Convert fadd..fmax_pseudo to ISLE (AArch64) (#4452)

Converted the existing implementations for the following Opcodes to ISLE on AArch64: - `fadd` - `fsub` - `fmul` - `fdiv` - `fmin` - `fmax` - `fmin_pseudo` - `fmax_pseudo` Copyright (c) 2022 Arm Limited
2022-07-19 20:03:05 +01:00
parent b18c9bee15
commit 00ac18c866
7 changed files with 447 additions and 154 deletions
--- a/cranelift/codegen/src/isa/aarch64/inst.isle
+++ b/cranelift/codegen/src/isa/aarch64/inst.isle
@@ -895,6 +895,16 @@
            Size64
            Size128))

+;; Helper for calculating the `ScalarSize` corresponding to a type
+(decl scalar_size (Type) ScalarSize)
+(rule (scalar_size $I8) (ScalarSize.Size8))
+(rule (scalar_size $I16) (ScalarSize.Size16))
+(rule (scalar_size $I32) (ScalarSize.Size32))
+(rule (scalar_size $I64) (ScalarSize.Size64))
+(rule (scalar_size $I128) (ScalarSize.Size128))
+(rule (scalar_size $F32) (ScalarSize.Size32))
+(rule (scalar_size $F64) (ScalarSize.Size64))
+
 (type Cond extern
  (enum
    (Eq)
@@ -1460,6 +1470,19 @@
            (_ Unit (emit (MInst.VecRRR op dst src1 src2 size))))
        dst))

+;; Helper for emitting `MInst.FpuRRR` instructions.
+(decl fpu_rrr (FPUOp2 Reg Reg ScalarSize) Reg)
+(rule (fpu_rrr op src1 src2 size)
+      (let ((dst WritableReg (temp_writable_reg $F64))
+            (_ Unit (emit (MInst.FpuRRR op size dst src1 src2))))
+        dst))
+
+;; Helper for emitting `MInst.FpuCmp` instructions.
+(decl fpu_cmp (ScalarSize Reg Reg) ProducesFlags)
+(rule (fpu_cmp size rn rm)
+      (ProducesFlags.ProducesFlagsSideEffect
+       (MInst.FpuCmp size rn rm)))
+
 ;; Helper for emitting `MInst.VecLanes` instructions.
 (decl vec_lanes (VecLanesOp Reg VectorSize) Reg)
 (rule (vec_lanes op src size)
@@ -1612,6 +1635,22 @@
            (_ Unit (emit (MInst.VecRRLong op dst src high_half))))
        dst))

+;; Helper for emitting `MInst.FpuCSel32` / `MInst.FpuCSel64`
+;; instructions.
+(decl fpu_csel (Type Cond Reg Reg) ConsumesFlags)
+(rule (fpu_csel $F32 cond if_true if_false)
+      (let ((dst WritableReg (temp_writable_reg $F32)))
+        (ConsumesFlags.ConsumesFlagsReturnsReg
+         (MInst.FpuCSel32 dst if_true if_false cond)
+         dst)))
+
+(rule (fpu_csel $F64 cond if_true if_false)
+      (let ((dst WritableReg (temp_writable_reg $F64)))
+        (ConsumesFlags.ConsumesFlagsReturnsReg
+         (MInst.FpuCSel64 dst if_true if_false cond)
+         dst)))
+
+
 ;; Helper for emitting `MInst.MovToFpu` instructions.
 (decl mov_to_fpu (Reg ScalarSize) Reg)
 (rule (mov_to_fpu x size)
--- a/cranelift/codegen/src/isa/aarch64/lower.isle
+++ b/cranelift/codegen/src/isa/aarch64/lower.isle
@@ -164,6 +164,72 @@
 (rule (lower (has_type (fits_in_32 ty) (iabs x)))
      (abs (OperandSize.Size32) (put_in_reg_sext32 x)))

+;;;; Rules for `fadd` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty @ (multi_lane _ _) (fadd rn rm)))
+      (vec_rrr (VecALUOp.Fadd) rn rm (vector_size ty)))
+
+(rule (lower (has_type (ty_scalar_float ty) (fadd rn rm)))
+      (fpu_rrr (FPUOp2.Add) rn rm (scalar_size ty)))
+
+;;;; Rules for `fsub` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty @ (multi_lane _ _) (fsub rn rm)))
+      (vec_rrr (VecALUOp.Fsub) rn rm (vector_size ty)))
+
+(rule (lower (has_type (ty_scalar_float ty) (fsub rn rm)))
+      (fpu_rrr (FPUOp2.Sub) rn rm (scalar_size ty)))
+
+;;;; Rules for `fmul` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty @ (multi_lane _ _) (fmul rn rm)))
+      (vec_rrr (VecALUOp.Fmul) rn rm (vector_size ty)))
+
+(rule (lower (has_type (ty_scalar_float ty) (fmul rn rm)))
+      (fpu_rrr (FPUOp2.Mul) rn rm (scalar_size ty)))
+
+;;;; Rules for `fdiv` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty @ (multi_lane _ _) (fdiv rn rm)))
+      (vec_rrr (VecALUOp.Fdiv) rn rm (vector_size ty)))
+
+(rule (lower (has_type (ty_scalar_float ty) (fdiv rn rm)))
+      (fpu_rrr (FPUOp2.Div) rn rm (scalar_size ty)))
+
+;;;; Rules for `fmin` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty @ (multi_lane _ _) (fmin rn rm)))
+      (vec_rrr (VecALUOp.Fmin) rn rm (vector_size ty)))
+
+(rule (lower (has_type (ty_scalar_float ty) (fmin rn rm)))
+      (fpu_rrr (FPUOp2.Min) rn rm (scalar_size ty)))
+
+;;;; Rules for `fmax` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty @ (multi_lane _ _) (fmax rn rm)))
+      (vec_rrr (VecALUOp.Fmax) rn rm (vector_size ty)))
+
+(rule (lower (has_type (ty_scalar_float ty) (fmax rn rm)))
+      (fpu_rrr (FPUOp2.Max) rn rm (scalar_size ty)))
+
+;;;; Rules for `fmin_pseudo` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty @ (multi_lane _ _) (fmin_pseudo rm rn)))
+      (bsl ty (vec_rrr (VecALUOp.Fcmgt) rm rn (vector_size ty)) rn rm))
+
+(rule (lower (has_type (ty_scalar_float ty) (fmin_pseudo rm rn)))
+      (with_flags (fpu_cmp (scalar_size ty) rm rn)
+                  (fpu_csel ty (Cond.Gt) rn rm)))
+
+;;;; Rules for `fmax_pseudo` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+
+(rule (lower (has_type ty @ (multi_lane _ _) (fmax_pseudo rm rn)))
+      (bsl ty (vec_rrr (VecALUOp.Fcmgt) rn rm (vector_size ty)) rn rm))
+
+(rule (lower (has_type (ty_scalar_float ty) (fmax_pseudo rm rn)))
+      (with_flags (fpu_cmp (scalar_size ty) rn rm)
+                  (fpu_csel ty (Cond.Gt) rn rm)))
+
 ;;;; Rules for `isub` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

 ;; `i64` and smaller
--- a/cranelift/codegen/src/isa/aarch64/lower_dynamic_neon.isle
+++ b/cranelift/codegen/src/isa/aarch64/lower_dynamic_neon.isle
@@ -19,6 +19,34 @@
 (rule (lower (has_type ty @ (dynamic_lane _ _) (fsub x y)))
      (value_reg (vec_rrr (VecALUOp.Fsub) (put_in_reg x) (put_in_reg y) (vector_size ty))))

+;;;; Rules for `fmul` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type ty @ (dynamic_lane _ _) (fmul x y)))
+      (value_reg (vec_rrr (VecALUOp.Fmul) (put_in_reg x) (put_in_reg y) (vector_size ty))))
+
+;;;; Rules for `fdiv` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type ty @ (dynamic_lane _ _) (fdiv x y)))
+      (value_reg (vec_rrr (VecALUOp.Fdiv) (put_in_reg x) (put_in_reg y) (vector_size ty))))
+
+;;;; Rules for `fmin` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type ty @ (dynamic_lane _ _) (fmin x y)))
+      (value_reg (vec_rrr (VecALUOp.Fmin) (put_in_reg x) (put_in_reg y) (vector_size ty))))
+
+;;;; Rules for `fmax` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type ty @ (dynamic_lane _ _) (fmax x y)))
+      (value_reg (vec_rrr (VecALUOp.Fmax) (put_in_reg x) (put_in_reg y) (vector_size ty))))
+
+;;;; Rules for `fmin_pseudo` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type ty @ (dynamic_lane _ _) (fmin_pseudo x y)))
+      (value_reg (bsl ty
+                  (vec_rrr (VecALUOp.Fcmgt) (put_in_reg x) (put_in_reg y)
+                   (vector_size ty)) (put_in_reg y) (put_in_reg x))))
+
+;;;; Rules for `fmax_pseudo` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
+(rule (lower (has_type ty @ (dynamic_lane _ _) (fmax_pseudo x y)))
+      (value_reg (bsl ty
+                  (vec_rrr (VecALUOp.Fcmgt) (put_in_reg y) (put_in_reg x)
+                   (vector_size ty)) (put_in_reg y) (put_in_reg x))))
+
 ;;; Rules for `dynamic_stack_addr` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 (rule (lower (dynamic_stack_addr stack_slot))
      (let ((dst WritableReg (temp_writable_reg $I64))
--- a/cranelift/codegen/src/isa/aarch64/lower_inst.rs
+++ b/cranelift/codegen/src/isa/aarch64/lower_inst.rs
@@ -1244,110 +1244,10 @@ pub(crate) fn lower_insn_to_regs<C: LowerCtx<I = Inst>>(
        }

        Opcode::Fadd | Opcode::Fsub | Opcode::Fmul | Opcode::Fdiv | Opcode::Fmin | Opcode::Fmax => {
-            let ty = ty.unwrap();
-            let rn = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
-            let rm = put_input_in_reg(ctx, inputs[1], NarrowValueMode::None);
-            let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
-            if !ty.is_vector() && !ty.is_dynamic_vector() {
-                let fpu_op = match op {
-                    Opcode::Fadd => FPUOp2::Add,
-                    Opcode::Fsub => FPUOp2::Sub,
-                    Opcode::Fmul => FPUOp2::Mul,
-                    Opcode::Fdiv => FPUOp2::Div,
-                    Opcode::Fmin => FPUOp2::Min,
-                    Opcode::Fmax => FPUOp2::Max,
-                    _ => unreachable!(),
-                };
-                ctx.emit(Inst::FpuRRR {
-                    fpu_op,
-                    size: ScalarSize::from_ty(ty),
-                    rd,
-                    rn,
-                    rm,
-                });
-            } else {
-                let alu_op = match op {
-                    Opcode::Fadd => VecALUOp::Fadd,
-                    Opcode::Fsub => VecALUOp::Fsub,
-                    Opcode::Fdiv => VecALUOp::Fdiv,
-                    Opcode::Fmax => VecALUOp::Fmax,
-                    Opcode::Fmin => VecALUOp::Fmin,
-                    Opcode::Fmul => VecALUOp::Fmul,
-                    _ => unreachable!(),
-                };
-
-                ctx.emit(Inst::VecRRR {
-                    rd,
-                    rn,
-                    rm,
-                    alu_op,
-                    size: VectorSize::from_ty(ty),
-                });
-            }
+            implemented_in_isle(ctx)
        }

-        Opcode::FminPseudo | Opcode::FmaxPseudo => {
-            let rd = get_output_reg(ctx, outputs[0]).only_reg().unwrap();
-            let rm = put_input_in_reg(ctx, inputs[0], NarrowValueMode::None);
-            let rn = put_input_in_reg(ctx, inputs[1], NarrowValueMode::None);
-            let (ra, rb) = if op == Opcode::FminPseudo {
-                (rm, rn)
-            } else {
-                (rn, rm)
-            };
-            let ty = ty.unwrap();
-            let lane_type = ty.lane_type();
-
-            debug_assert!(lane_type == F32 || lane_type == F64);
-
-            if ty.is_vector() || ty.is_dynamic_vector() {
-                let size = VectorSize::from_ty(ty);
-
-                // pmin(a,b) => bitsel(b, a, cmpgt(a, b))
-                // pmax(a,b) => bitsel(b, a, cmpgt(b, a))
-                // Since we're going to write the output register `rd` anyway, we might as well
-                // first use it to hold the comparison result.  This has the slightly unusual
-                // effect that we modify the output register in the first instruction (`fcmgt`)
-                // but read both the inputs again in the second instruction (`bsl`), which means
-                // that the output register can't be either of the input registers.  Regalloc
-                // should handle this correctly, nevertheless.
-                ctx.emit(Inst::VecRRR {
-                    alu_op: VecALUOp::Fcmgt,
-                    rd,
-                    rn: ra,
-                    rm: rb,
-                    size,
-                });
-                ctx.emit(Inst::VecRRR {
-                    alu_op: VecALUOp::Bsl,
-                    rd,
-                    rn,
-                    rm,
-                    size,
-                });
-            } else {
-                ctx.emit(Inst::FpuCmp {
-                    size: ScalarSize::from_ty(lane_type),
-                    rn: ra,
-                    rm: rb,
-                });
-                if lane_type == F32 {
-                    ctx.emit(Inst::FpuCSel32 {
-                        rd,
-                        rn,
-                        rm,
-                        cond: Cond::Gt,
-                    });
-                } else {
-                    ctx.emit(Inst::FpuCSel64 {
-                        rd,
-                        rn,
-                        rm,
-                        cond: Cond::Gt,
-                    });
-                }
-            }
-        }
+        Opcode::FminPseudo | Opcode::FmaxPseudo => implemented_in_isle(ctx),

        Opcode::Sqrt | Opcode::Fneg | Opcode::Fabs | Opcode::Fpromote | Opcode::Fdemote => {
            let ty = ty.unwrap();
--- a/cranelift/filetests/filetests/isa/aarch64/dynamic-simd-neon.clif
+++ b/cranelift/filetests/filetests/isa/aarch64/dynamic-simd-neon.clif
@@ -102,3 +102,107 @@ block0(v0: f64, v1: f64):
 ; nextln: dup v6.2d, v1.d[0]
 ; nextln: fsub v0.2d, v4.2d, v6.2d
 ; nextln: ret
+
+function %f64x2_splat_mul(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmul v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+
+; check:  dup v4.2d, v0.d[0]
+; nextln: dup v6.2d, v1.d[0]
+; nextln: fmul v0.2d, v4.2d, v6.2d
+; nextln: ret
+
+function %f64x2_splat_div(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fdiv v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+
+; check:  dup v4.2d, v0.d[0]
+; nextln: dup v6.2d, v1.d[0]
+; nextln: fdiv v0.2d, v4.2d, v6.2d
+; nextln: ret
+
+function %f64x2_splat_min(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmin v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+
+; check:  dup v4.2d, v0.d[0]
+; nextln: dup v6.2d, v1.d[0]
+; nextln: fmin v0.2d, v4.2d, v6.2d
+; nextln: ret
+
+function %f64x2_splat_max(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmax v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+
+; check:  dup v4.2d, v0.d[0]
+; nextln: dup v6.2d, v1.d[0]
+; nextln: fmax v0.2d, v4.2d, v6.2d
+; nextln: ret
+
+function %f64x2_splat_min_pseudo(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmin_pseudo v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+
+; check:  dup v4.2d, v0.d[0]
+; nextln: dup v6.2d, v1.d[0]
+; nextln: fcmgt v0.2d, v4.2d, v6.2d
+; nextln: bsl v0.16b, v6.16b, v4.16b
+; nextln: ret
+
+function %f64x2_splat_max_pseudo(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmax_pseudo v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+
+; check:  dup v4.2d, v0.d[0]
+; nextln: dup v6.2d, v1.d[0]
+; nextln: fcmgt v0.2d, v6.2d, v4.2d
+; nextln: bsl v0.16b, v6.16b, v4.16b
+; nextln: ret
--- a/cranelift/filetests/filetests/isa/aarch64/prologue.clif
+++ b/cranelift/filetests/filetests/isa/aarch64/prologue.clif
@@ -82,68 +82,68 @@ block0(v0: f64):
 ;   stp d10, d11, [sp, #-16]!
 ;   stp d8, d9, [sp, #-16]!
 ; block0:
-;   fadd d2, d0, d0
-;   fadd d4, d0, d0
-;   fadd d6, d0, d0
-;   fadd d8, d0, d0
-;   fadd d10, d0, d0
-;   fadd d12, d0, d0
-;   fadd d14, d0, d0
 ;   fadd d1, d0, d0
+;   fadd d2, d0, d0
 ;   fadd d3, d0, d0
+;   fadd d4, d0, d0
 ;   fadd d5, d0, d0
+;   fadd d6, d0, d0
 ;   fadd d7, d0, d0
+;   fadd d8, d0, d0
 ;   fadd d9, d0, d0
+;   fadd d10, d0, d0
 ;   fadd d11, d0, d0
+;   fadd d12, d0, d0
 ;   fadd d13, d0, d0
-;   fadd d30, d0, d0
+;   fadd d14, d0, d0
 ;   fadd d15, d0, d0
-;   fadd d18, d0, d0
-;   fadd d20, d0, d0
-;   fadd d22, d0, d0
-;   fadd d24, d0, d0
-;   fadd d26, d0, d0
-;   fadd d28, d0, d0
-;   fadd d31, d0, d0
 ;   fadd d16, d0, d0
-;   fadd d19, d0, d0
-;   fadd d21, d0, d0
-;   fadd d23, d0, d0
-;   fadd d25, d0, d0
-;   fadd d27, d0, d0
-;   fadd d29, d0, d0
 ;   fadd d17, d0, d0
-;   fadd d0, d0, d2
-;   fadd d2, d4, d6
-;   fadd d4, d8, d10
-;   fadd d6, d12, d14
-;   fadd d8, d1, d3
-;   fadd d10, d5, d7
-;   fadd d12, d9, d11
-;   fadd d14, d13, d30
-;   fadd d1, d15, d18
-;   fadd d3, d20, d22
-;   fadd d5, d24, d26
-;   fadd d7, d28, d31
-;   fadd d9, d16, d19
-;   fadd d11, d21, d23
-;   fadd d13, d25, d27
-;   fadd d15, d29, d17
-;   fadd d0, d0, d2
-;   fadd d2, d4, d6
-;   fadd d4, d8, d10
-;   fadd d6, d12, d14
-;   fadd d8, d1, d3
-;   fadd d10, d5, d7
-;   fadd d12, d9, d11
-;   fadd d14, d13, d15
-;   fadd d0, d0, d2
-;   fadd d2, d4, d6
-;   fadd d4, d8, d10
-;   fadd d6, d12, d14
-;   fadd d8, d0, d2
-;   fadd d10, d4, d6
-;   fadd d0, d8, d10
+;   fadd d18, d0, d0
+;   fadd d19, d0, d0
+;   fadd d20, d0, d0
+;   fadd d21, d0, d0
+;   fadd d22, d0, d0
+;   fadd d23, d0, d0
+;   fadd d24, d0, d0
+;   fadd d25, d0, d0
+;   fadd d26, d0, d0
+;   fadd d27, d0, d0
+;   fadd d28, d0, d0
+;   fadd d29, d0, d0
+;   fadd d30, d0, d0
+;   fadd d31, d0, d0
+;   fadd d0, d0, d1
+;   fadd d1, d2, d3
+;   fadd d2, d4, d5
+;   fadd d3, d6, d7
+;   fadd d4, d8, d9
+;   fadd d5, d10, d11
+;   fadd d6, d12, d13
+;   fadd d7, d14, d15
+;   fadd d8, d16, d17
+;   fadd d9, d18, d19
+;   fadd d10, d20, d21
+;   fadd d11, d22, d23
+;   fadd d12, d24, d25
+;   fadd d13, d26, d27
+;   fadd d14, d28, d29
+;   fadd d15, d30, d31
+;   fadd d0, d0, d1
+;   fadd d1, d2, d3
+;   fadd d2, d4, d5
+;   fadd d3, d6, d7
+;   fadd d4, d8, d9
+;   fadd d5, d10, d11
+;   fadd d6, d12, d13
+;   fadd d7, d14, d15
+;   fadd d0, d0, d1
+;   fadd d1, d2, d3
+;   fadd d2, d4, d5
+;   fadd d3, d6, d7
+;   fadd d0, d0, d1
+;   fadd d1, d2, d3
+;   fadd d0, d0, d1
 ;   ldp d8, d9, [sp], #16
 ;   ldp d10, d11, [sp], #16
 ;   ldp d12, d13, [sp], #16
--- a/cranelift/filetests/filetests/runtests/dynamic-simd-arithmetic.clif
+++ b/cranelift/filetests/filetests/runtests/dynamic-simd-arithmetic.clif
@@ -195,3 +195,159 @@ block0(v0: f64, v1: f64):
  return v5
 }
 ; run: %f64x2_splat_sub(0x1.0, 0x3.0) == [-0x2.0 -0x2.0]
+
+function %f32x4_splat_mul(f32, f32) -> f32x4 {
+  gv0 = dyn_scale_target_const.f32x4
+  dt0 = f32x4*gv0
+
+block0(v0: f32, v1: f32):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmul v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f32x4_splat_mul(0x2.0, 0x3.0) == [0x6.0 0x6.0 0x6.0 0x6.0]
+
+function %f64x2_splat_mul(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmul v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f64x2_splat_sub(-0x2.0, 0x3.0) == [-0x6.0 -0x6.0]
+
+function %f32x4_splat_div(f32, f32) -> f32x4 {
+  gv0 = dyn_scale_target_const.f32x4
+  dt0 = f32x4*gv0
+
+block0(v0: f32, v1: f32):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fdiv v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f32x4_splat_div(0x6.6, 0x2.2) == [0x3.0 0x3.0 0x3.0 0x3.0]
+
+function %f64x2_splat_div(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fdiv v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f64x2_splat_div(-0x6.6, 0x2.2) == [-0x3.0 -0x3.0]
+
+function %f32x4_splat_min(f32, f32) -> f32x4 {
+  gv0 = dyn_scale_target_const.f32x4
+  dt0 = f32x4*gv0
+
+block0(v0: f32, v1: f32):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmin v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f32x4_splat_min(0x6.6, 0x2.2) == [0x2.2 0x2.2 0x2.2 0x2.2]
+
+function %f64x2_splat_min(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmin v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f64x2_splat_min(-0x6.6, 0x2.2) == [-0x6.6 -0x6.6]
+
+function %f32x4_splat_max(f32, f32) -> f32x4 {
+  gv0 = dyn_scale_target_const.f32x4
+  dt0 = f32x4*gv0
+
+block0(v0: f32, v1: f32):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmax v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f32x4_splat_max(0x6.6, 0x2.2) == [0x6.6 0x6.6 0x6.6 0x6.6]
+
+function %f64x2_splat_max(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmax v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f64x2_splat_max(-0x6.6, 0x2.2) == [0x2.2 0x2.2]
+
+function %f32x4_splat_min_pseudo(f32, f32) -> f32x4 {
+  gv0 = dyn_scale_target_const.f32x4
+  dt0 = f32x4*gv0
+
+block0(v0: f32, v1: f32):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmin_pseudo v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f32x4_splat_min_pseudo(0x6.6, 0x2.2) == [0x2.2 0x2.2 0x2.2 0x2.2]
+
+function %f64x2_splat_min_pseudo(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmin_pseudo v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f64x2_splat_min_pseudo(-0x6.6, 0x2.2) == [-0x6.6 -0x6.6]
+
+function %f32x4_splat_max_pseudo(f32, f32) -> f32x4 {
+  gv0 = dyn_scale_target_const.f32x4
+  dt0 = f32x4*gv0
+
+block0(v0: f32, v1: f32):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmax_pseudo v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f32x4_splat_max_pseudo(0x6.6, 0x2.2) == [0x6.6 0x6.6 0x6.6 0x6.6]
+
+function %f64x2_splat_max_pseudo(f64, f64) -> f64x2 {
+  gv0 = dyn_scale_target_const.f64x2
+  dt0 = f64x2*gv0
+
+block0(v0: f64, v1: f64):
+  v2 = splat.dt0 v0
+  v3 = splat.dt0 v1
+  v4 = fmax_pseudo v2, v3
+  v5 = extract_vector v4, 0
+  return v5
+}
+; run: %f64x2_splat_max_pseudo(-0x6.6, 0x2.2) == [0x2.2 0x2.2]