Legalize [u|s]widen_high for x86

Use `x86_palignr` and `[u|s]widen_low` for legalizing this instruction.

Author: abrown

cranelift/codegen/meta/src/isa/x86/legalize.rs

@@ -407,13 +407,18 @@ fn define_simd(
     let umax = insts.by_name("umax");
     let umin = insts.by_name("umin");
     let snarrow = insts.by_name("snarrow");
+    let swiden_high = insts.by_name("swiden_high");
+    let swiden_low = insts.by_name("swiden_low");
     let ushr_imm = insts.by_name("ushr_imm");
     let ushr = insts.by_name("ushr");
+    let uwiden_high = insts.by_name("uwiden_high");
+    let uwiden_low = insts.by_name("uwiden_low");
     let vconst = insts.by_name("vconst");
     let vall_true = insts.by_name("vall_true");
     let vany_true = insts.by_name("vany_true");
     let vselect = insts.by_name("vselect");
 
+    let x86_palignr = x86_instructions.by_name("x86_palignr");
     let x86_pmaxs = x86_instructions.by_name("x86_pmaxs");
     let x86_pmaxu = x86_instructions.by_name("x86_pmaxu");
     let x86_pmins = x86_instructions.by_name("x86_pmins");
@@ -786,6 +791,26 @@ fn define_simd(
         );
     }
 
+    // SIMD widen
+    for ty in &[I8, I16] {
+        let swiden_high = swiden_high.bind(vector(*ty, sse_vector_size));
+        narrow.legalize(
+            def!(b = swiden_high(a)),
+            vec![
+                def!(c = x86_palignr(a, a, uimm8_eight)),
+                def!(b = swiden_low(c)),
+            ],
+        );
+        let uwiden_high = uwiden_high.bind(vector(*ty, sse_vector_size));
+        narrow.legalize(
+            def!(b = uwiden_high(a)),
+            vec![
+                def!(c = x86_palignr(a, a, uimm8_eight)),
+                def!(b = uwiden_low(c)),
+            ],
+        );
+    }
+
     narrow.custom_legalize(shuffle, "convert_shuffle");
     narrow.custom_legalize(extractlane, "convert_extractlane");
     narrow.custom_legalize(insertlane, "convert_insertlane");

cranelift/filetests/filetests/isa/x86/simd-conversion-legalize.clif

@@ -52,3 +52,19 @@ block0(v0:f32x4):
     ; nextln: v1 = iadd v12, v11
     return v1
 }
+
+function %uwiden_high(i8x16) -> i16x8 {
+block0(v0: i8x16):
+    v1 = uwiden_high v0
+    ; check: v2 = x86_palignr v0, v0, 8
+    ; nextln: v1 = uwiden_low v2
+    return v1
+}
+
+function %swiden_high(i16x8) -> i32x4 {
+block0(v0: i16x8):
+    v1 = swiden_high v0
+    ; check: v2 = x86_palignr v0, v0, 8
+    ; nextln: v1 = swiden_low v2
+    return v1
+}