Add optimizations for srem

Author: KGrewal1

cranelift/codegen/src/isle_prelude.rs

@@ -59,6 +59,28 @@ macro_rules! isle_common_prelude_methods {
             Some(Imm64::new(result))
         }
 
+        #[inline]
+        fn imm64_srem(&mut self, ty: Type, x: Imm64, y: Imm64) -> Option<Imm64> {
+            // Sign extend `x` and `y`.
+            let shift = u32::checked_sub(64, ty.bits()).unwrap_or(0);
+            let x = (x.bits() << shift) >> shift;
+            let y = (y.bits() << shift) >> shift;
+
+            // NB: We can't rely on `checked_rem` to detect `ty::MIN / -1`
+            // (which overflows and should trap) because we are working with
+            // `i64` values here, and `i32::MIN != i64::MIN`, for
+            // example. Therefore, we have to explicitly check for this case
+            // ourselves.
+            let min = ((self.ty_smin(ty) as i64) << shift) >> shift;
+            if x == min && y == -1 {
+                return None;
+            }
+
+            let ty_mask = self.ty_mask(ty) as i64;
+            let result = x.checked_rem(y)? & ty_mask;
+            Some(Imm64::new(result))
+        }
+
         #[inline]
         fn imm64_shl(&mut self, ty: Type, x: Imm64, y: Imm64) -> Imm64 {
             // Mask off any excess shift bits.

cranelift/codegen/src/opts/cprop.isle

@@ -35,6 +35,12 @@
       (if-let d (imm64_sdiv ty k1 k2))
       (iconst ty d))
 
+(rule (simplify_skeleton
+       (srem (iconst ty k1)
+             (iconst _ k2)))
+      (if-let d (imm64_srem ty k1 k2))
+      (iconst ty d))
+
 (rule (simplify_skeleton
        (udiv (iconst_u ty k1)
              (iconst_u ty k2)))

cranelift/codegen/src/prelude.isle

@@ -73,6 +73,9 @@
 (decl pure partial imm64_sdiv (Type Imm64 Imm64) Imm64)
 (extern constructor imm64_sdiv imm64_sdiv)
 
+(decl pure partial imm64_srem (Type Imm64 Imm64) Imm64)
+(extern constructor imm64_srem imm64_srem)
+
 (decl pure imm64_shl (Type Imm64 Imm64) Imm64)
 (extern constructor imm64_shl imm64_shl)
 

cranelift/filetests/filetests/egraph/skeleton.clif

@@ -110,6 +110,21 @@ block0:
 ;     return v11  ; v11 = -1
 ; }
 
+function %cprop_srem() -> i32 {
+block0:
+    v0 = iconst.i32 -17
+    v1 = iconst.i32 7
+    v2 = srem v0, v1
+    return v2
+}
+
+; function %cprop_srem() -> i32 fast {
+; block0:
+;     v28 = iconst.i32 -3
+;     v2 -> v28
+;     return v28  ; v28 = -3
+; }
+
 function %udiv_by_one(i32) -> i32 {
 block0(v0: i32):
     v1 = iconst.i32 1
@@ -228,4 +243,3 @@ block0:
 ;     v2 = uadd_overflow_trap v0, v1, user42  ; v0 = -1, v1 = 1
 ;     return v2
 ; }
-