Add support sqrtps and sqrtpd (#10768)

Will pursue sqrtss and sqrtsd in a follow-up patch after the isle
support needed there is understood.

Author: jlb6740

cranelift/assembler-x64/meta/src/instructions.rs

@@ -8,6 +8,7 @@ mod mul;
 mod neg;
 mod or;
 mod shift;
+mod sqrt;
 mod sub;
 mod xor;
 
@@ -24,6 +25,7 @@ pub fn list() -> Vec<Inst> {
     all.extend(neg::list());
     all.extend(or::list());
     all.extend(shift::list());
+    all.extend(sqrt::list());
     all.extend(sub::list());
     all.extend(xor::list());
     all

cranelift/assembler-x64/meta/src/instructions/sqrt.rs

@@ -0,0 +1,11 @@
+use crate::dsl::{Feature::*, Inst, Location::*};
+use crate::dsl::{align, fmt, inst, r, rex, w};
+
+#[rustfmt::skip] // Keeps instructions on a single line.
+pub fn list() -> Vec<Inst> {
+    vec![
+        // Vector instructions.
+        inst("sqrtpd", fmt("A", [w(xmm), r(align(xmm_m128))]), rex([0x66, 0x0F, 0x51]).r(), _64b | compat | sse2),
+        inst("sqrtps", fmt("A", [w(xmm), r(align(xmm_m128))]), rex([0x0F, 0x51]).r(), _64b | compat | sse),
+    ]
+}

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

@@ -4728,17 +4728,19 @@
 
 ;; Helper for creating `sqrtps` instructions.
 (decl x64_sqrtps (XmmMem) Xmm)
-(rule (x64_sqrtps x) (xmm_unary_rm_r (SseOpcode.Sqrtps) x))
 (rule 1 (x64_sqrtps x)
         (if-let true (use_avx))
         (xmm_unary_rm_r_vex (AvxOpcode.Vsqrtps) x))
+(rule 0 (x64_sqrtps x)
+      (x64_sqrtps_a x))
 
 ;; Helper for creating `sqrtpd` instructions.
 (decl x64_sqrtpd (XmmMem) Xmm)
-(rule (x64_sqrtpd x) (xmm_unary_rm_r (SseOpcode.Sqrtpd) x))
 (rule 1 (x64_sqrtpd x)
         (if-let true (use_avx))
         (xmm_unary_rm_r_vex (AvxOpcode.Vsqrtpd) x))
+(rule 0 (x64_sqrtpd x)
+      (x64_sqrtpd_a x))
 
 ;; Helper for creating `cvtss2sd` instructions.
 ;;

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

@@ -1009,8 +1009,6 @@ pub enum SseOpcode {
     Roundsd,
     Rsqrtss,
     Shufps,
-    Sqrtps,
-    Sqrtpd,
     Sqrtss,
     Sqrtsd,
     Ucomiss,
@@ -1052,7 +1050,6 @@ impl SseOpcode {
             | SseOpcode::Rcpss
             | SseOpcode::Rsqrtss
             | SseOpcode::Shufps
-            | SseOpcode::Sqrtps
             | SseOpcode::Sqrtss
             | SseOpcode::Ucomiss
             | SseOpcode::Unpcklps
@@ -1105,7 +1102,6 @@ impl SseOpcode {
             | SseOpcode::Punpckhwd
             | SseOpcode::Punpcklbw
             | SseOpcode::Punpcklwd
-            | SseOpcode::Sqrtpd
             | SseOpcode::Sqrtsd
             | SseOpcode::Ucomisd
             | SseOpcode::Punpckldq
@@ -1313,8 +1309,6 @@ impl fmt::Debug for SseOpcode {
             SseOpcode::Roundsd => "roundsd",
             SseOpcode::Rsqrtss => "rsqrtss",
             SseOpcode::Shufps => "shufps",
-            SseOpcode::Sqrtps => "sqrtps",
-            SseOpcode::Sqrtpd => "sqrtpd",
             SseOpcode::Sqrtss => "sqrtss",
             SseOpcode::Sqrtsd => "sqrtsd",
             SseOpcode::Ucomiss => "ucomiss",

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

@@ -1796,8 +1796,6 @@ pub(crate) fn emit(
                 SseOpcode::Pmovzxwd => (LegacyPrefixes::_66, 0x0F3833, 3),
                 SseOpcode::Pmovzxwq => (LegacyPrefixes::_66, 0x0F3834, 3),
                 SseOpcode::Pmovzxdq => (LegacyPrefixes::_66, 0x0F3835, 3),
-                SseOpcode::Sqrtps => (LegacyPrefixes::None, 0x0F51, 2),
-                SseOpcode::Sqrtpd => (LegacyPrefixes::_66, 0x0F51, 2),
                 SseOpcode::Movddup => (LegacyPrefixes::_F2, 0x0F12, 2),
                 _ => unimplemented!("Opcode {:?} not implemented", op),
             };