From f05babc74414b59a8eb2d92e13ce4708c7482711 Mon Sep 17 00:00:00 2001
From: Alex Crichton <alex@alexcrichton.com>
Date: Wed, 1 Mar 2023 15:49:00 -0600
Subject: [PATCH] x64: Add `shuffle` cases for `punpck{h,l}bw` (#5905)

* x64: Add `shuffle` cases for `punpck{h,l}bw`

I noticed this difference between LLVM and Cranelift for something I was
looking at recently, and while it's probably not all that common I
figured I'd add it here since it should be somewhat useful nevertheless.

* Review feedback

* Use u128 extractor instead
---
 cranelift/codegen/src/isa/x64/lower.isle      | 10 ++++
 .../filetests/filetests/isa/x64/shuffle.clif  | 54 +++++++++++++++++++
 2 files changed, 64 insertions(+)
 create mode 100644 cranelift/filetests/filetests/isa/x64/shuffle.clif

diff --git a/cranelift/codegen/src/isa/x64/lower.isle b/cranelift/codegen/src/isa/x64/lower.isle
index 38c1f2de79..9f83ccbfb9 100644
--- a/cranelift/codegen/src/isa/x64/lower.isle
+++ b/cranelift/codegen/src/isa/x64/lower.isle
@@ -3510,6 +3510,16 @@
 
 ;; 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 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)))
+      (x64_punpcklbw a b))
+
 ;; 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
diff --git a/cranelift/filetests/filetests/isa/x64/shuffle.clif b/cranelift/filetests/filetests/isa/x64/shuffle.clif
new file mode 100644
index 0000000000..529b95cc5d
--- /dev/null
+++ b/cranelift/filetests/filetests/isa/x64/shuffle.clif
@@ -0,0 +1,54 @@
+test compile precise-output
+set enable_simd
+target x86_64
+
+function %punpcklbw(i8x16, i8x16) -> i8x16 {
+block0(v0: i8x16, v1: i8x16):
+    v2 = shuffle v0, v1, 0x17071606150514041303120211011000
+    return v2
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   punpcklbw %xmm0, %xmm1, %xmm0
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   punpcklbw %xmm1, %xmm0
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+
+function %punpckhbw(i8x16, i8x16) -> i8x16 {
+block0(v0: i8x16, v1: i8x16):
+    v2 = shuffle v0, v1, 0x1f0f1e0e1d0d1c0c1b0b1a0a19091808
+    return v2
+}
+
+; VCode:
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   punpckhbw %xmm0, %xmm1, %xmm0
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+; 
+; Disassembled:
+; block0: ; offset 0x0
+;   pushq %rbp
+;   movq %rsp, %rbp
+; block1: ; offset 0x4
+;   punpckhbw %xmm1, %xmm0
+;   movq %rbp, %rsp
+;   popq %rbp
+;   retq
+