Beginning to clean up x86 stacks

mkeeter · Mar 11, 2024 · 992a1f9 · 992a1f9
1 parent a51678f
commit 992a1f9
Show file tree

Hide file tree

Showing 5 changed files with 13 additions and 19 deletions.
diff --git a/fidget/src/jit/mod.rs b/fidget/src/jit/mod.rs
@@ -280,7 +280,6 @@ impl<T> AssemblerData<T> {
         }
     }
 
-    #[cfg(target_arch = "aarch64")]
     fn prepare_stack(&mut self, slot_count: usize, stack_size: usize) {
         // We always use the stack, if only to store callee-saved registers
         let mem = slot_count.saturating_sub(REGISTER_LIMIT)
@@ -289,24 +288,19 @@ impl<T> AssemblerData<T> {
 
         // Round up to the nearest multiple of 16 bytes, for alignment
         self.mem_offset = ((mem + 15) / 16) * 16;
+        self.push_stack();
+    }
+
+    #[cfg(target_arch = "aarch64")]
+    fn push_stack(&mut self) {
         assert!(self.mem_offset < 4096);
         dynasm!(self.ops
             ; sub sp, sp, #(self.mem_offset as u32)
         );
     }
 
     #[cfg(target_arch = "x86_64")]
-    fn prepare_stack(&mut self, slot_count: usize) {
-        // We always use the stack on x86_64, if only to store X/Y/Z
-        let stack_slots = slot_count.saturating_sub(REGISTER_LIMIT);
-
-        // We put X/Y/Z values at the top of the stack, where they can be
-        // accessed with `movss [rbp - i*size_of(T)] xmm`.  This frees up the
-        // incoming registers (xmm0-2) in the point evaluator.
-        let mem = (stack_slots + 4) * std::mem::size_of::<T>();
-
-        // Round up to the nearest multiple of 16 bytes, for alignment
-        self.mem_offset = ((mem + 15) / 16) * 16;
+    fn push_stack(&mut self) {
         dynasm!(self.ops
             ; sub rsp, self.mem_offset as i32
         );

diff --git a/fidget/src/jit/x86_64/float_slice.rs b/fidget/src/jit/x86_64/float_slice.rs
@@ -33,14 +33,14 @@ impl Assembler for FloatSliceAssembler {
         let mut out = AssemblerData::new(mmap);
         dynasm!(out.ops
             ; push rbp
+            ; mov rbp, rsp
             ; push r12
             ; push r13
             ; push r14
             ; push r15
-            ; mov rbp, rsp
             ; vzeroupper
         );
-        out.prepare_stack(slot_count);
+        out.prepare_stack(slot_count, 4);
         dynasm!(out.ops
             // The loop returns here, and we check whether to keep looping
             ; ->L:

diff --git a/fidget/src/jit/x86_64/grad_slice.rs b/fidget/src/jit/x86_64/grad_slice.rs
@@ -33,7 +33,7 @@ impl Assembler for GradSliceAssembler {
             ; mov rbp, rsp
             ; vzeroupper
         );
-        out.prepare_stack(slot_count);
+        out.prepare_stack(slot_count, 4);
         dynasm!(out.ops
             // The loop returns here, and we check whether to keep looping
             ; ->L:

diff --git a/fidget/src/jit/x86_64/interval.rs b/fidget/src/jit/x86_64/interval.rs
@@ -28,18 +28,18 @@ impl Assembler for IntervalAssembler {
         let mut out = AssemblerData::new(mmap);
         dynasm!(out.ops
             ; push rbp
+            ; mov rbp, rsp
             ; push r12
             ; push r13
             ; push r14
-            ; mov rbp, rsp
             ; vzeroupper
 
             // Put X/Y/Z on the stack so we can use those registers
             ; movq [rbp - 8], xmm0
             ; movq [rbp - 16], xmm1
             ; movq [rbp - 24], xmm2
         );
-        out.prepare_stack(slot_count);
+        out.prepare_stack(slot_count, 4);
         Self(out)
     }
     fn build_load(&mut self, dst_reg: u8, src_mem: u32) {

diff --git a/fidget/src/jit/x86_64/point.rs b/fidget/src/jit/x86_64/point.rs
@@ -30,18 +30,18 @@ impl Assembler for PointAssembler {
         let mut out = AssemblerData::new(mmap);
         dynasm!(out.ops
             ; push rbp
+            ; mov rbp, rsp
             ; push r12
             ; push r13
             ; push r14
             ; push r15
             ; vzeroupper
-            ; mov rbp, rsp
             // Put X/Y/Z on the stack so we can use those registers
             ; vmovss [rbp - 4], xmm0
             ; vmovss [rbp - 8], xmm1
             ; vmovss [rbp - 12], xmm2
         );
-        out.prepare_stack(slot_count);
+        out.prepare_stack(slot_count, 4);
         Self(out)
     }