Add f16 inline ASM support for 32-bit ARM

rust-lang · Jun 21, 2024 · 753fb07 · 753fb07
1 parent 12b33d3
commit 753fb07
Show file tree

Hide file tree

Showing 3 changed files with 365 additions and 183 deletions.
diff --git a/compiler/rustc_codegen_llvm/src/asm.rs b/compiler/rustc_codegen_llvm/src/asm.rs
@@ -1020,6 +1020,19 @@ fn llvm_fixup_input<'ll, 'tcx>(
                 value
             }
         }
+        (
+            InlineAsmRegClass::Arm(
+                ArmInlineAsmRegClass::dreg
+                | ArmInlineAsmRegClass::dreg_low8
+                | ArmInlineAsmRegClass::dreg_low16
+                | ArmInlineAsmRegClass::qreg
+                | ArmInlineAsmRegClass::qreg_low4
+                | ArmInlineAsmRegClass::qreg_low8,
+            ),
+            Abi::Vector { element, count: count @ (4 | 8) },
+        ) if element.primitive() == Primitive::Float(Float::F16) => {
+            bx.bitcast(value, bx.type_vector(bx.type_i16(), count))
+        }
         (InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg), Abi::Scalar(s)) => {
             match s.primitive() {
                 // MIPS only supports register-length arithmetics.
@@ -1130,6 +1143,19 @@ fn llvm_fixup_output<'ll, 'tcx>(
                 value
             }
         }
+        (
+            InlineAsmRegClass::Arm(
+                ArmInlineAsmRegClass::dreg
+                | ArmInlineAsmRegClass::dreg_low8
+                | ArmInlineAsmRegClass::dreg_low16
+                | ArmInlineAsmRegClass::qreg
+                | ArmInlineAsmRegClass::qreg_low4
+                | ArmInlineAsmRegClass::qreg_low8,
+            ),
+            Abi::Vector { element, count: count @ (4 | 8) },
+        ) if element.primitive() == Primitive::Float(Float::F16) => {
+            bx.bitcast(value, bx.type_vector(bx.type_f16(), count))
+        }
         (InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg), Abi::Scalar(s)) => {
             match s.primitive() {
                 // MIPS only supports register-length arithmetics.
@@ -1233,6 +1259,19 @@ fn llvm_fixup_output_type<'ll, 'tcx>(
                 layout.llvm_type(cx)
             }
         }
+        (
+            InlineAsmRegClass::Arm(
+                ArmInlineAsmRegClass::dreg
+                | ArmInlineAsmRegClass::dreg_low8
+                | ArmInlineAsmRegClass::dreg_low16
+                | ArmInlineAsmRegClass::qreg
+                | ArmInlineAsmRegClass::qreg_low4
+                | ArmInlineAsmRegClass::qreg_low8,
+            ),
+            Abi::Vector { element, count: count @ (4 | 8) },
+        ) if element.primitive() == Primitive::Float(Float::F16) => {
+            cx.type_vector(cx.type_i16(), count)
+        }
         (InlineAsmRegClass::Mips(MipsInlineAsmRegClass::reg), Abi::Scalar(s)) => {
             match s.primitive() {
                 // MIPS only supports register-length arithmetics.

diff --git a/compiler/rustc_target/src/asm/arm.rs b/compiler/rustc_target/src/asm/arm.rs
@@ -47,16 +47,18 @@ impl ArmInlineAsmRegClass {
         _arch: InlineAsmArch,
     ) -> &'static [(InlineAsmType, Option<Symbol>)] {
         match self {
-            Self::reg => types! { _: I8, I16, I32, F32; },
-            Self::sreg | Self::sreg_low16 => types! { vfp2: I32, F32; },
+            Self::reg => types! { _: I8, I16, I32, F16, F32; },
+            Self::sreg | Self::sreg_low16 => types! { vfp2: I32, F16, F32; },
             Self::dreg_low16 | Self::dreg_low8 => types! {
-                vfp2: I64, F64, VecI8(8), VecI16(4), VecI32(2), VecI64(1), VecF32(2);
+                vfp2: I64, F64;
+                neon: VecI8(8), VecI16(4), VecI32(2), VecI64(1), VecF16(4), VecF32(2);
             },
             Self::dreg => types! {
-                d32: I64, F64, VecI8(8), VecI16(4), VecI32(2), VecI64(1), VecF32(2);
+                d32: I64, F64;
+                neon: VecI8(8), VecI16(4), VecI32(2), VecI64(1), VecF16(4), VecF32(2);
             },
             Self::qreg | Self::qreg_low8 | Self::qreg_low4 => types! {
-                neon: VecI8(16), VecI16(8), VecI32(4), VecI64(2), VecF32(4);
+                neon: VecI8(16), VecI16(8), VecI32(4), VecI64(2), VecF16(8), VecF32(4);
             },
         }
     }