x64: Add native lowering for scalar fma (#4539)

Use `vfmadd213{ss,sd}` for these lowerings.

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

@@ -1095,7 +1095,9 @@
 (extern extractor cc_nz_or_z cc_nz_or_z)
 
 (type AvxOpcode extern
-      (enum Vfmadd213ps
+      (enum Vfmadd213ss
+            Vfmadd213sd
+            Vfmadd213ps
             Vfmadd213pd))
 
 (type Avx512Opcode extern
@@ -1389,6 +1391,9 @@
 (decl use_popcnt () Type)
 (extern extractor use_popcnt use_popcnt)
 
+(decl use_fma () Type)
+(extern extractor use_fma use_fma)
+
 ;;;; Helpers for Merging and Sinking Immediates/Loads  ;;;;;;;;;;;;;;;;;;;;;;;;;
 
 ;; Extract a constant `Imm8Reg.Imm8` from a value operand.
@@ -2935,6 +2940,16 @@
                                            dst))))
         dst))
 
+;; Helper for creating `vfmadd213ss` instructions.
+(decl x64_vfmadd213ss (Xmm Xmm XmmMem) Xmm)
+(rule (x64_vfmadd213ss x y z)
+      (xmm_rmr_vex (AvxOpcode.Vfmadd213ss) x y z))
+
+;; Helper for creating `vfmadd213sd` instructions.
+(decl x64_vfmadd213sd (Xmm Xmm XmmMem) Xmm)
+(rule (x64_vfmadd213sd x y z)
+      (xmm_rmr_vex (AvxOpcode.Vfmadd213sd) x y z))
+
 ;; Helper for creating `vfmadd213ps` instructions.
 (decl x64_vfmadd213ps (Xmm Xmm XmmMem) Xmm)
 (rule (x64_vfmadd213ps x y z)

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

@@ -1383,6 +1383,8 @@ impl fmt::Display for SseOpcode {
 
 #[derive(Clone, PartialEq)]
 pub enum AvxOpcode {
+    Vfmadd213ss,
+    Vfmadd213sd,
     Vfmadd213ps,
     Vfmadd213pd,
 }
@@ -1391,8 +1393,10 @@ impl AvxOpcode {
     /// Which `InstructionSet`s support the opcode?
     pub(crate) fn available_from(&self) -> SmallVec<[InstructionSet; 2]> {
         match self {
-            AvxOpcode::Vfmadd213ps => smallvec![InstructionSet::FMA],
+            AvxOpcode::Vfmadd213ss
-            AvxOpcode::Vfmadd213pd => smallvec![InstructionSet::FMA],
+            | AvxOpcode::Vfmadd213sd
+            | AvxOpcode::Vfmadd213ps
+            | AvxOpcode::Vfmadd213pd => smallvec![InstructionSet::FMA],
         }
     }
 }
@@ -1400,6 +1404,8 @@ impl AvxOpcode {
 impl fmt::Debug for AvxOpcode {
     fn fmt(&self, fmt: &mut fmt::Formatter) -> fmt::Result {
         let name = match self {
+            AvxOpcode::Vfmadd213ss => "vfmadd213ss",
+            AvxOpcode::Vfmadd213sd => "vfmadd213sd",
             AvxOpcode::Vfmadd213ps => "vfmadd213ps",
             AvxOpcode::Vfmadd213pd => "vfmadd213pd",
         };

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

@@ -1742,6 +1742,8 @@ pub(crate) fn emit(
             let src3 = src3.clone().to_reg_mem().with_allocs(allocs);
 
             let (w, opcode) = match op {
+                AvxOpcode::Vfmadd213ss => (false, 0xA9),
+                AvxOpcode::Vfmadd213sd => (true, 0xA9),
                 AvxOpcode::Vfmadd213ps => (false, 0xA8),
                 AvxOpcode::Vfmadd213pd => (true, 0xA8),
             };

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

@@ -3531,6 +3531,18 @@ fn test_x64_emit() {
     // ========================================================
     // XMM FMA
 
+    insns.push((
+        Inst::xmm_rm_r_vex(AvxOpcode::Vfmadd213ss, RegMem::reg(xmm2), xmm1, w_xmm0),
+        "C4E271A9C2",
+        "vfmadd213ss %xmm0, %xmm1, %xmm2, %xmm0",
+    ));
+
+    insns.push((
+        Inst::xmm_rm_r_vex(AvxOpcode::Vfmadd213sd, RegMem::reg(xmm5), xmm4, w_xmm3),
+        "C4E2D9A9DD",
+        "vfmadd213sd %xmm3, %xmm4, %xmm5, %xmm3",
+    ));
+
     insns.push((
         Inst::xmm_rm_r_vex(AvxOpcode::Vfmadd213ps, RegMem::reg(xmm2), xmm1, w_xmm0),
         "C4E271A8C2",

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

@@ -1847,7 +1847,12 @@ fn x64_get_operands<F: Fn(VReg) -> VReg>(inst: &Inst, collector: &mut OperandCol
             // Vfmadd uses and defs the dst reg, that is not the case with all
             // AVX's ops, if you're adding a new op, make sure to correctly define
             // register uses.
-            assert!(*op == AvxOpcode::Vfmadd213ps || *op == AvxOpcode::Vfmadd213pd);
+            assert!(
+                *op == AvxOpcode::Vfmadd213ss
+                    || *op == AvxOpcode::Vfmadd213sd
+                    || *op == AvxOpcode::Vfmadd213ps
+                    || *op == AvxOpcode::Vfmadd213pd
+            );
 
             collector.reg_use(src1.to_reg());
             collector.reg_reuse_def(dst.to_writable_reg(), 0);

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

@@ -2504,9 +2504,13 @@
       (libcall_3 (LibCall.FmaF32) x y z))
 (rule (lower (has_type $F64 (fma x y z)))
       (libcall_3 (LibCall.FmaF64) x y z))
-(rule (lower (has_type $F32X4 (fma x y z)))
+(rule 1 (lower (has_type (and (use_fma) $F32) (fma x y z)))
+      (x64_vfmadd213ss x y z))
+(rule 1 (lower (has_type (and (use_fma) $F64) (fma x y z)))
+      (x64_vfmadd213sd x y z))
+(rule (lower (has_type (and (use_fma) $F32X4) (fma x y z)))
       (x64_vfmadd213ps x y z))
-(rule (lower (has_type $F64X2 (fma x y z)))
+(rule (lower (has_type (and (use_fma) $F64X2) (fma x y z)))
       (x64_vfmadd213pd x y z))
 
 ;; Rules for `load*` ;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;;

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

@@ -248,6 +248,15 @@ where
         }
     }
 
+    #[inline]
+    fn use_fma(&mut self, _: Type) -> Option<()> {
+        if self.isa_flags.use_fma() {
+            Some(())
+        } else {
+            None
+        }
+    }
+
     #[inline]
     fn imm8_from_value(&mut self, val: Value) -> Option<Imm8Reg> {
         let inst = self.lower_ctx.dfg().value_def(val).inst()?;

cranelift/filetests/filetests/isa/x64/fma-call.clif

@@ -0,0 +1,33 @@
+test compile precise-output
+target x86_64 has_avx=false has_fma=false
+
+function %fma_f32(f32, f32, f32) -> f32 {
+block0(v0: f32, v1: f32, v2: f32):
+    v3 = fma v0, v1, v2
+    return v3
+}
+
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   load_ext_name %FmaF32+0, %rax
+;   call    *%rax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+
+function %fma_f64(f64, f64, f64) -> f64 {
+block0(v0: f64, v1: f64, v2: f64):
+    v3 = fma v0, v1, v2
+    return v3
+}
+
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   load_ext_name %FmaF64+0, %rax
+;   call    *%rax
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+

cranelift/filetests/filetests/isa/x64/fma-inst.clif

@@ -0,0 +1,31 @@
+test compile precise-output
+target x86_64 has_avx=true has_fma=true
+
+function %fma_f32(f32, f32, f32) -> f32 {
+block0(v0: f32, v1: f32, v2: f32):
+    v3 = fma v0, v1, v2
+    return v3
+}
+
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   vfmadd213ss %xmm0, %xmm1, %xmm2, %xmm0
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+
+function %fma_f64(f64, f64, f64) -> f64 {
+block0(v0: f64, v1: f64, v2: f64):
+    v3 = fma v0, v1, v2
+    return v3
+}
+
+;   pushq   %rbp
+;   movq    %rsp, %rbp
+; block0:
+;   vfmadd213sd %xmm0, %xmm1, %xmm2, %xmm0
+;   movq    %rbp, %rsp
+;   popq    %rbp
+;   ret
+

cranelift/filetests/filetests/runtests/fma.clif

@@ -2,6 +2,7 @@ test interpret
 test run
 target aarch64
 target s390x
+target x86_64 has_avx has_fma
 target x86_64 has_avx=false has_fma=false
 
 function %fma_f32(f32, f32, f32) -> f32 {