optimize: pow when base is a power of two

if base == 2 ** k, then
   (2 ** k) ** n
== 2 ** (k * n)
== 1 << (k * n)

Author: Kmeakin

library/core/src/num/uint_macros.rs

@@ -2085,11 +2085,23 @@ macro_rules! uint_impl {
                       without modifying the original"]
         #[inline]
         pub const fn checked_pow(self, mut exp: u32) -> Option<Self> {
+            let mut base = self;
+            let mut acc: Self = 1;
+
+            if intrinsics::is_val_statically_known(base) && base.is_power_of_two() {
+                // change of base:
+                // if base == 2 ** k, then
+                //    (2 ** k) ** n
+                // == 2 ** (k * n)
+                // == 1 << (k * n)
+                let k = base.ilog2();
+                let shift = try_opt!(k.checked_mul(exp));
+                return (1 as Self).checked_shl(shift);
+            }
+
             if exp == 0 {
                 return Some(1);
             }
-            let mut base = self;
-            let mut acc: Self = 1;
 
             if intrinsics::is_val_statically_known(exp) {
                 while exp > 1 {
@@ -3255,14 +3267,28 @@ macro_rules! uint_impl {
                       without modifying the original"]
         #[inline]
         pub const fn overflowing_pow(self, mut exp: u32) -> (Self, bool) {
-            if exp == 0 {
-                return (1, false);
-            }
             let mut base = self;
             let mut acc: Self = 1;
             let mut overflow = false;
             let mut tmp_overflow;
 
+            if intrinsics::is_val_statically_known(base) && base.is_power_of_two() {
+                // change of base:
+                // if base == 2 ** k, then
+                //    (2 ** k) ** n
+                // == 2 ** (k * n)
+                // == 1 << (k * n)
+                let k = base.ilog2();
+                let Some(shift) = k.checked_mul(exp) else {
+                    return (0, true)
+                };
+                return ((1 as Self).unbounded_shl(shift), shift >= Self::BITS)
+            }
+
+            if exp == 0 {
+                return (1, false);
+            }
+
             if intrinsics::is_val_statically_known(exp) {
                 while exp > 1 {
                     if (exp & 1) == 1 {

tests/codegen-llvm/pow_known_base.rs

@@ -0,0 +1,43 @@
+//@ compile-flags: -Copt-level=3
+// Test that `pow` can use a faster implementation when `base` is a
+// known power of two
+
+#![crate_type = "lib"]
+
+// CHECK-LABEL: @pow2
+#[no_mangle]
+pub fn pow2(exp: u32) -> u32 {
+    // CHECK: %[[OVERFLOW:.+]] = icmp ugt i32 %exp, 31
+    // CHECK: %[[POW:.+]] = shl nuw i32 1, %exp
+    // CHECK: %[[RET:.+]] = select i1 %[[OVERFLOW]], i32 0, i32 %[[POW]]
+    // CHECK: ret i32 %[[RET]]
+    2u32.pow(exp)
+}
+
+// 4 ** n == 2 ** (2 * n) == 1 << (2 * n)
+// CHECK-LABEL: @pow4
+#[no_mangle]
+pub fn pow4(exp: u32) -> u32 {
+    // CHECK: %[[SHIFT_AMOUNT:.+]] = shl i32 %exp, 1
+    // CHECK: %[[ICMP1:.+]] = icmp slt i32 %exp, 0
+    // CHECK: %[[ICMP2:.+]] = icmp ugt i32 %[[SHIFT_AMOUNT]], 31
+    // CHECK: %[[OVERFLOW:.+]] = or i1 %[[ICMP1]], %[[ICMP2]]
+    // CHECK: %[[POW:.+]] = shl nuw i32 1, %[[SHIFT_AMOUNT]]
+    // CHECK: %[[RET:.+]] = select i1 %[[OVERFLOW]], i32 0, i32 %[[POW]]
+    // CHECK: ret i32 %[[RET]]
+    4u32.pow(exp)
+}
+
+// 16 ** n == 2 ** (4 * n) == 1 << (4 * n)
+// CHECK-LABEL: @pow16
+#[no_mangle]
+pub fn pow16(exp: u32) -> u32 {
+    // CHECK: %[[SHIFT_AMOUNT:.+]] = shl i32 %exp, 2
+    // CHECK: %[[ICMP1:.+]] = icmp ugt i32 %exp, 1073741823
+    // CHECK: %[[ICMP2:.+]] = icmp ugt i32 %[[SHIFT_AMOUNT]], 31
+    // CHECK: %[[OVERFLOW:.+]] = or i1 %[[ICMP1]], %[[ICMP2]]
+    // CHECK: %[[POW:.+]] = shl nuw i32 1, %[[SHIFT_AMOUNT]]
+    // CHECK: %[[RET:.+]] = select i1 %[[OVERFLOW]], i32 0, i32 %[[POW]]
+    // CHECK: ret i32 %[[RET]]
+    16u32.pow(exp)
+}

tests/ui/numbers-arithmetic/overflowing-pow-signed.rs

@@ -1,6 +1,6 @@
 //@ run-fail
 //@ regex-error-pattern: thread 'main'.*panicked
-//@ error-pattern: attempt to multiply with overflow
+//@ regex-error-pattern: attempt to (multiply|shift left) with overflow
 //@ needs-subprocess
 //@ compile-flags: -C debug-assertions
 

tests/ui/numbers-arithmetic/overflowing-pow-unsigned.rs

@@ -1,6 +1,6 @@
 //@ run-fail
 //@ regex-error-pattern: thread 'main'.*panicked
-//@ error-pattern: attempt to multiply with overflow
+//@ regex-error-pattern: attempt to (multiply|shift left) with overflow
 //@ needs-subprocess
 //@ compile-flags: -C debug-assertions