aarch64: Add specialized shuffle lowerings

cranelift/codegen/src/isa/aarch64/inst.isle

@@ -553,7 +553,8 @@
        (VecDupFromFpu
         (rd WritableReg)
         (rn Reg)
-        (size VectorSize))
+        (size VectorSize)
+        (lane u8))
 
        ;; Duplicate FP immediate to vector.
        (VecDupFPImm
@@ -1390,8 +1391,18 @@
     (Addp)
     ;; Zip vectors (primary) [meaning, high halves]
     (Zip1)
+    ;; Zip vectors (secondary)
+    (Zip2)
     ;; Signed saturating rounding doubling multiply returning high half
     (Sqrdmulh)
+    ;; Unzip vectors (primary)
+    (Uzp1)
+    ;; Unzip vectors (secondary)
+    (Uzp2)
+    ;; Transpose vectors (primary)
+    (Trn1)
+    ;; Transpose vectors (secondary)
+    (Trn2)
 ))
 
 ;; A Vector ALU operation which modifies a source register.
@@ -1420,6 +1431,10 @@
     (Fneg)
     ;; Floating-point square root
     (Fsqrt)
+    ;; Reverse elements in 16-bit lanes
+    (Rev16)
+    ;; Reverse elements in 32-bit lanes
+    (Rev32)
     ;; Reverse elements in 64-bit doublewords
     (Rev64)
     ;; Floating-point convert to signed integer, rounding toward zero
@@ -1887,10 +1902,10 @@
         dst))
 
 ;; Helper for emitting `MInst.VecDupFromFpu` instructions.
-(decl vec_dup_from_fpu (Reg VectorSize) Reg)
-(rule (vec_dup_from_fpu src size)
+(decl vec_dup_from_fpu (Reg VectorSize u8) Reg)
+(rule (vec_dup_from_fpu src size lane)
       (let ((dst WritableReg (temp_writable_reg $I8X16))
-            (_ Unit (emit (MInst.VecDupFromFpu dst src size))))
+            (_ Unit (emit (MInst.VecDupFromFpu dst src size lane))))
         dst))
 
 ;; Helper for emitting `MInst.AluRRImm12` instructions.
@@ -2386,6 +2401,14 @@
 (decl neg (Reg VectorSize) Reg)
 (rule (neg x size) (vec_misc (VecMisc2.Neg) x size))
 
+;; Helper for generating `rev16` instructions.
+(decl rev16 (Reg VectorSize) Reg)
+(rule (rev16 x size) (vec_misc (VecMisc2.Rev16) x size))
+
+;; Helper for generating `rev32` instructions.
+(decl rev32 (Reg VectorSize) Reg)
+(rule (rev32 x size) (vec_misc (VecMisc2.Rev32) x size))
+
 ;; Helper for generating `rev64` instructions.
 (decl rev64 (Reg VectorSize) Reg)
 (rule (rev64 x size) (vec_misc (VecMisc2.Rev64) x size))
@@ -3767,3 +3790,27 @@
        (emit_side_effect (with_flags_side_effect
             (cmp (OperandSize.Size32) ridx jt_size)
             (jt_sequence ridx jt_info)))))
+
+;; Helper for emitting the `uzp1` instruction
+(decl vec_uzp1 (Reg Reg VectorSize) Reg)
+(rule (vec_uzp1 rn rm size) (vec_rrr (VecALUOp.Uzp1) rn rm size))
+
+;; Helper for emitting the `uzp2` instruction
+(decl vec_uzp2 (Reg Reg VectorSize) Reg)
+(rule (vec_uzp2 rn rm size) (vec_rrr (VecALUOp.Uzp2) rn rm size))
+
+;; Helper for emitting the `zip1` instruction
+(decl vec_zip1 (Reg Reg VectorSize) Reg)
+(rule (vec_zip1 rn rm size) (vec_rrr (VecALUOp.Zip1) rn rm size))
+
+;; Helper for emitting the `zip2` instruction
+(decl vec_zip2 (Reg Reg VectorSize) Reg)
+(rule (vec_zip2 rn rm size) (vec_rrr (VecALUOp.Zip2) rn rm size))
+
+;; Helper for emitting the `trn1` instruction
+(decl vec_trn1 (Reg Reg VectorSize) Reg)
+(rule (vec_trn1 rn rm size) (vec_rrr (VecALUOp.Trn1) rn rm size))
+
+;; Helper for emitting the `trn2` instruction
+(decl vec_trn2 (Reg Reg VectorSize) Reg)
+(rule (vec_trn2 rn rm size) (vec_rrr (VecALUOp.Trn2) rn rm size))

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

@@ -1977,8 +1977,20 @@ impl MachInstEmit for Inst {
                         );
                         (0b1, 0b11111, enc_size)
                     }
+                    VecMisc2::Rev16 => {
+                        debug_assert_eq!(size, VectorSize::Size8x16);
+                        (0b0, 0b00001, enc_size)
+                    }
+                    VecMisc2::Rev32 => {
+                        debug_assert!(size == VectorSize::Size8x16 || size == VectorSize::Size16x8);
+                        (0b1, 0b00000, enc_size)
+                    }
                     VecMisc2::Rev64 => {
-                        debug_assert_ne!(VectorSize::Size64x2, size);
+                        debug_assert!(
+                            size == VectorSize::Size8x16
+                                || size == VectorSize::Size16x8
+                                || size == VectorSize::Size32x4
+                        );
                         (0b0, 0b00000, enc_size)
                     }
                     VecMisc2::Fcvtzs => {
@@ -2493,13 +2505,27 @@ impl MachInstEmit for Inst {
                         | machreg_to_vec(rd.to_reg()),
                 );
             }
-            &Inst::VecDupFromFpu { rd, rn, size } => {
+            &Inst::VecDupFromFpu { rd, rn, size, lane } => {
                 let rd = allocs.next_writable(rd);
                 let rn = allocs.next(rn);
                 let q = size.is_128bits() as u32;
                 let imm5 = match size.lane_size() {
-                    ScalarSize::Size32 => 0b00100,
-                    ScalarSize::Size64 => 0b01000,
+                    ScalarSize::Size8 => {
+                        assert!(lane < 16);
+                        0b00001 | (u32::from(lane) << 1)
+                    }
+                    ScalarSize::Size16 => {
+                        assert!(lane < 8);
+                        0b00010 | (u32::from(lane) << 2)
+                    }
+                    ScalarSize::Size32 => {
+                        assert!(lane < 4);
+                        0b00100 | (u32::from(lane) << 3)
+                    }
+                    ScalarSize::Size64 => {
+                        assert!(lane < 2);
+                        0b01000 | (u32::from(lane) << 4)
+                    }
                     _ => unimplemented!(),
                 };
                 sink.put4(
@@ -2870,6 +2896,7 @@ impl MachInstEmit for Inst {
                     VecALUOp::Fmul => (0b001_01110_00_1, 0b110111),
                     VecALUOp::Addp => (0b000_01110_00_1 | enc_size << 1, 0b101111),
                     VecALUOp::Zip1 => (0b01001110_00_0 | enc_size << 1, 0b001110),
+                    VecALUOp::Zip2 => (0b01001110_00_0 | enc_size << 1, 0b011110),
                     VecALUOp::Sqrdmulh => {
                         debug_assert!(
                             size.lane_size() == ScalarSize::Size16
@@ -2878,6 +2905,10 @@ impl MachInstEmit for Inst {
 
                         (0b001_01110_00_1 | enc_size << 1, 0b101101)
                     }
+                    VecALUOp::Uzp1 => (0b01001110_00_0 | enc_size << 1, 0b000110),
+                    VecALUOp::Uzp2 => (0b01001110_00_0 | enc_size << 1, 0b010110),
+                    VecALUOp::Trn1 => (0b01001110_00_0 | enc_size << 1, 0b001010),
+                    VecALUOp::Trn2 => (0b01001110_00_0 | enc_size << 1, 0b011010),
                 };
                 let top11 = if is_float {
                     top11 | size.enc_float_size() << 1

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

@@ -2657,6 +2657,7 @@ fn test_aarch64_binemit() {
             rd: writable_vreg(14),
             rn: vreg(19),
             size: VectorSize::Size32x4,
+            lane: 0,
         },
         "6E06044E",
         "dup v14.4s, v19.s[0]",
@@ -2666,6 +2667,7 @@ fn test_aarch64_binemit() {
             rd: writable_vreg(18),
             rn: vreg(10),
             size: VectorSize::Size64x2,
+            lane: 0,
         },
         "5205084E",
         "dup v18.2d, v10.d[0]",

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

@@ -2123,9 +2123,9 @@ impl Inst {
                 let rn = pretty_print_ireg(rn, size.operand_size(), allocs);
                 format!("dup {}, {}", rd, rn)
             }
-            &Inst::VecDupFromFpu { rd, rn, size } => {
+            &Inst::VecDupFromFpu { rd, rn, size, lane } => {
                 let rd = pretty_print_vreg_vector(rd.to_reg(), size, allocs);
-                let rn = pretty_print_vreg_element(rn, 0, size.lane_size(), allocs);
+                let rn = pretty_print_vreg_element(rn, lane.into(), size.lane_size(), allocs);
                 format!("dup {}, {}", rd, rn)
             }
             &Inst::VecDupFPImm { rd, imm, size } => {
@@ -2345,7 +2345,12 @@ impl Inst {
                     VecALUOp::Fmul => ("fmul", size),
                     VecALUOp::Addp => ("addp", size),
                     VecALUOp::Zip1 => ("zip1", size),
+                    VecALUOp::Zip2 => ("zip2", size),
                     VecALUOp::Sqrdmulh => ("sqrdmulh", size),
+                    VecALUOp::Uzp1 => ("uzp1", size),
+                    VecALUOp::Uzp2 => ("uzp2", size),
+                    VecALUOp::Trn1 => ("trn1", size),
+                    VecALUOp::Trn2 => ("trn2", size),
                 };
                 let rd = pretty_print_vreg_vector(rd.to_reg(), size, allocs);
                 let rn = pretty_print_vreg_vector(rn, size, allocs);
@@ -2471,6 +2476,8 @@ impl Inst {
                     VecMisc2::Fabs => ("fabs", size, ""),
                     VecMisc2::Fneg => ("fneg", size, ""),
                     VecMisc2::Fsqrt => ("fsqrt", size, ""),
+                    VecMisc2::Rev16 => ("rev16", size, ""),
+                    VecMisc2::Rev32 => ("rev32", size, ""),
                     VecMisc2::Rev64 => ("rev64", size, ""),
                     VecMisc2::Fcvtzs => ("fcvtzs", size, ""),
                     VecMisc2::Fcvtzu => ("fcvtzu", size, ""),

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

@@ -118,6 +118,118 @@
 
 ;;;; Rules for `shuffle` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
+;; When a single element of one vector is broadcast to all the destination
+;; lanes then the `dup` instruction can be used for this operation. Note that
+;; for now this only matches lane selection from the first vector `a`, but
+;; if necessary in the future rules can be added to select from `b` as well.
+(rule 6 (lower (shuffle a b (shuffle_dup8_from_imm n)))
+        (vec_dup_from_fpu a (VectorSize.Size8x16) n))
+(rule 5 (lower (shuffle a b (shuffle_dup16_from_imm n)))
+        (vec_dup_from_fpu a (VectorSize.Size16x8) n))
+(rule 4 (lower (shuffle a b (shuffle_dup32_from_imm n)))
+        (vec_dup_from_fpu a (VectorSize.Size32x4) n))
+(rule 3 (lower (shuffle a b (shuffle_dup64_from_imm n)))
+        (vec_dup_from_fpu a (VectorSize.Size64x2) n))
+
+;; If the `Immediate` specified to the extractor looks like a duplication of the
+;; `n`th lane of the first vector of size K-byte lanes, then each extractor
+;; returns the `n` value as a `u8` to be used as part of a `vec_dup_from_fpu`
+;; instruction. Note that there's a different extractor for each bit-width of
+;; lane.
+(decl shuffle_dup8_from_imm (u8) Immediate)
+(extern extractor shuffle_dup8_from_imm shuffle_dup8_from_imm)
+(decl shuffle_dup16_from_imm (u8) Immediate)
+(extern extractor shuffle_dup16_from_imm shuffle_dup16_from_imm)
+(decl shuffle_dup32_from_imm (u8) Immediate)
+(extern extractor shuffle_dup32_from_imm shuffle_dup32_from_imm)
+(decl shuffle_dup64_from_imm (u8) Immediate)
+(extern extractor shuffle_dup64_from_imm shuffle_dup64_from_imm)
+
+;; When the shuffle looks like "concatenate `a` and `b` and shift right by n*8
+;; bytes", that's an `ext` instruction.
+(rule 2 (lower (shuffle a b (vec_extract_imm4_from_immediate n)))
+        (vec_extract a b n))
+
+;; Attempts to extract `n` from the specified shuffle `Immediate` where each
+;; byte of the `Immediate` is a consecutive sequence starting from `n`. This
+;; value of `n` is used as part of the `vec_extract` instruction which extracts
+;; consecutive bytes from two vectors into one final vector, offset by `n`
+;; bytes.
+(decl vec_extract_imm4_from_immediate (u8) Immediate)
+(extern extractor vec_extract_imm4_from_immediate vec_extract_imm4_from_immediate)
+
+;; Rules for the `uzp1` and `uzp2` instructions which gather even-numbered lanes
+;; or odd-numbered lanes
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1e1c_1a18_1614_1210_0e0c_0a08_0604_0200)))
+      (vec_uzp1 a b (VectorSize.Size8x16)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1f1d_1b19_1715_1311_0f0d_0b09_0705_0301)))
+      (vec_uzp2 a b (VectorSize.Size8x16)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1d1c_1918_1514_1110_0d0c_0908_0504_0100)))
+      (vec_uzp1 a b (VectorSize.Size16x8)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1f1e_1b1a_1716_1312_0f0e_0b0a_0706_0302)))
+      (vec_uzp2 a b (VectorSize.Size16x8)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1b1a1918_13121110_0b0a0908_03020100)))
+      (vec_uzp1 a b (VectorSize.Size32x4)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1f1e1d1c_17161514_0f0e0d0c_07060504)))
+      (vec_uzp2 a b (VectorSize.Size32x4)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1716151413121110_0706050403020100)))
+      (vec_uzp1 a b (VectorSize.Size64x2)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1f1e1d1c1b1a1918_0f0e0d0c0b0a0908)))
+      (vec_uzp2 a b (VectorSize.Size64x2)))
+
+;; Rules for the `zip1` and `zip2` instructions which interleave lanes in the
+;; low or high halves of the two input vectors.
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1707_1606_1505_1404_1303_1202_1101_1000)))
+      (vec_zip1 a b (VectorSize.Size8x16)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1f0f_1e0e_1d0d_1c0c_1b0b_1a0a_1909_1808)))
+      (vec_zip2 a b (VectorSize.Size8x16)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1716_0706_1514_0504_1312_0302_1110_0100)))
+      (vec_zip1 a b (VectorSize.Size16x8)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1f1e_0f0e_1d1c_0d0c_1b1a_0b0a_1918_0908)))
+      (vec_zip2 a b (VectorSize.Size16x8)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x17161514_07060504_13121110_03020100)))
+      (vec_zip1 a b (VectorSize.Size32x4)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1f1e1d1c_0f0e0d0c_1b1a1918_0b0a0908)))
+      (vec_zip2 a b (VectorSize.Size32x4)))
+;; Note that zip1/zip2 for i64x2 vectors is omitted since it's already covered
+;; by the i64x2 cases of uzp1/uzp2 above where both zip and uzp have the same
+;; semantics for 64-bit lanes.
+
+;; Rules for the `trn1` and `trn2` instructions which interleave odd or even
+;; lanes in the two input vectors.
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1e0e_1c0c_1a0a_1808_1606_1404_1202_1000)))
+      (vec_trn1 a b (VectorSize.Size8x16)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1f0f_1d0d_1b0b_1909_1707_1505_1303_1101)))
+      (vec_trn2 a b (VectorSize.Size8x16)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1d1c_0d0c_1918_0908_1514_0504_1110_0100)))
+      (vec_trn1 a b (VectorSize.Size16x8)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1f1e_0f0e_1b1a_0b0a_1716_0706_1312_0302)))
+      (vec_trn2 a b (VectorSize.Size16x8)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1b1a1918_0b0a0908_13121110_03020100)))
+      (vec_trn1 a b (VectorSize.Size32x4)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x1f1e1d1c_0f0e0d0c_17161514_07060504)))
+      (vec_trn2 a b (VectorSize.Size32x4)))
+;; Note that trn1/trn2 for i64x2 vectors is omitted since it's already covered
+;; by the i64x2 cases of uzp1/uzp2 above where both trn and uzp have the same
+;; semantics for 64-bit lanes.
+
+;; Rules for the `rev{16,32,64}` instructions where reversals happen at either
+;; the byte level, the 16-bit level, or 32-bit level. Note that all of these
+;; patterns only match reversals in the first operand, but they can
+;; theoretically be extended if necessary to reversals in the second operand.
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x0e0f_0c0d_0a0b_0809_0607_0405_0203_0001)))
+      (rev16 a (VectorSize.Size8x16)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x0c0d0e0f_08090a0b_04050607_00010203)))
+      (rev32 a (VectorSize.Size8x16)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x0d0c0f0e_09080b0a_05040706_01000302)))
+      (rev32 a (VectorSize.Size16x8)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x08090a0b0c0d0e0f_0001020304050607)))
+      (rev64 a (VectorSize.Size8x16)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x09080b0a0d0c0f0e_0100030205040706)))
+      (rev64 a (VectorSize.Size16x8)))
+(rule 1 (lower (shuffle a b (u128_from_immediate 0x0b0a09080f0e0d0c_0302010007060504)))
+      (rev64 a (VectorSize.Size32x4)))
+
 (rule (lower (has_type ty (shuffle rn rn2 (u128_from_immediate mask))))
       (let ((mask_reg Reg (constant_f128 mask)))
        (vec_tbl2 rn rn2 mask_reg ty)))
@@ -1840,7 +1952,7 @@
       (vec_dup x (vector_size ty)))
 
 (rule -2 (lower (has_type ty (splat x @ (value_type (ty_scalar_float _)))))
-      (vec_dup_from_fpu x (vector_size ty)))
+      (vec_dup_from_fpu x (vector_size ty) 0))
 
 (rule (lower (has_type ty (splat (f32const (u64_from_ieee32 n)))))
       (splat_const n (vector_size ty)))

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

@@ -742,4 +742,47 @@ impl Context for IsleContext<'_, '_, MInst, AArch64Backend> {
             );
         }
     }
+
+    fn vec_extract_imm4_from_immediate(&mut self, imm: Immediate) -> Option<u8> {
+        let bytes = self.lower_ctx.get_immediate_data(imm).as_slice();
+
+        if bytes.windows(2).all(|a| a[0] + 1 == a[1]) && bytes[0] < 16 {
+            Some(bytes[0])
+        } else {
+            None
+        }
+    }
+
+    fn shuffle_dup8_from_imm(&mut self, imm: Immediate) -> Option<u8> {
+        let bytes = self.lower_ctx.get_immediate_data(imm).as_slice();
+        if bytes.iter().all(|b| *b == bytes[0]) && bytes[0] < 16 {
+            Some(bytes[0])
+        } else {
+            None
+        }
+    }
+    fn shuffle_dup16_from_imm(&mut self, imm: Immediate) -> Option<u8> {
+        let (a, b, c, d, e, f, g, h) = self.shuffle16_from_imm(imm)?;
+        if a == b && b == c && c == d && d == e && e == f && f == g && g == h && a < 8 {
+            Some(a)
+        } else {
+            None
+        }
+    }
+    fn shuffle_dup32_from_imm(&mut self, imm: Immediate) -> Option<u8> {
+        let (a, b, c, d) = self.shuffle32_from_imm(imm)?;
+        if a == b && b == c && c == d && a < 4 {
+            Some(a)
+        } else {
+            None
+        }
+    }
+    fn shuffle_dup64_from_imm(&mut self, imm: Immediate) -> Option<u8> {
+        let (a, b) = self.shuffle64_from_imm(imm)?;
+        if a == b && a < 2 {
+            Some(a)
+        } else {
+            None
+        }
+    }
 }