Avoid extra register movement when lowering the x86 scalar_to_vector of a float value

cranelift/codegen/meta/src/cdsl/types.rs

@@ -264,6 +264,27 @@ impl LaneType {
             ValueType::Vector(VectorType::new(*self, lanes.into()))
         }
     }
+
+    pub fn is_float(&self) -> bool {
+        match self {
+            LaneType::FloatType(_) => true,
+            _ => false,
+        }
+    }
+
+    pub fn is_int(&self) -> bool {
+        match self {
+            LaneType::IntType(_) => true,
+            _ => false,
+        }
+    }
+
+    pub fn is_bool(&self) -> bool {
+        match self {
+            LaneType::BoolType(_) => true,
+            _ => false,
+        }
+    }
 }
 
 impl fmt::Display for LaneType {

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

@@ -4,8 +4,8 @@ use std::collections::HashMap;
 
 use crate::cdsl::encodings::{Encoding, EncodingBuilder};
 use crate::cdsl::instructions::{
-    BoundInstruction, InstSpec, Instruction, InstructionGroup, InstructionPredicate,
-    InstructionPredicateNode, InstructionPredicateRegistry,
+    InstSpec, Instruction, InstructionGroup, InstructionPredicate, InstructionPredicateNode,
+    InstructionPredicateRegistry,
 };
 use crate::cdsl::recipes::{EncodingRecipe, EncodingRecipeNumber, Recipes};
 use crate::cdsl::settings::{SettingGroup, SettingPredicateNumber};
@@ -279,6 +279,17 @@ impl PerCpuModeEncodings {
         }
     }
 
+    /// Add the same encoding/recipe pairing to both X86_32 and X86_64
+    fn enc_32_64_rec(
+        &mut self,
+        inst: impl Clone + Into<InstSpec>,
+        recipe: &EncodingRecipe,
+        bits: u16,
+    ) {
+        self.enc32_rec(inst.clone(), recipe, bits);
+        self.enc64_rec(inst, recipe, bits);
+    }
+
     /// Add the same encoding to both X86_32 and X86_64; assumes configuration (e.g. REX, operand
     /// binding) has already happened.
     fn enc_32_64_maybe_isap(
@@ -1761,6 +1772,9 @@ pub(crate) fn define(
     // written to the low doubleword of the register and the regiser is zero-extended to 128 bits."
     for ty in ValueType::all_lane_types().filter(allowed_simd_type) {
         let instruction = scalar_to_vector.bind_vector_from_lane(ty, sse_vector_size);
+        if ty.is_float() {
+            e.enc_32_64_rec(instruction, rec_null_fpr, 0);
+        } else {
             let template = rec_frurm.opcodes(vec![0x66, 0x0f, 0x6e]); // MOVD/MOVQ
             if ty.lane_bits() < 64 {
                 // no 32-bit encodings for 64-bit widths
@@ -1768,6 +1782,7 @@ pub(crate) fn define(
             }
             e.enc_x86_64(instruction, template);
         }
+    }
 
     // SIMD insertlane
     let mut insertlane_mapping: HashMap<u64, (Vec<u8>, Option<SettingPredicateNumber>)> =
@@ -1811,37 +1826,34 @@ pub(crate) fn define(
         }
     }
 
-    // helper for generating null encodings for FPRs on both 32- and 64-bit architectures
-    let mut null_encode_32_64 = |instruction: BoundInstruction| {
-        e.enc32_rec(instruction.clone(), rec_null_fpr, 0);
-        e.enc64_rec(instruction, rec_null_fpr, 0);
-    };
-
     // SIMD bitcast all 128-bit vectors to each other (for legalizing splat.x16x8)
     for from_type in ValueType::all_lane_types().filter(allowed_simd_type) {
         for to_type in
             ValueType::all_lane_types().filter(|t| allowed_simd_type(t) && *t != from_type)
         {
-            null_encode_32_64(
-                raw_bitcast
+            let instruction = raw_bitcast
                 .bind_vector_from_lane(to_type, sse_vector_size)
-                    .bind_vector_from_lane(from_type, sse_vector_size),
-            );
+                .bind_vector_from_lane(from_type, sse_vector_size);
+            e.enc_32_64_rec(instruction, rec_null_fpr, 0);
         }
     }
 
     // SIMD raw bitcast floats to vector (and back); assumes that floats are already stored in an XMM register
     for float_type in &[F32, F64] {
         for lane_type in ValueType::all_lane_types().filter(allowed_simd_type) {
-            null_encode_32_64(
+            e.enc_32_64_rec(
                 raw_bitcast
                     .bind_vector_from_lane(lane_type, sse_vector_size)
                     .bind(*float_type),
+                rec_null_fpr,
+                0,
             );
-            null_encode_32_64(
+            e.enc_32_64_rec(
                 raw_bitcast
                     .bind(*float_type)
                     .bind_vector_from_lane(lane_type, sse_vector_size),
+                rec_null_fpr,
+                0,
             );
         }
     }

cranelift/filetests/filetests/isa/x86/extractlane-run.clif

@@ -24,8 +24,7 @@ ebb0:
     v0 = f32const 0x42.42
     v1 = vconst.f32x4 [0x00.00 0x00.00 0x00.00 0x42.42]
     v2 = extractlane v1, 3
-    v10 = f32const 0x42.42  ; TODO this should not be necessary, v0 should be re-usable
-    v3 = fcmp eq v2, v10
+    v3 = fcmp eq v2, v0
     return v3
 }
 ; run

cranelift/filetests/filetests/isa/x86/scalar_to_vector-binemit.clif

@@ -17,10 +17,10 @@ ebb0:
             return
 }
 
-function %test_scalar_to_vector_f32() {
+function %test_scalar_to_vector_b32() {
 ebb0:
-[-, %rcx]   v0 = f32const 0x0.42
-[-, %xmm3]  v1 = scalar_to_vector.f32x4 v0    ; bin: 66 0f 6e d9
+[-, %rcx]   v0 = bconst.b32 false
+[-, %xmm3]  v1 = scalar_to_vector.b32x4 v0    ; bin: 66 0f 6e d9
             return
 }
 

cranelift/filetests/filetests/isa/x86/scalar_to_vector-compile.clif

@@ -0,0 +1,19 @@
+test compile
+set opt_level=best
+set probestack_enabled=false
+set enable_simd
+target x86_64
+
+; ensure that scalar_to_vector emits no instructions for floats (already exist in an XMM register)
+function %test_scalar_to_vector_f32() -> f32x4 baldrdash_system_v {
+ebb0:
+    v0 = f32const 0x0.42
+    v1 = scalar_to_vector.f32x4 v0
+    return v1
+}
+
+; check: ebb0
+; nextln: v2 = iconst.i32 0x3e84_0000
+; nextln: v0 = bitcast.f32 v2
+; nextln: [null_fpr#00,%xmm0]                 v1 = scalar_to_vector.f32x4 v0
+; nextln: return v1