AArch64 SIMD: pattern-match load+splat into LD1R instruction.

cranelift/codegen/src/isa/aarch64/inst/args.rs

@@ -209,6 +209,19 @@ impl AMode {
     pub fn label(label: MemLabel) -> AMode {
         AMode::Label(label)
     }
+
+    /// Does the address resolve to just a register value, with no offset or
+    /// other computation?
+    pub fn is_reg(&self) -> Option<Reg> {
+        match self {
+            &AMode::UnsignedOffset(r, uimm12) if uimm12.value() == 0 => Some(r),
+            &AMode::Unscaled(r, imm9) if imm9.value() == 0 => Some(r),
+            &AMode::RegOffset(r, off, _) if off == 0 => Some(r),
+            &AMode::FPOffset(off, _) if off == 0 => Some(fp_reg()),
+            &AMode::SPOffset(off, _) if off == 0 => Some(stack_reg()),
+            _ => None,
+        }
+    }
 }
 
 /// A memory argument to a load/store-pair.

cranelift/codegen/src/isa/aarch64/inst/mod.rs

@@ -1463,6 +1463,17 @@ impl Inst {
             }
         }
     }
+
+    /// Generate a LoadAddr instruction (load address of an amode into
+    /// register). Elides when possible (when amode is just a register). Returns
+    /// destination register: either `rd` or a register directly from the amode.
+    pub fn gen_load_addr(rd: Writable<Reg>, mem: AMode) -> (Reg, Option<Inst>) {
+        if let Some(r) = mem.is_reg() {
+            (r, None)
+        } else {
+            (rd.to_reg(), Some(Inst::LoadAddr { rd, mem }))
+        }
+    }
 }
 
 //=============================================================================

cranelift/codegen/src/isa/aarch64/lower.rs

@@ -1169,6 +1169,59 @@ pub(crate) fn normalize_bool_result<C: LowerCtx<I = Inst>>(
     }
 }
 
+/// This is target-word-size dependent.  And it excludes booleans and reftypes.
+pub(crate) fn is_valid_atomic_transaction_ty(ty: Type) -> bool {
+    match ty {
+        I8 | I16 | I32 | I64 => true,
+        _ => false,
+    }
+}
+
+fn load_op_to_ty(op: Opcode) -> Option<Type> {
+    match op {
+        Opcode::Sload8 | Opcode::Uload8 | Opcode::Sload8Complex | Opcode::Uload8Complex => Some(I8),
+        Opcode::Sload16 | Opcode::Uload16 | Opcode::Sload16Complex | Opcode::Uload16Complex => {
+            Some(I16)
+        }
+        Opcode::Sload32 | Opcode::Uload32 | Opcode::Sload32Complex | Opcode::Uload32Complex => {
+            Some(I32)
+        }
+        Opcode::Load | Opcode::LoadComplex => None,
+        Opcode::Sload8x8 | Opcode::Uload8x8 | Opcode::Sload8x8Complex | Opcode::Uload8x8Complex => {
+            Some(I8X8)
+        }
+        Opcode::Sload16x4
+        | Opcode::Uload16x4
+        | Opcode::Sload16x4Complex
+        | Opcode::Uload16x4Complex => Some(I16X4),
+        Opcode::Sload32x2
+        | Opcode::Uload32x2
+        | Opcode::Sload32x2Complex
+        | Opcode::Uload32x2Complex => Some(I32X2),
+        _ => None,
+    }
+}
+
+/// Helper to lower a load instruction; this is used in several places, because
+/// a load can sometimes be merged into another operation.
+pub(crate) fn lower_load<C: LowerCtx<I = Inst>, F: FnMut(&mut C, Writable<Reg>, Type, AMode)>(
+    ctx: &mut C,
+    ir_inst: IRInst,
+    inputs: &[InsnInput],
+    output: InsnOutput,
+    mut f: F,
+) {
+    let op = ctx.data(ir_inst).opcode();
+
+    let elem_ty = load_op_to_ty(op).unwrap_or_else(|| ctx.output_ty(ir_inst, 0));
+
+    let off = ctx.data(ir_inst).load_store_offset().unwrap();
+    let mem = lower_address(ctx, elem_ty, &inputs[..], off);
+    let rd = get_output_reg(ctx, output);
+
+    f(ctx, rd, elem_ty, mem);
+}
+
 //=============================================================================
 // Lowering-backend trait implementation.
 

cranelift/codegen/src/isa/aarch64/lower_inst.rs

@@ -17,30 +17,17 @@ use regalloc::{RegClass, Writable};
 use alloc::boxed::Box;
 use alloc::vec::Vec;
 use core::convert::TryFrom;
-use smallvec::SmallVec;
 
 use super::lower::*;
 
-/// This is target-word-size dependent.  And it excludes booleans and reftypes.
-fn is_valid_atomic_transaction_ty(ty: Type) -> bool {
-    match ty {
-        I8 | I16 | I32 | I64 => true,
-        _ => false,
-    }
-}
-
 /// Actually codegen an instruction's results into registers.
 pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
     ctx: &mut C,
     insn: IRInst,
 ) -> CodegenResult<()> {
     let op = ctx.data(insn).opcode();
-    let inputs: SmallVec<[InsnInput; 4]> = (0..ctx.num_inputs(insn))
+    let inputs = insn_inputs(ctx, insn);
-        .map(|i| InsnInput { insn, input: i })
+    let outputs = insn_outputs(ctx, insn);
-        .collect();
-    let outputs: SmallVec<[InsnOutput; 2]> = (0..ctx.num_outputs(insn))
-        .map(|i| InsnOutput { insn, output: i })
-        .collect();
     let ty = if outputs.len() > 0 {
         Some(ctx.output_ty(insn, 0))
     } else {
@@ -1134,34 +1121,6 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
         | Opcode::Sload16x4Complex
         | Opcode::Uload32x2Complex
         | Opcode::Sload32x2Complex => {
-            let off = ctx.data(insn).load_store_offset().unwrap();
-            let elem_ty = match op {
-                Opcode::Sload8 | Opcode::Uload8 | Opcode::Sload8Complex | Opcode::Uload8Complex => {
-                    I8
-                }
-                Opcode::Sload16
-                | Opcode::Uload16
-                | Opcode::Sload16Complex
-                | Opcode::Uload16Complex => I16,
-                Opcode::Sload32
-                | Opcode::Uload32
-                | Opcode::Sload32Complex
-                | Opcode::Uload32Complex => I32,
-                Opcode::Load | Opcode::LoadComplex => ctx.output_ty(insn, 0),
-                Opcode::Sload8x8
-                | Opcode::Uload8x8
-                | Opcode::Sload8x8Complex
-                | Opcode::Uload8x8Complex => I8X8,
-                Opcode::Sload16x4
-                | Opcode::Uload16x4
-                | Opcode::Sload16x4Complex
-                | Opcode::Uload16x4Complex => I16X4,
-                Opcode::Sload32x2
-                | Opcode::Uload32x2
-                | Opcode::Sload32x2Complex
-                | Opcode::Uload32x2Complex => I32X2,
-                _ => unreachable!(),
-            };
             let sign_extend = match op {
                 Opcode::Sload8
                 | Opcode::Sload8Complex
@@ -1171,72 +1130,59 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
                 | Opcode::Sload32Complex => true,
                 _ => false,
             };
-            let is_float = ty_has_float_or_vec_representation(elem_ty);
 
-            let mem = lower_address(ctx, elem_ty, &inputs[..], off);
+            lower_load(
-            let rd = get_output_reg(ctx, outputs[0]);
+                ctx,
+                insn,
+                &inputs[..],
+                outputs[0],
+                |ctx, rd, elem_ty, mem| {
+                    let is_float = ty_has_float_or_vec_representation(elem_ty);
+                    ctx.emit(match (ty_bits(elem_ty), sign_extend, is_float) {
+                        (1, _, _) => Inst::ULoad8 { rd, mem },
+                        (8, false, _) => Inst::ULoad8 { rd, mem },
+                        (8, true, _) => Inst::SLoad8 { rd, mem },
+                        (16, false, _) => Inst::ULoad16 { rd, mem },
+                        (16, true, _) => Inst::SLoad16 { rd, mem },
+                        (32, false, false) => Inst::ULoad32 { rd, mem },
+                        (32, true, false) => Inst::SLoad32 { rd, mem },
+                        (32, _, true) => Inst::FpuLoad32 { rd, mem },
+                        (64, _, false) => Inst::ULoad64 { rd, mem },
+                        // Note that we treat some of the vector loads as scalar floating-point loads,
+                        // which is correct in a little endian environment.
+                        (64, _, true) => Inst::FpuLoad64 { rd, mem },
+                        (128, _, _) => Inst::FpuLoad128 { rd, mem },
+                        _ => panic!("Unsupported size in load"),
+                    });
 
-            ctx.emit(match (ty_bits(elem_ty), sign_extend, is_float) {
+                    let vec_extend = match op {
-                (1, _, _) => Inst::ULoad8 { rd, mem },
+                        Opcode::Sload8x8 => Some(VecExtendOp::Sxtl8),
-                (8, false, _) => Inst::ULoad8 { rd, mem },
+                        Opcode::Sload8x8Complex => Some(VecExtendOp::Sxtl8),
-                (8, true, _) => Inst::SLoad8 { rd, mem },
+                        Opcode::Uload8x8 => Some(VecExtendOp::Uxtl8),
-                (16, false, _) => Inst::ULoad16 { rd, mem },
+                        Opcode::Uload8x8Complex => Some(VecExtendOp::Uxtl8),
-                (16, true, _) => Inst::SLoad16 { rd, mem },
+                        Opcode::Sload16x4 => Some(VecExtendOp::Sxtl16),
-                (32, false, false) => Inst::ULoad32 { rd, mem },
+                        Opcode::Sload16x4Complex => Some(VecExtendOp::Sxtl16),
-                (32, true, false) => Inst::SLoad32 { rd, mem },
+                        Opcode::Uload16x4 => Some(VecExtendOp::Uxtl16),
-                (32, _, true) => Inst::FpuLoad32 { rd, mem },
+                        Opcode::Uload16x4Complex => Some(VecExtendOp::Uxtl16),
-                (64, _, false) => Inst::ULoad64 { rd, mem },
+                        Opcode::Sload32x2 => Some(VecExtendOp::Sxtl32),
-                // Note that we treat some of the vector loads as scalar floating-point loads,
+                        Opcode::Sload32x2Complex => Some(VecExtendOp::Sxtl32),
-                // which is correct in a little endian environment.
+                        Opcode::Uload32x2 => Some(VecExtendOp::Uxtl32),
-                (64, _, true) => Inst::FpuLoad64 { rd, mem },
+                        Opcode::Uload32x2Complex => Some(VecExtendOp::Uxtl32),
-                (128, _, _) => Inst::FpuLoad128 { rd, mem },
+                        _ => None,
-                _ => panic!("Unsupported size in load"),
+                    };
-            });
 
-            let vec_extend = match op {
+                    if let Some(t) = vec_extend {
-                Opcode::Sload8x8 => Some(VecExtendOp::Sxtl8),
+                        ctx.emit(Inst::VecExtend {
-                Opcode::Sload8x8Complex => Some(VecExtendOp::Sxtl8),
+                            t,
-                Opcode::Uload8x8 => Some(VecExtendOp::Uxtl8),
+                            rd,
-                Opcode::Uload8x8Complex => Some(VecExtendOp::Uxtl8),
+                            rn: rd.to_reg(),
-                Opcode::Sload16x4 => Some(VecExtendOp::Sxtl16),
+                            high_half: false,
-                Opcode::Sload16x4Complex => Some(VecExtendOp::Sxtl16),
+                        });
-                Opcode::Uload16x4 => Some(VecExtendOp::Uxtl16),
+                    }
-                Opcode::Uload16x4Complex => Some(VecExtendOp::Uxtl16),
+                },
-                Opcode::Sload32x2 => Some(VecExtendOp::Sxtl32),
+            );
-                Opcode::Sload32x2Complex => Some(VecExtendOp::Sxtl32),
-                Opcode::Uload32x2 => Some(VecExtendOp::Uxtl32),
-                Opcode::Uload32x2Complex => Some(VecExtendOp::Uxtl32),
-                _ => None,
-            };
-
-            if let Some(t) = vec_extend {
-                ctx.emit(Inst::VecExtend {
-                    t,
-                    rd,
-                    rn: rd.to_reg(),
-                    high_half: false,
-                });
-            }
         }
 
-        /*
-        Opcode::LoadSplat => {
-            let off = ctx.data(insn).load_store_offset().unwrap();
-            let ty = ty.unwrap();
-            let mem = lower_address(ctx, ty.lane_type(), &inputs[..], off);
-            let rd = get_output_reg(ctx, outputs[0]);
-            let size = VectorSize::from_ty(ty);
-            let tmp = ctx.alloc_tmp(RegClass::I64, I64);
-
-            ctx.emit(Inst::LoadAddr { rd: tmp, mem });
-            ctx.emit(Inst::VecLoadReplicate {
-                rd,
-                rn: tmp.to_reg(),
-                size,
-            });
-        }
-        */
-
         Opcode::Store
         | Opcode::Istore8
         | Opcode::Istore16
@@ -2028,6 +1974,36 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
                 maybe_input_insn_via_conv(ctx, inputs[0], Opcode::Bconst, Opcode::Breduce)
             {
                 lower_splat_const(ctx, rd, ctx.get_constant(insn).unwrap(), size);
+            } else if let Some((_, insn)) = maybe_input_insn_multi(
+                ctx,
+                inputs[0],
+                &[
+                    Opcode::Uload8,
+                    Opcode::Sload8,
+                    Opcode::Uload16,
+                    Opcode::Sload16,
+                    Opcode::Uload32,
+                    Opcode::Sload32,
+                    Opcode::Load,
+                ],
+            ) {
+                ctx.sink_inst(insn);
+                let load_inputs = insn_inputs(ctx, insn);
+                let load_outputs = insn_outputs(ctx, insn);
+                lower_load(
+                    ctx,
+                    insn,
+                    &load_inputs[..],
+                    load_outputs[0],
+                    |ctx, _rd, _elem_ty, mem| {
+                        let tmp = ctx.alloc_tmp(RegClass::I64, I64);
+                        let (addr, addr_inst) = Inst::gen_load_addr(tmp, mem);
+                        if let Some(addr_inst) = addr_inst {
+                            ctx.emit(addr_inst);
+                        }
+                        ctx.emit(Inst::VecLoadReplicate { rd, rn: addr, size });
+                    },
+                );
             } else {
                 let input_ty = ctx.input_ty(insn, 0);
                 let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);

cranelift/codegen/src/isa/x64/lower.rs

@@ -3661,21 +3661,13 @@ fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
             emit_extract_lane(ctx, src, dst, lane, ty);
         }
 
-        Opcode::Splat | Opcode::LoadSplat => {
+        Opcode::Splat => {
             let ty = ty.unwrap();
             assert_eq!(ty.bits(), 128);
             let src_ty = ctx.input_ty(insn, 0);
             assert!(src_ty.bits() < 128);
 
-            let src = match op {
+            let src = input_to_reg_mem(ctx, inputs[0]);
-                Opcode::Splat => input_to_reg_mem(ctx, inputs[0]),
-                Opcode::LoadSplat => {
-                    let offset = ctx.data(insn).load_store_offset().unwrap();
-                    let amode = lower_to_amode(ctx, inputs[0], offset);
-                    RegMem::mem(amode)
-                }
-                _ => unreachable!(),
-            };
             let dst = get_output_reg(ctx, outputs[0]);
 
             // We know that splat will overwrite all of the lanes of `dst` but it takes several

cranelift/codegen/src/machinst/inst_common.rs

@@ -1,6 +1,8 @@
 //! A place to park MachInst::Inst fragments which are common across multiple architectures.
 
+use super::{LowerCtx, VCodeInst};
 use crate::ir::{self, Inst as IRInst};
+use smallvec::SmallVec;
 
 //============================================================================
 // Instruction input "slots".
@@ -22,6 +24,24 @@ pub(crate) struct InsnOutput {
     pub(crate) output: usize,
 }
 
+pub(crate) fn insn_inputs<I: VCodeInst, C: LowerCtx<I = I>>(
+    ctx: &C,
+    insn: IRInst,
+) -> SmallVec<[InsnInput; 4]> {
+    (0..ctx.num_inputs(insn))
+        .map(|i| InsnInput { insn, input: i })
+        .collect()
+}
+
+pub(crate) fn insn_outputs<I: VCodeInst, C: LowerCtx<I = I>>(
+    ctx: &C,
+    insn: IRInst,
+) -> SmallVec<[InsnOutput; 4]> {
+    (0..ctx.num_outputs(insn))
+        .map(|i| InsnOutput { insn, output: i })
+        .collect()
+}
+
 //============================================================================
 // Atomic instructions.
 

cranelift/codegen/src/machinst/lower.rs

@@ -147,9 +147,10 @@ pub trait LowerCtx {
     /// Emit a machine instruction that is a safepoint.
     fn emit_safepoint(&mut self, mach_inst: Self::I);
     /// Indicate that the side-effect of an instruction has been sunk to the
-    /// current scan location. This can only be done to an instruction with no
+    /// current scan location. This should only be done with the instruction's
-    /// uses of its result register(s), because it will cause the instruction
+    /// original results are not used (i.e., `put_input_in_reg` is not invoked
-    /// not to be codegen'd at its original location.
+    /// for the input produced by the sunk instruction), otherwise the
+    /// side-effect will occur twice.
     fn sink_inst(&mut self, ir_inst: Inst);
     /// Retrieve constant data given a handle.
     fn get_constant_data(&self, constant_handle: Constant) -> &ConstantData;

cranelift/filetests/filetests/isa/aarch64/simd.clif

@@ -61,3 +61,69 @@ block0(v0: i32, v1: i8x16, v2: i8x16):
 ; nextln:  mov sp, fp
 ; nextln:  ldp fp, lr, [sp], #16
 ; nextln:  ret
+
+function %f5(i64) -> i8x16 {
+block0(v0: i64):
+  v1 = load.i8 v0
+  v2 = splat.i8x16 v1
+  return v2
+}
+
+; check:  stp fp, lr, [sp, #-16]!
+; nextln:  mov fp, sp
+; nextln:  ld1r { v0.16b }, [x0]
+; nextln:  mov sp, fp
+; nextln:  ldp fp, lr, [sp], #16
+; nextln:  ret
+
+function %f6(i64, i64) -> i8x16, i8x16 {
+block0(v0: i64, v1: i64):
+  v2 = load.i8 v0
+  v3 = load.i8 v1
+  v4 = splat.i8x16 v2
+  v5 = splat.i8x16 v3
+  return v4, v5
+}
+
+; check:  stp fp, lr, [sp, #-16]!
+; nextln:  mov fp, sp
+; nextln:  ld1r { v0.16b }, [x0]
+; nextln:  ld1r { v1.16b }, [x1]
+; nextln:  mov sp, fp
+; nextln:  ldp fp, lr, [sp], #16
+; nextln:  ret
+
+function %f7(i64, i64) -> i8x16, i8x16 {
+block0(v0: i64, v1: i64):
+  v2 = load.i8 v0
+  v3 = load.i8 v1
+  v4 = splat.i8x16 v3
+  v5 = splat.i8x16 v2
+  return v4, v5
+}
+
+; check:  stp fp, lr, [sp, #-16]!
+; nextln:  mov fp, sp
+; nextln:  ldrb w0, [x0]
+; nextln:  ld1r { v0.16b }, [x1]
+; nextln:  dup v1.16b, w0
+; nextln:  mov sp, fp
+; nextln:  ldp fp, lr, [sp], #16
+; nextln:  ret
+
+function %f8(i64, i64) -> i8x16, i8x16 {
+block0(v0: i64, v1: i64):
+  v2 = load.i8 v0
+  v3 = splat.i8x16 v2
+  v4 = splat.i8x16 v2
+  return v3, v4
+}
+
+; check:  stp fp, lr, [sp, #-16]!
+; nextln:  mov fp, sp
+; nextln:  ldrb w0, [x0]
+; nextln:  dup v0.16b, w0
+; nextln:  dup v1.16b, w0
+; nextln:  mov sp, fp
+; nextln:  ldp fp, lr, [sp], #16
+; nextln:  ret