x64: Improve codegen for splats

This commit goes through the lowerings for the CLIF `splat` instruction
and improves the support for each operator. Many of these lowerings are
mirrored from v8/SpiderMonkey and there are a number of improvements:

* AVX2 `v{p,}broadcast*` instructions are added and used when available.
* Float-based splats are much simpler and always a single-instruction
* Integer-based splats don't insert into an uninit xmm value and instead
  start out with a `movd` to move into an `xmm` register. This
  thoeretically breaks dependencies with prior instructions since `movd`
  creates a fresh new value in the destination register.
* Loads are now sunk into all of the instructions. A new extractor,
  `sinkable_load_exact`, was added to sink the i8/i16 loads.

cranelift/codegen/src/isa/x64/inst.isle

@@ -919,6 +919,7 @@
             Pshuflw
             Pshufhw
             Pblendw
+            Movddup
           ))
 
 (type CmpOpcode extern
@@ -1292,6 +1293,11 @@
             Vpextrd
             Vpextrq
             Vpblendw
+            Vmovddup
+            Vpbroadcastb
+            Vpbroadcastw
+            Vpbroadcastd
+            Vbroadcastss
           ))
 
 (type Avx512Opcode extern
@@ -1622,6 +1628,9 @@
 (decl pure has_avx () bool)
 (extern constructor has_avx has_avx)
 
+(decl pure has_avx2 () bool)
+(extern constructor has_avx2 has_avx2)
+
 ;;;; Helpers for Merging and Sinking Immediates/Loads  ;;;;;;;;;;;;;;;;;;;;;;;;;
 
 ;; Extract a constant `Imm8Reg.Imm8` from a value operand.
@@ -1656,9 +1665,21 @@
 
 ;; Extract a `SinkableLoad` that works with `RegMemImm.Mem` from a value
 ;; operand.
+;;
+;; Note that this will only work for 32-bit-types-or-larger since this is
+;; pervasively used with operations that load a minimum of 32-bits. For
+;; instructions which load exactly the type width necessary use
+;; `sinkable_load_exact`.
 (decl sinkable_load (SinkableLoad) Value)
 (extern extractor sinkable_load sinkable_load)
 
+;; Same as `sinkable_load` except that all type widths of loads are supported.
+;;
+;; Only use this when the instruction which performs the load is guaranteed to
+;; load the precisely correct size.
+(decl sinkable_load_exact (SinkableLoad) Value)
+(extern extractor sinkable_load_exact sinkable_load_exact)
+
 ;; Sink a `SinkableLoad` into a `SyntheticAmode`.
 ;;
 ;; This is a side-effectful operation that notifies the context that the
@@ -1678,6 +1699,9 @@
 (decl sink_load_to_reg_mem (SinkableLoad) RegMem)
 (rule (sink_load_to_reg_mem load) (RegMem.Mem load))
 
+(decl sink_load_to_gpr_mem (SinkableLoad) GprMem)
+(rule (sink_load_to_gpr_mem load) (RegMem.Mem load))
+
 (decl sink_load_to_reg_mem_imm (SinkableLoad) RegMemImm)
 (rule (sink_load_to_reg_mem_imm load) (RegMemImm.Mem load))
 
@@ -4103,6 +4127,34 @@
 (rule (trap_if_fcmp (FcmpCondResult.OrCondition producer cc1 cc2) tc)
       (with_flags_side_effect producer (trap_if_or cc1 cc2 tc)))
 
+;; Helper for creating `movddup` instructions
+(decl x64_movddup (XmmMem) Xmm)
+(rule (x64_movddup src)
+      (xmm_unary_rm_r_unaligned (SseOpcode.Movddup) src))
+(rule 1 (x64_movddup src)
+        (if-let $true (has_avx))
+        (xmm_unary_rm_r_vex (AvxOpcode.Vmovddup) src))
+
+;; Helper for creating `vpbroadcastb` instructions
+(decl x64_vpbroadcastb (XmmMem) Xmm)
+(rule (x64_vpbroadcastb src)
+      (xmm_unary_rm_r_vex (AvxOpcode.Vpbroadcastb) src))
+
+;; Helper for creating `vpbroadcastw` instructions
+(decl x64_vpbroadcastw (XmmMem) Xmm)
+(rule (x64_vpbroadcastw src)
+      (xmm_unary_rm_r_vex (AvxOpcode.Vpbroadcastw) src))
+
+;; Helper for creating `vpbroadcastd` instructions
+(decl x64_vpbroadcastd (XmmMem) Xmm)
+(rule (x64_vpbroadcastd src)
+      (xmm_unary_rm_r_vex (AvxOpcode.Vpbroadcastd) src))
+
+;; Helper for creating `vbroadcastss` instructions
+(decl x64_vbroadcastss (XmmMem) Xmm)
+(rule (x64_vbroadcastss src)
+      (xmm_unary_rm_r_vex (AvxOpcode.Vbroadcastss) src))
+
 ;;;; Jumps ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
 ;; Unconditional jump.
@@ -4664,6 +4716,7 @@
 (convert AtomicRmwOp MachAtomicRmwOp atomic_rmw_op_to_mach_atomic_rmw_op)
 
 (convert SinkableLoad RegMem sink_load_to_reg_mem)
+(convert SinkableLoad GprMem sink_load_to_gpr_mem)
 (convert SinkableLoad RegMemImm sink_load_to_reg_mem_imm)
 (convert SinkableLoad GprMemImm sink_load_to_gpr_mem_imm)
 (convert SinkableLoad XmmMem sink_load_to_xmm_mem)

cranelift/codegen/src/isa/x64/inst/args.rs

@@ -928,6 +928,7 @@ pub(crate) enum InstructionSet {
     BMI2,
     FMA,
     AVX,
+    AVX2,
     AVX512BITALG,
     AVX512DQ,
     AVX512F,
@@ -1126,6 +1127,7 @@ pub enum SseOpcode {
     Pshuflw,
     Pshufhw,
     Pblendw,
+    Movddup,
 }
 
 impl SseOpcode {
@@ -1280,7 +1282,8 @@ impl SseOpcode {
             | SseOpcode::Pmulhrsw
             | SseOpcode::Pshufb
             | SseOpcode::Phaddw
-            | SseOpcode::Phaddd => SSSE3,
+            | SseOpcode::Phaddd
+            | SseOpcode::Movddup => SSSE3,
 
             SseOpcode::Blendvpd
             | SseOpcode::Blendvps
@@ -1524,6 +1527,7 @@ impl fmt::Debug for SseOpcode {
             SseOpcode::Pshuflw => "pshuflw",
             SseOpcode::Pshufhw => "pshufhw",
             SseOpcode::Pblendw => "pblendw",
+            SseOpcode::Movddup => "movddup",
         };
         write!(fmt, "{}", name)
     }
@@ -1709,9 +1713,15 @@ impl AvxOpcode {
             | AvxOpcode::Vpextrw
             | AvxOpcode::Vpextrd
             | AvxOpcode::Vpextrq
-            | AvxOpcode::Vpblendw => {
+            | AvxOpcode::Vpblendw
+            | AvxOpcode::Vmovddup
+            | AvxOpcode::Vbroadcastss => {
                 smallvec![InstructionSet::AVX]
             }
+
+            AvxOpcode::Vpbroadcastb | AvxOpcode::Vpbroadcastw | AvxOpcode::Vpbroadcastd => {
+                smallvec![InstructionSet::AVX2]
+            }
         }
     }
 }

cranelift/codegen/src/isa/x64/inst/emit.rs

@@ -122,6 +122,7 @@ pub(crate) fn emit(
             InstructionSet::BMI2 => info.isa_flags.has_bmi2(),
             InstructionSet::FMA => info.isa_flags.has_fma(),
             InstructionSet::AVX => info.isa_flags.has_avx(),
+            InstructionSet::AVX2 => info.isa_flags.has_avx2(),
             InstructionSet::AVX512BITALG => info.isa_flags.has_avx512bitalg(),
             InstructionSet::AVX512DQ => info.isa_flags.has_avx512dq(),
             InstructionSet::AVX512F => info.isa_flags.has_avx512f(),
@@ -1826,6 +1827,7 @@ pub(crate) fn emit(
                 SseOpcode::Sqrtpd => (LegacyPrefixes::_66, 0x0F51, 2),
                 SseOpcode::Sqrtss => (LegacyPrefixes::_F3, 0x0F51, 2),
                 SseOpcode::Sqrtsd => (LegacyPrefixes::_F2, 0x0F51, 2),
+                SseOpcode::Movddup => (LegacyPrefixes::_F2, 0x0F12, 2),
                 _ => unimplemented!("Opcode {:?} not implemented", op),
             };
 
@@ -2450,6 +2452,13 @@ pub(crate) fn emit(
                     RegisterOrAmode::Amode(_) => (LegacyPrefixes::_F2, OpcodeMap::_0F, 0x10),
                     _ => unreachable!(),
                 },
+
+                AvxOpcode::Vpbroadcastb => (LegacyPrefixes::_66, OpcodeMap::_0F38, 0x78),
+                AvxOpcode::Vpbroadcastw => (LegacyPrefixes::_66, OpcodeMap::_0F38, 0x79),
+                AvxOpcode::Vpbroadcastd => (LegacyPrefixes::_66, OpcodeMap::_0F38, 0x58),
+                AvxOpcode::Vbroadcastss => (LegacyPrefixes::_66, OpcodeMap::_0F38, 0x18),
+                AvxOpcode::Vmovddup => (LegacyPrefixes::_F2, OpcodeMap::_0F, 0x12),
+
                 _ => panic!("unexpected rmr_imm_vex opcode {op:?}"),
             };
 

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

@@ -3915,47 +3915,89 @@
 
 ;; Rules for `splat` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-(rule (lower (has_type (multi_lane 8 16) (splat src)))
-      (let ((vec Xmm (vec_insert_lane $I8X16 (xmm_uninit_value) src 0))
-            (zeros Xmm (xmm_zero $I8X16)))
-        ;; Shuffle the lowest byte lane to all other lanes.
-        (x64_pshufb vec zeros)))
-
-(rule (lower (has_type (multi_lane 16 8) (splat src)))
-      (let (;; Force the input into a register so that we don't create a
-            ;; VCodeConstant.
-            (src RegMem (RegMem.Reg src))
-            (vec Xmm (vec_insert_lane $I16X8 (xmm_uninit_value) src 0))
-            (vec Xmm (vec_insert_lane $I16X8 vec src 1)))
-        ;; Shuffle the lowest two lanes to all other lanes.
-        (x64_pshufd vec 0)))
-
-(rule 1 (lower (has_type (multi_lane 32 4) (splat src @ (value_type (ty_scalar_float _)))))
-      (lower_splat_32x4 $F32X4 src))
-
-(rule (lower (has_type (multi_lane 32 4) (splat src)))
-      (lower_splat_32x4 $I32X4 src))
-
-(decl lower_splat_32x4 (Type Value) Xmm)
-(rule (lower_splat_32x4 ty src)
-      (let ((src RegMem src)
-            (vec Xmm (vec_insert_lane ty (xmm_uninit_value) src 0)))
-        ;; Shuffle the lowest lane to all other lanes.
-        (x64_pshufd vec 0)))
-
-(rule 1 (lower (has_type (multi_lane 64 2) (splat src @ (value_type (ty_scalar_float _)))))
-      (lower_splat_64x2 $F64X2 src))
-
-(rule (lower (has_type (multi_lane 64 2) (splat src)))
-      (lower_splat_64x2 $I64X2 src))
-
-(decl lower_splat_64x2 (Type Value) Xmm)
-(rule (lower_splat_64x2 ty src)
-      (let (;; Force the input into a register so that we don't create a
-            ;; VCodeConstant.
-            (src RegMem (RegMem.Reg src))
-            (vec Xmm (vec_insert_lane ty (xmm_uninit_value) src 0)))
-        (vec_insert_lane ty vec src 1)))
+;; For all the splat rules below one of the goals is that splatting a value
+;; doesn't end up accidentally depending on the previous value in a register.
+;; This means that instructions are chosen to avoid false dependencies where
+;; new values are created fresh or otherwise overwrite previous register
+;; contents where possible.
+;;
+;; Additionally splats are specialized to special-case load-and-splat which
+;; has a number of micro-optimizations available.
+
+;; i8x16 splats: use `vpbroadcastb` on AVX2 and otherwise `pshufb` broadcasts
+;; with a mask of zero which is calculated with an xor-against-itself register.
+(rule 0 (lower (has_type $I8X16 (splat src)))
+        (x64_pshufb (bitcast_gpr_to_xmm $I32 src) (xmm_zero $I8X16)))
+(rule 1 (lower (has_type $I8X16 (splat src)))
+        (if-let $true (has_avx2))
+        (x64_vpbroadcastb (bitcast_gpr_to_xmm $I32 src)))
+(rule 2 (lower (has_type $I8X16 (splat (sinkable_load_exact addr))))
+        (x64_pshufb (x64_pinsrb (xmm_uninit_value) addr 0) (xmm_zero $I8X16)))
+(rule 3 (lower (has_type $I8X16 (splat (sinkable_load_exact addr))))
+        (if-let $true (has_avx2))
+        (x64_vpbroadcastb addr))
+
+;; i16x8 splats: use `vpbroadcastw` on AVX2 and otherwise a 16-bit value is
+;; loaded into an xmm register, `pshuflw` broadcasts the low 16-bit lane
+;; to the low four lanes, and `pshufd` broadcasts the low 32-bit lane (which
+;; at that point is two of the 16-bit values we want to broadcast) to all the
+;; lanes.
+(rule 0 (lower (has_type $I16X8 (splat src)))
+        (x64_pshufd (x64_pshuflw (bitcast_gpr_to_xmm $I32 src) 0) 0))
+(rule 1 (lower (has_type $I16X8 (splat src)))
+        (if-let $true (has_avx2))
+        (x64_vpbroadcastw (bitcast_gpr_to_xmm $I32 src)))
+(rule 2 (lower (has_type $I16X8 (splat (sinkable_load_exact addr))))
+        (x64_pshufd (x64_pshuflw (x64_pinsrw (xmm_uninit_value) addr 0) 0) 0))
+(rule 3 (lower (has_type $I16X8 (splat (sinkable_load_exact addr))))
+        (if-let $true (has_avx2))
+        (x64_vpbroadcastw addr))
+
+;; i32x4.splat - use `vpbroadcastd` on AVX2 and otherwise `pshufd` can be
+;; used to broadcast the low lane to all other lanes.
+;;
+;; Note that sinkable-load cases come later
+(rule 0 (lower (has_type $I32X4 (splat src)))
+        (x64_pshufd (bitcast_gpr_to_xmm $I32 src) 0))
+(rule 1 (lower (has_type $I32X4 (splat src)))
+        (if-let $true (has_avx2))
+        (x64_vpbroadcastd (bitcast_gpr_to_xmm $I32 src)))
+
+;; f32x4.splat - the source is already in an xmm register so `shufps` is all
+;; that's necessary to complete the splat. This is specialized to `vbroadcastss`
+;; on AVX2 to leverage that specific instruction for this operation.
+(rule 0 (lower (has_type $F32X4 (splat src)))
+        (let ((tmp Xmm src))
+          (x64_shufps src src 0)))
+(rule 1 (lower (has_type $F32X4 (splat src)))
+        (if-let $true (has_avx2))
+        (x64_vbroadcastss src))
+
+;; t32x4.splat of a load - use a `movss` to load into an xmm register and then
+;; `shufps` broadcasts to the other lanes. Note that this is used for both i32
+;; and f32 splats.
+;;
+;; With AVX the `vbroadcastss` instruction suits this purpose precisely. Note
+;; that the memory-operand encoding of `vbroadcastss` is usable with AVX, but
+;; the register-based encoding is only available with AVX2. With the
+;; `sinkable_load` extractor this should be guaranteed to use the memory-based
+;; encoding hence the `has_avx` test.
+(rule 4 (lower (has_type (multi_lane 32 4) (splat (sinkable_load addr))))
+        (let ((tmp Xmm (x64_movss_load addr)))
+          (x64_shufps tmp tmp 0)))
+(rule 5 (lower (has_type (multi_lane 32 4) (splat (sinkable_load addr))))
+        (if-let $true (has_avx))
+        (x64_vbroadcastss addr))
+
+;; t64x2.splat - use `movddup` which is exactly what we want and there's a
+;; minor specialization for sinkable loads to avoid going through a gpr for i64
+;; splats
+(rule 0 (lower (has_type $I64X2 (splat src)))
+        (x64_movddup (bitcast_gpr_to_xmm $I64 src)))
+(rule 0 (lower (has_type $F64X2 (splat src)))
+        (x64_movddup src))
+(rule 5 (lower (has_type (multi_lane 64 2) (splat (sinkable_load addr))))
+        (x64_movddup addr))
 
 ;; Rules for `vany_true` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 

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

@@ -70,26 +70,42 @@ fn put_input_in_reg(ctx: &mut Lower<Inst>, spec: InsnInput) -> Reg {
         .expect("Multi-register value not expected")
 }
 
+enum MergeableLoadSize {
+    /// The load size performed by a sinkable load merging operation is
+    /// precisely the size necessary for the type in question.
+    Exact,
+
+    /// Narrower-than-32-bit values are handled by ALU insts that are at least
+    /// 32 bits wide, which is normally OK as we ignore upper buts; but, if we
+    /// generate, e.g., a direct-from-memory 32-bit add for a byte value and
+    /// the byte is the last byte in a page, the extra data that we load is
+    /// incorrectly accessed. So we only allow loads to merge for
+    /// 32-bit-and-above widths.
+    Min32,
+}
+
 /// Determines whether a load operation (indicated by `src_insn`) can be merged
 /// into the current lowering point. If so, returns the address-base source (as
 /// an `InsnInput`) and an offset from that address from which to perform the
 /// load.
-fn is_mergeable_load(ctx: &mut Lower<Inst>, src_insn: IRInst) -> Option<(InsnInput, i32)> {
+fn is_mergeable_load(
+    ctx: &mut Lower<Inst>,
+    src_insn: IRInst,
+    size: MergeableLoadSize,
+) -> Option<(InsnInput, i32)> {
     let insn_data = ctx.data(src_insn);
     let inputs = ctx.num_inputs(src_insn);
     if inputs != 1 {
         return None;
     }
 
+    // If this type is too small to get a merged load, don't merge the load.
     let load_ty = ctx.output_ty(src_insn, 0);
     if ty_bits(load_ty) < 32 {
-        // Narrower values are handled by ALU insts that are at least 32 bits
-        // wide, which is normally OK as we ignore upper buts; but, if we
-        // generate, e.g., a direct-from-memory 32-bit add for a byte value and
-        // the byte is the last byte in a page, the extra data that we load is
-        // incorrectly accessed. So we only allow loads to merge for
-        // 32-bit-and-above widths.
-        return None;
+        match size {
+            MergeableLoadSize::Exact => {}
+            MergeableLoadSize::Min32 => return None,
+        }
     }
 
     // Just testing the opcode is enough, because the width will always match if