stage2,llvm: handle softfloats in @intToFloat and @floatToInt

If the hw doesn't have support for exotic floating-point types such
as `f80`, we lower the call to a compiler-rt function call instead.

I've added a behavior test specifically targeting this use case which
now passes on `aarch64-macos`.

Additionally, this commit makes it possible to successfully build
stage3 on `aarch64-macos`. We can print the compiler's help message,
however, building with it needs a little bit more love still.

1 files changed, 149 insertions, 14 deletions

diff --git a/src/codegen/llvm.zig b/src/codegen/llvm.zig
index 426ef7c378..374ae9b38e 100644
--- a/src/codegen/llvm.zig
+++ b/src/codegen/llvm.zig
@@ -4229,41 +4229,166 @@ pub const FuncGen = struct {
         return self.builder.buildInsertValue(partial, len, 1, "");
     }
 
+    inline fn isPowerOfTwo(bits: u64) bool {
+        return bits != 0 and ((bits & (~bits + 1)) == bits);
+    }
+
+    fn intTypeFromBitsAndSignRounded(self: *FuncGen, bits: u16, signed: bool) error{OutOfMemory}!Type {
+        const next_pow_two = math.log2_int_ceil(u16, bits);
+        const rounded_bits = @as(u32, 1) << next_pow_two;
+        return switch (rounded_bits) {
+            8, 16, 32 => if (signed) Type.initTag(.i32) else Type.initTag(.u32),
+            64 => if (signed) Type.initTag(.i64) else Type.initTag(.u64),
+            128 => if (signed) Type.initTag(.i128) else Type.initTag(.u128),
+            else => |big| if (signed)
+                Type.Tag.int_signed.create(
+                    self.dg.object.type_map_arena.allocator(),
+                    @intCast(u16, big),
+                )
+            else
+                Type.Tag.int_unsigned.create(
+                    self.dg.object.type_map_arena.allocator(),
+                    @intCast(u16, big),
+                ),
+        };
+    }
+
     fn airIntToFloat(self: *FuncGen, inst: Air.Inst.Index) !?*const llvm.Value {
         if (self.liveness.isUnused(inst))
             return null;
 
+        const target = self.dg.module.getTarget();
         const ty_op = self.air.instructions.items(.data)[inst].ty_op;
-        const operand = try self.resolveInst(ty_op.operand);
-        const operand_ty = self.air.typeOf(ty_op.operand);
-        const operand_scalar_ty = operand_ty.scalarType();
+
+        var operand = try self.resolveInst(ty_op.operand);
+        var operand_ty = self.air.typeOf(ty_op.operand);
+        var operand_scalar_ty = operand_ty.scalarType();
+
+        {
+            const operand_bits = @intCast(u16, operand_scalar_ty.bitSize(target));
+            const is_signed = operand_scalar_ty.isSignedInt();
+
+            if (!isPowerOfTwo(operand_bits) or operand_bits < 32) {
+                const wider_ty = try self.intTypeFromBitsAndSignRounded(operand_bits, is_signed);
+                const wider_llvm_ty = try self.dg.llvmType(wider_ty);
+                if (is_signed) {
+                    operand = self.builder.buildSExt(operand, wider_llvm_ty, "");
+                } else {
+                    operand = self.builder.buildZExt(operand, wider_llvm_ty, "");
+                }
+                operand_ty = wider_ty;
+                operand_scalar_ty = operand_ty.scalarType();
+            }
+        }
+
         const dest_ty = self.air.typeOfIndex(inst);
+        const dest_scalar_ty = dest_ty.scalarType();
         const dest_llvm_ty = try self.dg.llvmType(dest_ty);
 
-        if (operand_scalar_ty.isSignedInt()) {
-            return self.builder.buildSIToFP(operand, dest_llvm_ty, "");
-        } else {
-            return self.builder.buildUIToFP(operand, dest_llvm_ty, "");
+        if (intrinsicsAllowed(dest_scalar_ty, target)) {
+            if (operand_scalar_ty.isSignedInt()) {
+                return self.builder.buildSIToFP(operand, dest_llvm_ty, "");
+            } else {
+                return self.builder.buildUIToFP(operand, dest_llvm_ty, "");
+            }
         }
+
+        const operand_bits = @intCast(u16, operand_scalar_ty.bitSize(target));
+        const compiler_rt_operand_abbrev = compilerRtIntAbbrev(operand_bits);
+
+        const dest_bits = dest_scalar_ty.floatBits(target);
+        const compiler_rt_dest_abbrev = compilerRtFloatAbbrev(dest_bits);
+
+        var fn_name_buf: [64]u8 = undefined;
+        const fn_name = if (operand_scalar_ty.isSignedInt())
+            std.fmt.bufPrintZ(&fn_name_buf, "__float{s}i{s}f", .{
+                compiler_rt_operand_abbrev,
+                compiler_rt_dest_abbrev,
+            }) catch unreachable
+        else
+            std.fmt.bufPrintZ(&fn_name_buf, "__floatun{s}i{s}f", .{
+                compiler_rt_operand_abbrev,
+                compiler_rt_dest_abbrev,
+            }) catch unreachable;
+
+        const operand_llvm_ty = try self.dg.llvmType(operand_ty);
+        const param_types = [1]*const llvm.Type{operand_llvm_ty};
+        const libc_fn = self.getLibcFunction(fn_name, &param_types, dest_llvm_ty);
+        const params = [1]*const llvm.Value{operand};
+
+        return self.builder.buildCall(libc_fn, &params, params.len, .C, .Auto, "");
     }
 
     fn airFloatToInt(self: *FuncGen, inst: Air.Inst.Index) !?*const llvm.Value {
         if (self.liveness.isUnused(inst))
             return null;
 
+        const target = self.dg.module.getTarget();
         const ty_op = self.air.instructions.items(.data)[inst].ty_op;
+
         const operand = try self.resolveInst(ty_op.operand);
-        const dest_ty = self.air.typeOfIndex(inst);
-        const dest_scalar_ty = dest_ty.scalarType();
+        const operand_ty = self.air.typeOf(ty_op.operand);
+        const operand_scalar_ty = operand_ty.scalarType();
+
+        var dest_ty = self.air.typeOfIndex(inst);
+        var dest_scalar_ty = dest_ty.scalarType();
+
+        if (intrinsicsAllowed(operand_scalar_ty, target)) {
+            // TODO set fast math flag
+            const dest_llvm_ty = try self.dg.llvmType(dest_ty);
+            if (dest_scalar_ty.isSignedInt()) {
+                return self.builder.buildFPToSI(operand, dest_llvm_ty, "");
+            } else {
+                return self.builder.buildFPToUI(operand, dest_llvm_ty, "");
+            }
+        }
+
+        const needs_truncating = blk: {
+            const dest_bits = @intCast(u16, dest_scalar_ty.bitSize(target));
+
+            if (!isPowerOfTwo(dest_bits) or dest_bits < 32) {
+                dest_ty = try self.intTypeFromBitsAndSignRounded(dest_bits, dest_scalar_ty.isSignedInt());
+                dest_scalar_ty = dest_ty.scalarType();
+                break :blk true;
+            }
+
+            break :blk false;
+        };
+
         const dest_llvm_ty = try self.dg.llvmType(dest_ty);
 
-        // TODO set fast math flag
+        const operand_bits = operand_scalar_ty.floatBits(target);
+        const compiler_rt_operand_abbrev = compilerRtFloatAbbrev(operand_bits);
 
-        if (dest_scalar_ty.isSignedInt()) {
-            return self.builder.buildFPToSI(operand, dest_llvm_ty, "");
-        } else {
-            return self.builder.buildFPToUI(operand, dest_llvm_ty, "");
+        const dest_bits = @intCast(u16, dest_scalar_ty.bitSize(target));
+        const compiler_rt_dest_abbrev = compilerRtIntAbbrev(dest_bits);
+
+        var fn_name_buf: [64]u8 = undefined;
+        const fn_name = if (dest_scalar_ty.isSignedInt())
+            std.fmt.bufPrintZ(&fn_name_buf, "__fix{s}f{s}i", .{
+                compiler_rt_operand_abbrev,
+                compiler_rt_dest_abbrev,
+            }) catch unreachable
+        else
+            std.fmt.bufPrintZ(&fn_name_buf, "__fixun{s}f{s}i", .{
+                compiler_rt_operand_abbrev,
+                compiler_rt_dest_abbrev,
+            }) catch unreachable;
+
+        const operand_llvm_ty = try self.dg.llvmType(operand_ty);
+        const param_types = [1]*const llvm.Type{operand_llvm_ty};
+        const libc_fn = self.getLibcFunction(fn_name, &param_types, dest_llvm_ty);
+        const params = [1]*const llvm.Value{operand};
+
+        const result = self.builder.buildCall(libc_fn, &params, params.len, .C, .Auto, "");
+
+        if (needs_truncating) {
+            const requested_ty = self.air.typeOfIndex(inst);
+            const requested_llvm_ty = try self.dg.llvmType(requested_ty);
+            return self.builder.buildTrunc(result, requested_llvm_ty, "");
         }
+
+        return result;
     }
 
     fn airSliceField(self: *FuncGen, inst: Air.Inst.Index, index: c_uint) !?*const llvm.Value {
@@ -5615,6 +5740,16 @@ pub const FuncGen = struct {
         };
     }
 
+    fn compilerRtIntAbbrev(bits: u16) []const u8 {
+        return switch (bits) {
+            16 => "h",
+            32 => "s",
+            64 => "d",
+            128 => "t",
+            else => "o", // Non-standard
+        };
+    }
+
     /// Creates a floating point comparison by lowering to the appropriate
     /// hardware instruction or softfloat routine for the target
     fn buildFloatCmp(