aarch64: Add more lowerings for the CLIF fma (#6150)

This commit adds new lowerings to the AArch64 backend of the element-based `fmla` and `fmls` instructions. These instructions have one of the multiplicands as an implicit broadcast of a single lane of another register and can help remove `shuffle` or `dup` instructions that would otherwise be used to implement them.
2023-04-05 12:22:55 -05:00
parent bf741955f0
commit 967543eb43
8 changed files with 321 additions and 15 deletions
--- a/cranelift/codegen/src/isa/aarch64/inst.isle
+++ b/cranelift/codegen/src/isa/aarch64/inst.isle
@@ -651,6 +651,16 @@
        (rm Reg)
        (size VectorSize))

+       ;; A vector ALU op modifying a source register.
+       (VecFmlaElem
+        (alu_op VecALUModOp)
+        (rd WritableReg)
+        (ri Reg)
+        (rn Reg)
+        (rm Reg)
+        (size VectorSize)
+        (idx u8))
+
       ;; Vector two register miscellaneous instruction.
       (VecMisc
        (op VecMisc2)
@@ -1850,7 +1860,7 @@
            (_ Unit (emit (MInst.FpuRR op size dst src))))
       dst))

-;; Helper for emitting `MInst.VecRRR` instructions which use three registers,
+;; Helper for emitting `MInst.VecRRRMod` instructions which use three registers,
 ;; one of which is both source and output.
 (decl vec_rrr_mod (VecALUModOp Reg Reg Reg VectorSize) Reg)
 (rule (vec_rrr_mod op src1 src2 src3 size)
@@ -1858,6 +1868,14 @@
            (_1 Unit (emit (MInst.VecRRRMod op dst src1 src2 src3 size))))
        dst))

+;; Helper for emitting `MInst.VecFmlaElem` instructions which use three registers,
+;; one of which is both source and output.
+(decl vec_fmla_elem (VecALUModOp Reg Reg Reg VectorSize u8) Reg)
+(rule (vec_fmla_elem op src1 src2 src3 size idx)
+      (let ((dst WritableReg (temp_writable_reg $I8X16))
+            (_1 Unit (emit (MInst.VecFmlaElem op dst src1 src2 src3 size idx))))
+        dst))
+
 (decl fpu_rri (FPUOpRI Reg) Reg)
 (rule (fpu_rri op src)
      (let ((dst WritableReg (temp_writable_reg $F64))
--- a/cranelift/codegen/src/isa/aarch64/inst/emit.rs
+++ b/cranelift/codegen/src/isa/aarch64/inst/emit.rs
@@ -2914,6 +2914,45 @@ impl MachInstEmit for Inst {
                };
                sink.put4(enc_vec_rrr(top11 | q << 9, rm, bit15_10, rn, rd));
            }
+            &Inst::VecFmlaElem {
+                rd,
+                ri,
+                rn,
+                rm,
+                alu_op,
+                size,
+                idx,
+            } => {
+                let rd = allocs.next_writable(rd);
+                let ri = allocs.next(ri);
+                debug_assert_eq!(rd.to_reg(), ri);
+                let rn = allocs.next(rn);
+                let rm = allocs.next(rm);
+                let idx = u32::from(idx);
+
+                let (q, _size) = size.enc_size();
+                let o2 = match alu_op {
+                    VecALUModOp::Fmla => 0b0,
+                    VecALUModOp::Fmls => 0b1,
+                    _ => unreachable!(),
+                };
+
+                let (h, l) = match size {
+                    VectorSize::Size32x4 => {
+                        assert!(idx < 4);
+                        (idx >> 1, idx & 1)
+                    }
+                    VectorSize::Size64x2 => {
+                        assert!(idx < 2);
+                        (idx, 0)
+                    }
+                    _ => unreachable!(),
+                };
+
+                let top11 = 0b000_011111_00 | (q << 9) | (size.enc_float_size() << 1) | l;
+                let bit15_10 = 0b000100 | (o2 << 4) | (h << 1);
+                sink.put4(enc_vec_rrr(top11, rm, bit15_10, rn, rd));
+            }
            &Inst::VecLoadReplicate {
                rd,
                rn,
--- a/cranelift/codegen/src/isa/aarch64/inst/mod.rs
+++ b/cranelift/codegen/src/isa/aarch64/inst/mod.rs
@@ -812,7 +812,7 @@ fn aarch64_get_operands<F: Fn(VReg) -> VReg>(inst: &Inst, collector: &mut Operan
            collector.reg_use(rn);
            collector.reg_use(rm);
        }
-        &Inst::VecRRRMod { rd, ri, rn, rm, .. } => {
+        &Inst::VecRRRMod { rd, ri, rn, rm, .. } | &Inst::VecFmlaElem { rd, ri, rn, rm, .. } => {
            collector.reg_reuse_def(rd, 1); // `rd` == `ri`.
            collector.reg_use(ri);
            collector.reg_use(rn);
@@ -2171,6 +2171,26 @@ impl Inst {
                let rm = pretty_print_vreg_vector(rm, size, allocs);
                format!("{} {}, {}, {}, {}", op, rd, ri, rn, rm)
            }
+            &Inst::VecFmlaElem {
+                rd,
+                ri,
+                rn,
+                rm,
+                alu_op,
+                size,
+                idx,
+            } => {
+                let (op, size) = match alu_op {
+                    VecALUModOp::Fmla => ("fmla", size),
+                    VecALUModOp::Fmls => ("fmls", size),
+                    _ => unreachable!(),
+                };
+                let rd = pretty_print_vreg_vector(rd.to_reg(), size, allocs);
+                let ri = pretty_print_vreg_vector(ri, size, allocs);
+                let rn = pretty_print_vreg_vector(rn, size, allocs);
+                let rm = pretty_print_vreg_element(rm, idx.into(), size.lane_size(), allocs);
+                format!("{} {}, {}, {}, {}", op, rd, ri, rn, rm)
+            }
            &Inst::VecRRRLong {
                rd,
                rn,
--- a/cranelift/codegen/src/isa/aarch64/lower.isle
+++ b/cranelift/codegen/src/isa/aarch64/lower.isle
@@ -513,18 +513,63 @@

 ;;;; Rules for `fma` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

-(rule (lower (has_type ty @ (multi_lane _ _) (fma x y z)))
-      (vec_rrr_mod (VecALUModOp.Fmla) z x y (vector_size ty)))
-
-(rule 1 (lower (has_type ty @ (multi_lane _ _) (fma (fneg x) y z)))
-      (vec_rrr_mod (VecALUModOp.Fmls) z x y (vector_size ty)))
-
-(rule 2 (lower (has_type ty @ (multi_lane _ _) (fma x (fneg y) z)))
-      (vec_rrr_mod (VecALUModOp.Fmls) z x y (vector_size ty)))
-
-(rule 3 (lower (has_type (ty_scalar_float ty) (fma x y z)))
+(rule (lower (has_type (ty_scalar_float ty) (fma x y z)))
       (fpu_rrrr (FPUOp3.MAdd) (scalar_size ty) x y z))

+;; Delegate vector-based lowerings to helpers below
+(rule 1 (lower (has_type ty @ (multi_lane _ _) (fma x y z)))
+        (lower_fmla (VecALUModOp.Fmla) x y z (vector_size ty)))
+
+;; Lowers a fused-multiply-add operation handling various forms of the
+;; instruction to get maximal coverage of what's available on AArch64.
+(decl lower_fmla (VecALUModOp Value Value Value VectorSize) Reg)
+
+;; Base case, emit the op requested.
+(rule (lower_fmla op x y z size)
+      (vec_rrr_mod op z x y size))
+
+;; Special case: if one of the multiplicands are a splat then the element-based
+;; fma can be used instead with 0 as the element index.
+(rule 1 (lower_fmla op (splat x) y z size)
+        (vec_fmla_elem op z y x size 0))
+(rule 2 (lower_fmla op x (splat y) z size)
+        (vec_fmla_elem op z x y size 0))
+
+;; Special case: if one of the multiplicands is a shuffle to broadcast a
+;; single element of a vector then the element-based fma can be used like splat
+;; above.
+;;
+;; Note that in Cranelift shuffle always has i8x16 inputs and outputs so
+;; a `bitcast` is matched here explicitly since that's the main way a shuffle
+;; output will be fed into this instruction.
+(rule 3 (lower_fmla op (bitcast _ (shuffle x x (shuffle32_from_imm n n n n))) y z size @ (VectorSize.Size32x4))
+        (if-let $true (u64_lt n 4))
+        (vec_fmla_elem op z y x size n))
+(rule 4 (lower_fmla op x (bitcast _ (shuffle y y (shuffle32_from_imm n n n n))) z size @ (VectorSize.Size32x4))
+        (if-let $true (u64_lt n 4))
+        (vec_fmla_elem op z x y size n))
+(rule 3 (lower_fmla op (bitcast _ (shuffle x x (shuffle64_from_imm n n))) y z size @ (VectorSize.Size64x2))
+        (if-let $true (u64_lt n 2))
+        (vec_fmla_elem op z y x size n))
+(rule 4 (lower_fmla op x (bitcast _ (shuffle y y (shuffle64_from_imm n n))) z size @ (VectorSize.Size64x2))
+        (if-let $true (u64_lt n 2))
+        (vec_fmla_elem op z x y size n))
+
+;; Special case: if one of the multiplicands is `fneg` then peel that away,
+;; reverse the operation being performed, and then recurse on `lower_fmla`
+;; again to generate the actual instruction.
+;;
+;; Note that these are the highest priority cases for `lower_fmla` to peel
+;; away as many `fneg` operations as possible.
+(rule 5 (lower_fmla op (fneg x) y z size)
+        (lower_fmla (neg_fmla op) x y z size))
+(rule 6 (lower_fmla op x (fneg y) z size)
+        (lower_fmla (neg_fmla op) x y z size))
+
+(decl neg_fmla (VecALUModOp) VecALUModOp)
+(rule (neg_fmla (VecALUModOp.Fmla)) (VecALUModOp.Fmls))
+(rule (neg_fmla (VecALUModOp.Fmls)) (VecALUModOp.Fmla))
+
 ;;;; Rules for `fcopysign` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

 (rule (lower (has_type ty (fcopysign x y)))
--- a/cranelift/codegen/src/isa/riscv64/inst.isle
+++ b/cranelift/codegen/src/isa/riscv64/inst.isle
@@ -708,8 +708,6 @@
 (decl u8_as_i32 (u8) i32)
 (extern constructor u8_as_i32 u8_as_i32)

-(convert u8 u64 u8_as_u64)
-
 (decl convert_valueregs_reg (ValueRegs) Reg)
 (rule (convert_valueregs_reg x)
  (value_regs_get x 0))
--- a/cranelift/codegen/src/prelude.isle
+++ b/cranelift/codegen/src/prelude.isle
@@ -82,6 +82,7 @@

 (decl pure u8_as_u64 (u8) u64)
 (extern constructor u8_as_u64 u8_as_u64)
+(convert u8 u64 u8_as_u64)

 (decl pure u16_as_u64 (u16) u64)
 (extern constructor u16_as_u64 u16_as_u64)
--- a/cranelift/filetests/filetests/isa/aarch64/fma.clif
+++ b/cranelift/filetests/filetests/isa/aarch64/fma.clif
@@ -157,3 +157,152 @@ block0(v0: f64x2, v1: f64x2, v2: f64x2):
 ;   fmls v0.2d, v5.2d, v1.2d
 ;   ret

+function %f32x4_splat0(f32, f32x4, f32x4) -> f32x4 {
+block0(v0: f32, v1: f32x4, v2: f32x4):
+    v3 = splat.f32x4 v0
+    v4 = fma v3, v1, v2
+    return v4
+}
+
+; VCode:
+; block0:
+;   mov v5.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmla v0.4s, v0.4s, v1.4s, v5.s[0]
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   mov v5.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmla v0.4s, v1.4s, v5.s[0]
+;   ret
+
+function %f32x4_splat1(f32x4, f32, f32x4) -> f32x4 {
+block0(v0: f32x4, v1: f32, v2: f32x4):
+    v3 = splat.f32x4 v1
+    v4 = fneg v0
+    v5 = fma v4, v3, v2
+    return v5
+}
+
+; VCode:
+; block0:
+;   mov v5.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmls v0.4s, v0.4s, v5.4s, v1.s[0]
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   mov v5.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmls v0.4s, v5.4s, v1.s[0]
+;   ret
+
+function %f32x4_splat2(f32x4, f32x4, f32x4) -> f32x4 {
+block0(v0: f32x4, v1: f32x4, v2: f32x4):
+    v3 = bitcast.i8x16 little v0
+    v4 = shuffle v3, v3, 0x07060504_07060504_07060504_07060504
+    v5 = bitcast.f32x4 little v4
+    v6 = fma v5, v1, v2
+    return v6
+}
+
+; VCode:
+; block0:
+;   mov v5.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmla v0.4s, v0.4s, v1.4s, v5.s[1]
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   mov v5.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmla v0.4s, v1.4s, v5.s[1]
+;   ret
+
+function %f32x4_splat3(f32x4, f32x4, f32x4) -> f32x4 {
+block0(v0: f32x4, v1: f32x4, v2: f32x4):
+    v3 = bitcast.i8x16 little v1
+    v4 = shuffle v3, v3, 0x0f0e0d0c_0f0e0d0c_0f0e0d0c_0f0e0d0c
+    v5 = bitcast.f32x4 little v4
+    v6 = fneg v5
+    v7 = fma v0, v6, v2
+    return v7
+}
+
+; VCode:
+; block0:
+;   mov v5.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmls v0.4s, v0.4s, v5.4s, v1.s[3]
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   mov v5.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmls v0.4s, v5.4s, v1.s[3]
+;   ret
+
+function %f32x4_splat4(f32x4, f32x4, f32x4) -> f32x4 {
+block0(v0: f32x4, v1: f32x4, v2: f32x4):
+    v3 = bitcast.i8x16 little v1
+    v4 = shuffle v3, v3, 0x1f1e1d1c_1f1e1d1c_1f1e1d1c_1f1e1d1c
+    v5 = bitcast.f32x4 little v4
+    v6 = fma v0, v5, v2
+    return v6
+}
+
+; VCode:
+; block0:
+;   mov v31.16b, v1.16b
+;   movz w6, #7452
+;   movk w6, w6, #7966, LSL #16
+;   dup v17.4s, w6
+;   mov v30.16b, v31.16b
+;   tbl v19.16b, { v30.16b, v31.16b }, v17.16b
+;   mov v23.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmla v0.4s, v0.4s, v23.4s, v19.4s
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   mov v31.16b, v1.16b
+;   mov w6, #0x1d1c
+;   movk w6, #0x1f1e, lsl #16
+;   dup v17.4s, w6
+;   mov v30.16b, v31.16b
+;   tbl v19.16b, {v30.16b, v31.16b}, v17.16b
+;   mov v23.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmla v0.4s, v23.4s, v19.4s
+;   ret
+
+function %f64x2_splat0(f64x2, f64x2, f64x2) -> f64x2 {
+block0(v0: f64x2, v1: f64x2, v2: f64x2):
+    v3 = bitcast.i8x16 little v1
+    v4 = shuffle v3, v3, 0x0f0e0d0c0b0a0908_0f0e0d0c0b0a0908
+    v5 = bitcast.f64x2 little v4
+    v6 = fneg v5
+    v7 = fma v0, v6, v2
+    return v7
+}
+
+; VCode:
+; block0:
+;   mov v5.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmls v0.2d, v0.2d, v5.2d, v1.d[1]
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   mov v5.16b, v0.16b
+;   mov v0.16b, v2.16b
+;   fmls v0.2d, v5.2d, v1.d[1]
+;   ret
+
--- a/cranelift/filetests/filetests/runtests/simd-fma.clif
+++ b/cranelift/filetests/filetests/runtests/simd-fma.clif
@@ -87,3 +87,39 @@ block0(v0: f64x2, v1: f64x2, v2: f64x2):
 ; run: %fma_is_nan_f64x2([0x0.0 0x0.0], [+NaN 0x0.0], [0x0.0 +NaN]) == 1
 ; run: %fma_is_nan_f64x2([-NaN 0x0.0], [0x0.0 -NaN], [0x0.0 0x0.0]) == 1
 ; run: %fma_is_nan_f64x2([0x0.0 NaN], [0x0.0 NaN], [-NaN NaN]) == 1
+
+function %fma_f32x4_splat1(f32x4, f32, f32x4) -> f32x4 {
+block0(v0: f32x4, v1: f32, v2: f32x4):
+    v3 = splat.f32x4 v1
+    v4 = fma v0, v3, v2
+    return v4
+}
+; run: %fma_f32x4_splat1([0x9.0 0x9.0 0x9.0 0x9.0], 0x9.0, [0x9.0 0x9.0 0x9.0 0x9.0]) == [0x1.680000p6 0x1.680000p6 0x1.680000p6 0x1.680000p6]
+; run: %fma_f32x4_splat1([0x1.0 0x2.0 0x3.0 0x4.0], 0x0.0, [0x5.0 0x6.0 0x7.0 0x8.0]) == [0x5.0 0x6.0 0x7.0 0x8.0]
+
+function %fma_f32x4_splat2(f32, f32x4, f32x4) -> f32x4 {
+block0(v0: f32, v1: f32x4, v2: f32x4):
+    v3 = splat.f32x4 v0
+    v4 = fma v3, v1, v2
+    return v4
+}
+; run: %fma_f32x4_splat2(0x9.0, [0x9.0 0x9.0 0x9.0 0x9.0], [0x9.0 0x9.0 0x9.0 0x9.0]) == [0x1.680000p6 0x1.680000p6 0x1.680000p6 0x1.680000p6]
+; run: %fma_f32x4_splat2(0x0.0, [0x1.0 0x2.0 0x3.0 0x4.0], [0x5.0 0x6.0 0x7.0 0x8.0]) == [0x5.0 0x6.0 0x7.0 0x8.0]
+
+function %fma_f64x2_splat1(f64x2, f64, f64x2) -> f64x2 {
+block0(v0: f64x2, v1: f64, v2: f64x2):
+    v3 = splat.f64x2 v1
+    v4 = fma v0, v3, v2
+    return v4
+}
+; run: %fma_f64x2_splat1([0x9.0 0x9.0], 0x9.0, [0x9.0 0x9.0]) == [0x1.680000p6 0x1.680000p6]
+; run: %fma_f64x2_splat1([0x1.0 0x2.0], 0x0.0, [0x5.0 0x6.0]) == [0x5.0 0x6.0]
+
+function %fma_f64x2_splat2(f64, f64x2, f64x2) -> f64x2 {
+block0(v0: f64, v1: f64x2, v2: f64x2):
+    v3 = splat.f64x2 v0
+    v4 = fma v3, v1, v2
+    return v4
+}
+; run: %fma_f64x2_splat2(0x9.0, [0x9.0 0x9.0], [0x9.0 0x9.0]) == [0x1.680000p6 0x1.680000p6]
+; run: %fma_f64x2_splat2(0x0.0, [0x1.0 0x2.0], [0x5.0 0x6.0]) == [0x5.0 0x6.0]