)
+    $v0 = COPY %1(<vscale x 1 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv2i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv2i1
+    ; CHECK: [[PseudoVMCLR_M_B2_:%[0-9]+]]:vr = PseudoVMCLR_M_B2 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B2_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s64) = G_CONSTANT i64 -1
+    %1:vrb(<vscale x 2 x s1>) = G_VMCLR_VL %0(s64)
+    $v0 = COPY %1(<vscale x 2 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv4i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv4i1
+    ; CHECK: [[PseudoVMCLR_M_B4_:%[0-9]+]]:vr = PseudoVMCLR_M_B4 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B4_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s64) = G_CONSTANT i64 -1
+    %1:vrb(<vscale x 4 x s1>) = G_VMCLR_VL %0(s64)
+    $v0 = COPY %1(<vscale x 4 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv8i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv8i1
+    ; CHECK: [[PseudoVMCLR_M_B8_:%[0-9]+]]:vr = PseudoVMCLR_M_B8 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B8_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s64) = G_CONSTANT i64 -1
+    %1:vrb(<vscale x 8 x s1>) = G_VMCLR_VL %0(s64)
+    $v0 = COPY %1(<vscale x 8 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv16i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv16i1
+    ; CHECK: [[PseudoVMCLR_M_B16_:%[0-9]+]]:vr = PseudoVMCLR_M_B16 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B16_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s64) = G_CONSTANT i64 -1
+    %1:vrb(<vscale x 16 x s1>) = G_VMCLR_VL %0(s64)
+    $v0 = COPY %1(<vscale x 16 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv32i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv32i1
+    ; CHECK: [[PseudoVMCLR_M_B32_:%[0-9]+]]:vr = PseudoVMCLR_M_B32 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B32_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s64) = G_CONSTANT i64 -1
+    %1:vrb(<vscale x 32 x s1>) = G_VMCLR_VL %0(s64)
+    $v0 = COPY %1(<vscale x 32 x s1>)
+    PseudoRET implicit $v0
+
+...
+---
+name:            splat_zero_nxv64i1
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: splat_zero_nxv64i1
+    ; CHECK: [[PseudoVMCLR_M_B64_:%[0-9]+]]:vr = PseudoVMCLR_M_B64 -1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B64_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s64) = G_CONSTANT i64 -1
+    %1:vrb(<vscale x 64 x s1>) = G_VMCLR_VL %0(s64)
+    $v0 = COPY %1(<vscale x 64 x s1>)
+    PseudoRET implicit $v0
+
+...
+

>From b0762dfdd9b2556799d544e743d4125f8f98c930 Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Tue, 1 Oct 2024 19:46:54 -0700
Subject: [PATCH 2/5] fixup! demonstrate the functionality supported in
 selectVLOp

---
 .../rvv/fail-select-vl-badnegative.mir        | 17 +++++++++++++++++
 .../rvv/fail-select-vl-nonconst.mir           | 19 +++++++++++++++++++
 .../rvv/fail-select-vl-nonzero.mir            | 19 +++++++++++++++++++
 .../rvv/fail-select-vl-zero.mir               | 18 ++++++++++++++++++
 4 files changed, 73 insertions(+)
 create mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-badnegative.mir
 create mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonconst.mir
 create mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonzero.mir
 create mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-zero.mir

diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-badnegative.mir b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-badnegative.mir
new file mode 100644
index 00000000000000..f71c0aab9c05ef
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-badnegative.mir
@@ -0,0 +1,17 @@
+# RUN: not --crash llc -mtriple=riscv32 -mattr=+v,+m -run-pass=instruction-select \
+# RUN:   -verify-machineinstrs %s 2>&1 | FileCheck %s
+
+---
+# CHECK: LLVM ERROR: cannot select: 
+name:            bad_negative_vl
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    %0:gprb(s32) = G_CONSTANT i32 -2
+    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 1 x s1>)
+    PseudoRET implicit $v0
+...
+
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonconst.mir b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonconst.mir
new file mode 100644
index 00000000000000..88d301f752a7dc
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonconst.mir
@@ -0,0 +1,19 @@
+# RUN: not --crash llc -mtriple=riscv32 -mattr=+v,+m -run-pass=instruction-select \
+# RUN:   -verify-machineinstrs %s 2>&1 | FileCheck %s
+
+# FIXME: This should be supported in the future
+---
+name:            not_implemented_vl_nonconst
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    liveins: $x10
+    %0:gprb(s32) = COPY $x10
+    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 1 x s1>)
+    PseudoRET implicit $v0
+
+# CHECK: LLVM ERROR: cannot select
+...
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonzero.mir b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonzero.mir
new file mode 100644
index 00000000000000..f12a800a4a2eef
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonzero.mir
@@ -0,0 +1,19 @@
+# RUN: not --crash llc -mtriple=riscv32 -mattr=+v,+m -run-pass=instruction-select \
+# RUN:   -verify-machineinstrs %s 2>&1 | FileCheck %s
+
+# FIXME: This should be supported in the future
+---
+name:            not_implemented_vl_nonzero
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    %0:gprb(s32) = G_CONSTANT i32 1
+    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 1 x s1>)
+    PseudoRET implicit $v0
+
+# CHECK: LLVM ERROR: cannot select
+...
+
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-zero.mir b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-zero.mir
new file mode 100644
index 00000000000000..9338e4fb32e3a2
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-zero.mir
@@ -0,0 +1,18 @@
+# RUN: not --crash llc -mtriple=riscv32 -mattr=+v,+m -run-pass=instruction-select \
+# RUN:   -verify-machineinstrs %s 2>&1 | FileCheck %s
+
+# FIXME: This should be supported in the future
+---
+# CHECK: LLVM ERROR: cannot select
+name:            not_implemented_vl0
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    %0:gprb(s32) = G_CONSTANT i32 0
+    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 1 x s1>)
+    PseudoRET implicit $v0
+
+...

>From 16d50d2a817536cfe67ccc0ac4da871d3ae49317 Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Tue, 1 Oct 2024 20:47:18 -0700
Subject: [PATCH 3/5] fixup! implement non-VLMAX cases in selectVLOp

---
 .../RISCV/GISel/RISCVInstructionSelector.cpp  | 26 ++++---
 .../rvv/fail-select-vl-nonconst.mir           | 19 -----
 .../rvv/fail-select-vl-nonzero.mir            | 19 -----
 .../rvv/fail-select-vl-zero.mir               | 18 -----
 .../instruction-select/rvv/select-vlop.mir    | 70 +++++++++++++++++++
 5 files changed, 85 insertions(+), 67 deletions(-)
 delete mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonconst.mir
 delete mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonzero.mir
 delete mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-zero.mir
 create mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/select-vlop.mir

diff --git a/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp b/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
index cb28f40d64b176..a5847d8c678a74 100644
--- a/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
+++ b/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
@@ -388,17 +388,21 @@ RISCVInstructionSelector::selectVLOp(MachineOperand &Root) const {
   assert(Root.isReg() && "Expected operand to be a Register");
   MachineInstr *RootDef = MRI.getVRegDef(Root.getReg());
 
-  if (RootDef->getOpcode() == TargetOpcode::G_CONSTANT &&
-      RootDef->getOperand(1).getCImm()->getSExtValue() == RISCV::VLMaxSentinel)
-    // If the operand is a G_CONSTANT with value VLMaxSentinel, convert it
-    // to an immediate with value VLMaxSentinel. This is recognized specially by
-    // the vsetvli insertion pass.
-    return {
-        {[=](MachineInstrBuilder &MIB) { MIB.addImm(RISCV::VLMaxSentinel); }}};
-
-  // FIXME: Implement non-VLMAX case. ISEL will fail gracefully by returning
-  // like this for now.
-  return std::nullopt;
+  if (RootDef->getOpcode() == TargetOpcode::G_CONSTANT) {
+    auto C = RootDef->getOperand(1).getCImm();
+    if (C->getSExtValue() == RISCV::VLMaxSentinel || C->getValue().isAllOnes())
+      // If the operand is a G_CONSTANT with value VLMaxSentinel or all ones,
+      // convert it to an immediate with value VLMaxSentinel. This is recognized
+      // specially by the vsetvli insertion pass.
+      return {{[=](MachineInstrBuilder &MIB) {
+        MIB.addImm(RISCV::VLMaxSentinel);
+      }}};
+
+    if (isUInt<5>(C->getZExtValue()))
+      return {
+          {[=](MachineInstrBuilder &MIB) { MIB.addImm(C->getZExtValue()); }}};
+  }
+  return {{[=](MachineInstrBuilder &MIB) { MIB.addReg(Root.getReg()); }}};
 }
 
 InstructionSelector::ComplexRendererFns
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonconst.mir b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonconst.mir
deleted file mode 100644
index 88d301f752a7dc..00000000000000
--- a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonconst.mir
+++ /dev/null
@@ -1,19 +0,0 @@
-# RUN: not --crash llc -mtriple=riscv32 -mattr=+v,+m -run-pass=instruction-select \
-# RUN:   -verify-machineinstrs %s 2>&1 | FileCheck %s
-
-# FIXME: This should be supported in the future
----
-name:            not_implemented_vl_nonconst
-legalized:       true
-regBankSelected: true
-tracksRegLiveness: true
-body:             |
-  bb.1:
-    liveins: $x10
-    %0:gprb(s32) = COPY $x10
-    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
-    $v0 = COPY %1(<vscale x 1 x s1>)
-    PseudoRET implicit $v0
-
-# CHECK: LLVM ERROR: cannot select
-...
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonzero.mir b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonzero.mir
deleted file mode 100644
index f12a800a4a2eef..00000000000000
--- a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-nonzero.mir
+++ /dev/null
@@ -1,19 +0,0 @@
-# RUN: not --crash llc -mtriple=riscv32 -mattr=+v,+m -run-pass=instruction-select \
-# RUN:   -verify-machineinstrs %s 2>&1 | FileCheck %s
-
-# FIXME: This should be supported in the future
----
-name:            not_implemented_vl_nonzero
-legalized:       true
-regBankSelected: true
-tracksRegLiveness: true
-body:             |
-  bb.1:
-    %0:gprb(s32) = G_CONSTANT i32 1
-    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
-    $v0 = COPY %1(<vscale x 1 x s1>)
-    PseudoRET implicit $v0
-
-# CHECK: LLVM ERROR: cannot select
-...
-
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-zero.mir b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-zero.mir
deleted file mode 100644
index 9338e4fb32e3a2..00000000000000
--- a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/fail-select-vl-zero.mir
+++ /dev/null
@@ -1,18 +0,0 @@
-# RUN: not --crash llc -mtriple=riscv32 -mattr=+v,+m -run-pass=instruction-select \
-# RUN:   -verify-machineinstrs %s 2>&1 | FileCheck %s
-
-# FIXME: This should be supported in the future
----
-# CHECK: LLVM ERROR: cannot select
-name:            not_implemented_vl0
-legalized:       true
-regBankSelected: true
-tracksRegLiveness: true
-body:             |
-  bb.1:
-    %0:gprb(s32) = G_CONSTANT i32 0
-    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
-    $v0 = COPY %1(<vscale x 1 x s1>)
-    PseudoRET implicit $v0
-
-...
diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/select-vlop.mir b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/select-vlop.mir
new file mode 100644
index 00000000000000..d43195c8b00968
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/instruction-select/rvv/select-vlop.mir
@@ -0,0 +1,70 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -mtriple=riscv32 -mattr=+v,+m -run-pass=instruction-select \
+# RUN:   -verify-machineinstrs %s -o - | FileCheck %s
+
+name:            negative_vl
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    %0:gprb(s32) = G_CONSTANT i32 -2
+    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 1 x s1>)
+    PseudoRET implicit $v0
+...
+---
+name:            nonconst_vl
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    liveins: $x10
+    ; CHECK-LABEL: name: nonconst_vl
+    ; CHECK: liveins: $x10
+    ; CHECK-NEXT: {{  $}}
+    ; CHECK-NEXT: [[COPY:%[0-9]+]]:gprnox0 = COPY $x10
+    ; CHECK-NEXT: [[PseudoVMCLR_M_B1_:%[0-9]+]]:vr = PseudoVMCLR_M_B1 [[COPY]], 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B1_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s32) = COPY $x10
+    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 1 x s1>)
+    PseudoRET implicit $v0
+...
+
+---
+name:            nonzero_vl
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: nonzero_vl
+    ; CHECK: [[PseudoVMCLR_M_B1_:%[0-9]+]]:vr = PseudoVMCLR_M_B1 1, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B1_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s32) = G_CONSTANT i32 1
+    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 1 x s1>)
+    PseudoRET implicit $v0
+...
+
+---
+name:            zero_vl
+legalized:       true
+regBankSelected: true
+tracksRegLiveness: true
+body:             |
+  bb.1:
+    ; CHECK-LABEL: name: zero_vl
+    ; CHECK: [[PseudoVMCLR_M_B1_:%[0-9]+]]:vr = PseudoVMCLR_M_B1 0, 0 /* e8 */
+    ; CHECK-NEXT: $v0 = COPY [[PseudoVMCLR_M_B1_]]
+    ; CHECK-NEXT: PseudoRET implicit $v0
+    %0:gprb(s32) = G_CONSTANT i32 0
+    %1:vrb(<vscale x 1 x s1>) = G_VMCLR_VL %0(s32)
+    $v0 = COPY %1(<vscale x 1 x s1>)
+    PseudoRET implicit $v0
+...
+

>From 2d1408e537071f8f2b6595ed27c45ed42b93fb1d Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Tue, 1 Oct 2024 20:52:51 -0700
Subject: [PATCH 4/5] fixup! remove check for sentinel

---
 llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp b/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
index a5847d8c678a74..c877efcf794819 100644
--- a/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
+++ b/llvm/lib/Target/RISCV/GISel/RISCVInstructionSelector.cpp
@@ -390,10 +390,10 @@ RISCVInstructionSelector::selectVLOp(MachineOperand &Root) const {
 
   if (RootDef->getOpcode() == TargetOpcode::G_CONSTANT) {
     auto C = RootDef->getOperand(1).getCImm();
-    if (C->getSExtValue() == RISCV::VLMaxSentinel || C->getValue().isAllOnes())
-      // If the operand is a G_CONSTANT with value VLMaxSentinel or all ones,
-      // convert it to an immediate with value VLMaxSentinel. This is recognized
-      // specially by the vsetvli insertion pass.
+    if (C->getValue().isAllOnes())
+      // If the operand is a G_CONSTANT with value of all ones it is larger than
+      // VLMAX. We convert it to an immediate with value VLMaxSentinel. This is
+      // recognized specially by the vsetvli insertion pass.
       return {{[=](MachineInstrBuilder &MIB) {
         MIB.addImm(RISCV::VLMaxSentinel);
       }}};

>From 7434b72e31e3fa5bb2982bf5aa858ea3f79cbefc Mon Sep 17 00:00:00 2001
From: Michael Maitland <michaeltmaitland at gmail.com>
Date: Tue, 1 Oct 2024 21:03:36 -0700
Subject: [PATCH 5/5] fixup! add end to end

---
 .../RISCV/GlobalISel/rvv/splat-vector.ll      | 73 +++++++++++++++++++
 1 file changed, 73 insertions(+)
 create mode 100644 llvm/test/CodeGen/RISCV/GlobalISel/rvv/splat-vector.ll

diff --git a/llvm/test/CodeGen/RISCV/GlobalISel/rvv/splat-vector.ll b/llvm/test/CodeGen/RISCV/GlobalISel/rvv/splat-vector.ll
new file mode 100644
index 00000000000000..4e58c4dcec2d85
--- /dev/null
+++ b/llvm/test/CodeGen/RISCV/GlobalISel/rvv/splat-vector.ll
@@ -0,0 +1,73 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 5
+; RUN: llc -mtriple=riscv32 -mattr=+v,+zvfh -global-isel \
+; RUN:   -verify-machineinstrs < %s | FileCheck -check-prefixes=CHECK,RV32 %s
+; RUN: llc -mtriple=riscv64 -mattr=+v,+zvfh -global-isel \
+; RUN:   -verify-machineinstrs < %s | FileCheck -check-prefixes=CHECK,RV64 %s
+
+define <vscale x 1 x i1> @splat_zero_nxv1i1() {
+; CHECK-LABEL: splat_zero_nxv1i1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e8, mf8, ta, ma
+; CHECK-NEXT:    vmclr.m v0
+; CHECK-NEXT:    ret
+  ret <vscale x 1 x i1> zeroinitializer
+}
+
+define <vscale x 2 x i1> @splat_zero_nxv2i1() {
+; CHECK-LABEL: splat_zero_nxv2i1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e8, mf4, ta, ma
+; CHECK-NEXT:    vmclr.m v0
+; CHECK-NEXT:    ret
+  ret <vscale x 2 x i1> zeroinitializer
+}
+
+define <vscale x 4 x i1> @splat_zero_nxv4i1() {
+; CHECK-LABEL: splat_zero_nxv4i1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e8, mf2, ta, ma
+; CHECK-NEXT:    vmclr.m v0
+; CHECK-NEXT:    ret
+  ret <vscale x 4 x i1> zeroinitializer
+}
+
+define <vscale x 8 x i1> @splat_zero_nxv8i1() {
+; CHECK-LABEL: splat_zero_nxv8i1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e8, m1, ta, ma
+; CHECK-NEXT:    vmclr.m v0
+; CHECK-NEXT:    ret
+  ret <vscale x 8 x i1> zeroinitializer
+}
+
+define <vscale x 16 x i1> @splat_zero_nxv16i1() {
+; CHECK-LABEL: splat_zero_nxv16i1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e8, m2, ta, ma
+; CHECK-NEXT:    vmclr.m v0
+; CHECK-NEXT:    ret
+  ret <vscale x 16 x i1> zeroinitializer
+}
+
+define <vscale x 32 x i1> @splat_zero_nxv32i1() {
+; CHECK-LABEL: splat_zero_nxv32i1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e8, m4, ta, ma
+; CHECK-NEXT:    vmclr.m v0
+; CHECK-NEXT:    ret
+  ret <vscale x 32 x i1> zeroinitializer
+}
+
+define <vscale x 64 x i1> @splat_zero_nxv64i1() {
+; CHECK-LABEL: splat_zero_nxv64i1:
+; CHECK:       # %bb.0:
+; CHECK-NEXT:    vsetvli a0, zero, e8, m8, ta, ma
+; CHECK-NEXT:    vmclr.m v0
+; CHECK-NEXT:    ret
+  ret <vscale x 64 x i1> zeroinitializer
+}
+
+
+;; NOTE: These prefixes are unused and the list is autogenerated. Do not add tests below this line:
+; RV32: {{.*}}
+; RV64: {{.*}}



More information about the llvm-commits mailing list