From 30ece8da06882b9cbca9f7e73f6804c669449a17 Mon Sep 17 00:00:00 2001
From: bjorn3 <17426603+bjorn3@users.noreply.github.com>
Date: Wed, 26 Feb 2025 13:06:01 +0000
Subject: [PATCH] Make tiny-skia work on arm64

---
 src/intrinsics/llvm.rs         |   8 +++
 src/intrinsics/llvm_aarch64.rs | 119 +++++++++++++++++++++++++++++++++
 2 files changed, 127 insertions(+)

diff --git a/src/intrinsics/llvm.rs b/src/intrinsics/llvm.rs
index 720a0d8fbf59..eb0dfbb69c3b 100644
--- a/src/intrinsics/llvm.rs
+++ b/src/intrinsics/llvm.rs
@@ -54,6 +54,14 @@ pub(crate) fn codegen_llvm_intrinsic_call<'tcx>(
             );
         }
 
+        "llvm.fptosi.sat.v4i32.v4f32" => {
+            intrinsic_args!(fx, args => (a); intrinsic);
+
+            simd_for_each_lane(fx, a, ret, &|fx, _lane_ty, _res_lane_ty, lane| {
+                fx.bcx.ins().fcvt_to_sint_sat(types::I32, lane)
+            });
+        }
+
         _ => {
             fx.tcx
                 .dcx()
diff --git a/src/intrinsics/llvm_aarch64.rs b/src/intrinsics/llvm_aarch64.rs
index 4c59c81296ba..b77c99fa2896 100644
--- a/src/intrinsics/llvm_aarch64.rs
+++ b/src/intrinsics/llvm_aarch64.rs
@@ -1,5 +1,9 @@
 //! Emulate AArch64 LLVM intrinsics
 
+use rustc_ast::ast::{InlineAsmOptions, InlineAsmTemplatePiece};
+use rustc_target::asm::*;
+
+use crate::inline_asm::{CInlineAsmOperand, codegen_inline_asm_inner};
 use crate::intrinsics::*;
 use crate::prelude::*;
 
@@ -49,6 +53,121 @@ pub(crate) fn codegen_aarch64_llvm_intrinsic_call<'tcx>(
             });
         }
 
+        "llvm.aarch64.neon.fcvtns.v4i32.v4f32" => {
+            intrinsic_args!(fx, args => (a); intrinsic);
+
+            // Note: Using inline asm instead of fcvt_to_sint as the latter rounds to zero rather than to nearest
+
+            let a_ptr = a.force_stack(fx).0.get_addr(fx);
+            let res_place = CPlace::new_stack_slot(fx, ret.layout());
+            let res_ptr = res_place.to_ptr().get_addr(fx);
+
+            codegen_inline_asm_inner(
+                fx,
+                &[InlineAsmTemplatePiece::String(
+                    "ldr     q0, [x0]
+                     fcvtns  v0.4s, v0.4s
+                     str     q0, [x1]"
+                        .into(),
+                )],
+                &[
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::AArch64(
+                            AArch64InlineAsmReg::x0,
+                        )),
+                        value: a_ptr,
+                    },
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::AArch64(
+                            AArch64InlineAsmReg::x1,
+                        )),
+                        value: res_ptr,
+                    },
+                ],
+                InlineAsmOptions::NOSTACK,
+            );
+            let res = res_place.to_cvalue(fx);
+            ret.write_cvalue_transmute(fx, res);
+        }
+
+        "llvm.aarch64.neon.frecpe.v4f32" => {
+            intrinsic_args!(fx, args => (a); intrinsic);
+
+            let a_ptr = a.force_stack(fx).0.get_addr(fx);
+            let res_place = CPlace::new_stack_slot(fx, ret.layout());
+            let res_ptr = res_place.to_ptr().get_addr(fx);
+
+            codegen_inline_asm_inner(
+                fx,
+                &[InlineAsmTemplatePiece::String(
+                    "ldr     q0, [x0]
+                     frecpe  v0.4s, v0.4s
+                     str     q0, [x1]"
+                        .into(),
+                )],
+                &[
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::AArch64(
+                            AArch64InlineAsmReg::x0,
+                        )),
+                        value: a_ptr,
+                    },
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::AArch64(
+                            AArch64InlineAsmReg::x1,
+                        )),
+                        value: res_ptr,
+                    },
+                ],
+                InlineAsmOptions::NOSTACK,
+            );
+            let res = res_place.to_cvalue(fx);
+            ret.write_cvalue_transmute(fx, res);
+        }
+
+        "llvm.aarch64.neon.frecps.v4f32" => {
+            intrinsic_args!(fx, args => (a, b); intrinsic);
+
+            let a_ptr = a.force_stack(fx).0.get_addr(fx);
+            let b_ptr = b.force_stack(fx).0.get_addr(fx);
+            let res_place = CPlace::new_stack_slot(fx, ret.layout());
+            let res_ptr = res_place.to_ptr().get_addr(fx);
+
+            codegen_inline_asm_inner(
+                fx,
+                &[InlineAsmTemplatePiece::String(
+                    "ldr     q0, [x0]
+                     ldr     q1, [x1]
+                     frecps  v0.4s, v0.4s, v1.4s
+                     str     q0, [x2]"
+                        .into(),
+                )],
+                &[
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::AArch64(
+                            AArch64InlineAsmReg::x0,
+                        )),
+                        value: a_ptr,
+                    },
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::AArch64(
+                            AArch64InlineAsmReg::x1,
+                        )),
+                        value: b_ptr,
+                    },
+                    CInlineAsmOperand::In {
+                        reg: InlineAsmRegOrRegClass::Reg(InlineAsmReg::AArch64(
+                            AArch64InlineAsmReg::x2,
+                        )),
+                        value: res_ptr,
+                    },
+                ],
+                InlineAsmOptions::NOSTACK,
+            );
+            let res = res_place.to_cvalue(fx);
+            ret.write_cvalue_transmute(fx, res);
+        }
+
         _ if intrinsic.starts_with("llvm.aarch64.neon.sqadd.v")
             || intrinsic.starts_with("llvm.aarch64.neon.uqadd.v") =>
         {