x64: Add a smattering of lowerings for shuffle specializations (#5930)

* x64: Add lowerings for `punpck{h,l}wd`

Add some special cases for `shuffle` for more specialized x86
instructions.

* x64: Add `shuffle` lowerings for `pshufd`

This commit adds special-cased lowerings for the x64 `shuffle`
instruction when the `pshufd` instruction alone is necessary. This is
possible when the shuffle immediate permutes 32-bit values within one of
the vector inputs of the `shuffle` instruction, but not both.

* x64: Add shuffle lowerings for `punpck{h,l}{q,}dq`

This adds specific permutations for some x86 instructions which
specifically interleave high/low bytes for 32 and 64-bit values. This
corresponds to the preexisting specific lowerings for interleaving 8 and
16-bit values.

* x64: Add `shuffle` lowerings for `shufps`

This commit adds targeted lowerings for the `shuffle` instruction that
match the pattern that `shufps` supports. The `shufps` instruction
selects two elements from the first vector and two elements from the
second vector which means while it's not generally applicable it should
still be more useful than the catch-all lowering of `shuffle`.

* x64: Add shuffle support for `pshuf{l,h}w`

This commit adds special lowering cases for these instructions which
permute 16-bit values within a 128-bit value either within the upper or
lower half of the 128-bit value.

* x64: Specialize `shuffle` with an all-zeros immediate

Instead of loading the all-zeros immediate from a rip-relative address
at the end of the function instead generate a zero with a `pxor`
instruction and then use `pshufb` to do the broadcast.

* Review comments

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

@@ -3529,16 +3529,98 @@
 
 ;; Rules for `shuffle` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;
 
-;; Special case for the `punpckhbw` instruction which interleaves the upper
-;; lanes of the two input registers.
-(rule 4 (lower (shuffle a b (u128_from_immediate 0x1f0f_1e0e_1d0d_1c0c_1b0b_1a0a_1909_1808)))
-      (x64_punpckhbw a b))
+;; Special case the `pshuf{l,h}w` instruction which shuffles four 16-bit
+;; integers within one value, preserving the other four 16-bit integers in that
+;; value (either the high or low half). The complicated logic is in the
+;; extractors here implemented in Rust and note that there's two cases for each
+;; instruction here to match when either the first or second shuffle operand is
+;; used.
+(rule 12 (lower (shuffle x y (pshuflw_lhs_imm imm)))
+      (x64_pshuflw x imm))
+(rule 11 (lower (shuffle x y (pshuflw_rhs_imm imm)))
+      (x64_pshuflw y imm))
+(rule 10 (lower (shuffle x y (pshufhw_lhs_imm imm)))
+      (x64_pshufhw x imm))
+(rule 9 (lower (shuffle x y (pshufhw_rhs_imm imm)))
+      (x64_pshufhw y imm))
 
-;; Special case for the `punpcklbw` instruction which interleaves the lower
-;; lanes of the two input registers.
-(rule 4 (lower (shuffle a b (u128_from_immediate 0x1707_1606_1505_1404_1303_1202_1101_1000)))
+(decl pshuflw_lhs_imm (u8) Immediate)
+(extern extractor pshuflw_lhs_imm pshuflw_lhs_imm)
+(decl pshuflw_rhs_imm (u8) Immediate)
+(extern extractor pshuflw_rhs_imm pshuflw_rhs_imm)
+(decl pshufhw_lhs_imm (u8) Immediate)
+(extern extractor pshufhw_lhs_imm pshufhw_lhs_imm)
+(decl pshufhw_rhs_imm (u8) Immediate)
+(extern extractor pshufhw_rhs_imm pshufhw_rhs_imm)
+
+;; Special case for the `pshufd` instruction which will permute 32-bit values
+;; within a single register. This is only applicable if the `imm` specified
+;; selects 32-bit values from either `x` or `y`, but not both. This means
+;; there's one rule for selecting from `x` and another rule for selecting from
+;; `y`.
+(rule 8 (lower (shuffle x y (pshufd_lhs_imm imm)))
+      (x64_pshufd x imm))
+(rule 7 (lower (shuffle x y (pshufd_rhs_imm imm)))
+      (x64_pshufd y imm))
+
+(decl pshufd_lhs_imm (u8) Immediate)
+(extern extractor pshufd_lhs_imm pshufd_lhs_imm)
+(decl pshufd_rhs_imm (u8) Immediate)
+(extern extractor pshufd_rhs_imm pshufd_rhs_imm)
+
+;; Special case for i8-level interleaving of upper/low bytes.
+(rule 6 (lower (shuffle a b (u128_from_immediate 0x1f0f_1e0e_1d0d_1c0c_1b0b_1a0a_1909_1808)))
+      (x64_punpckhbw a b))
+(rule 6 (lower (shuffle a b (u128_from_immediate 0x1707_1606_1505_1404_1303_1202_1101_1000)))
       (x64_punpcklbw a b))
 
+;; Special case for i16-level interleaving of upper/low bytes.
+(rule 6 (lower (shuffle a b (u128_from_immediate 0x1f1e_0f0e_1d1c_0d0c_1b1a_0b0a_1918_0908)))
+      (x64_punpckhwd a b))
+(rule 6 (lower (shuffle a b (u128_from_immediate 0x1716_0706_1514_0504_1312_0302_1110_0100)))
+      (x64_punpcklwd a b))
+
+;; Special case for i32-level interleaving of upper/low bytes.
+(rule 6 (lower (shuffle a b (u128_from_immediate 0x1f1e1d1c_0f0e0d0c_1b1a1918_0b0a0908)))
+      (x64_punpckhdq a b))
+(rule 6 (lower (shuffle a b (u128_from_immediate 0x17161514_07060504_13121110_03020100)))
+      (x64_punpckldq a b))
+
+;; Special case for i64-level interleaving of upper/low bytes.
+(rule 6 (lower (shuffle a b (u128_from_immediate 0x1f1e1d1c1b1a1918_0f0e0d0c0b0a0908)))
+      (x64_punpckhqdq a b))
+(rule 6 (lower (shuffle a b (u128_from_immediate 0x1716151413121110_0706050403020100)))
+      (x64_punpcklqdq a b))
+
+;; If the vector shift mask is all 0s then that means the first byte of the
+;; first operand is broadcast to all bytes. Falling through would load an
+;; all-zeros constant from a rip-relative location but it should be slightly
+;; more efficient to execute the `pshufb` here-and-now with an xor'd-to-be-zero
+;; register.
+(rule 6 (lower (shuffle a _ (u128_from_immediate 0)))
+      (x64_pshufb a (xmm_zero $I8X16)))
+
+;; Special case for the `shufps` instruction which will select two 32-bit values
+;; from the first operand and two 32-bit values from the second operand. Note
+;; that there is a second case here as well for when the operands can be
+;; swapped.
+;;
+;; Note that the priority of this instruction is currently lower than the above
+;; special cases since `shufps` handles many of them and for now it's
+;; hypothesized that the dedicated instructions are better than `shufps`.
+;; Someone with more knowledge about x86 timings should perhaps reorder the
+;; rules here eventually though.
+(rule 5 (lower (shuffle x y (shufps_imm imm)))
+      (x64_shufps x y imm))
+(rule 4 (lower (shuffle x y (shufps_rev_imm imm)))
+      (x64_shufps y x imm))
+
+(decl shufps_imm(u8) Immediate)
+(extern extractor shufps_imm shufps_imm)
+(decl shufps_rev_imm(u8) Immediate)
+(extern extractor shufps_rev_imm shufps_rev_imm)
+
+
 ;; If `lhs` and `rhs` are the same we can use a single PSHUFB to shuffle the XMM
 ;; register. We statically build `constructed_mask` to zero out any unknown lane
 ;; indices (may not be completely necessary: verification could fail incorrect