diff --git a/README.md b/README.md
index 37d7ab2e..fde542ee 100644
--- a/README.md
+++ b/README.md
@@ -250,7 +250,7 @@ These builtins are needed to support `f16` and `f128`, which are in the process
 - [ ] floatunditf.c
 - [ ] floatunsitf.c
 - [x] multf3.c
-- [ ] powitf2.c
+- [x] powitf2.c
 - [ ] ppc/fixtfdi.c
 - [ ] ppc/fixunstfdi.c
 - [ ] ppc/floatditf.c
diff --git a/build.rs b/build.rs
index ec830ecb..60dc8715 100644
--- a/build.rs
+++ b/build.rs
@@ -544,7 +544,6 @@ mod c {
                 ("__floatunditf", "floatunditf.c"),
                 ("__floatunsitf", "floatunsitf.c"),
                 ("__divtf3", "divtf3.c"),
-                ("__powitf2", "powitf2.c"),
                 ("__fe_getround", "fp_mode.c"),
                 ("__fe_raise_inexact", "fp_mode.c"),
             ]);
diff --git a/src/float/pow.rs b/src/float/pow.rs
index 3103fe6f..1e0f2289 100644
--- a/src/float/pow.rs
+++ b/src/float/pow.rs
@@ -35,4 +35,16 @@ intrinsics! {
     pub extern "C" fn __powidf2(a: f64, b: i32) -> f64 {
         pow(a, b)
     }
+
+    #[avr_skip]
+    #[cfg(not(any(feature = "no-f16-f128", target_arch = "powerpc", target_arch = "powerpc64")))]
+    pub extern "C" fn __powitf2(a: f128, b: i32) -> f128 {
+        pow(a, b)
+    }
+
+    #[avr_skip]
+    #[cfg(all(not(feature = "no-f16-f128"), any(target_arch = "powerpc", target_arch = "powerpc64")))]
+    pub extern "C" fn __powikf2(a: f128, b: i32) -> f128 {
+        pow(a, b)
+    }
 }
diff --git a/testcrate/tests/misc.rs b/testcrate/tests/misc.rs
index cdc37e2a..2fca23c6 100644
--- a/testcrate/tests/misc.rs
+++ b/testcrate/tests/misc.rs
@@ -1,5 +1,6 @@
 // makes configuration easier
 #![allow(unused_macros)]
+#![feature(f128)]
 
 use testcrate::*;
 
@@ -96,7 +97,7 @@ fn leading_zeros() {
 // https://github.com/rust-lang/rust/issues/73920.
 // TODO how do we resolve this indeterminacy?
 macro_rules! pow {
-    ($($f:ty, $tolerance:expr, $fn:ident);*;) => {
+    ($($f:ty, $tolerance:expr, $fn:ident);* ;) => {
         $(
             fuzz_float_2(N, |x: $f, y: $f| {
                 if !(Float::is_subnormal(x) || Float::is_subnormal(y) || x.is_nan()) {
@@ -118,12 +119,12 @@ macro_rules! pow {
                             b < $tolerance
                         } else {
                             let quo = b / a;
-                            (quo < (1. + $tolerance)) && (quo > (1. - $tolerance))
+                            (quo < (1. + black_box($tolerance))) && (quo > (1. - black_box($tolerance)))
                         }
                     };
                     if !good {
                         panic!(
-                            "{}({}, {}): std: {}, builtins: {}",
+                            "{}({:?}, {:?}): std: {:?}, builtins: {:?}",
                             stringify!($fn), x, n, tmp0, tmp1
                         );
                     }
@@ -139,8 +140,27 @@ fn float_pow() {
     use compiler_builtins::float::pow::{__powidf2, __powisf2};
     use compiler_builtins::float::Float;
 
+    fn black_box<T>(val: T) -> T {
+        val
+    }
+
     pow!(
         f32, 1e-4, __powisf2;
         f64, 1e-12, __powidf2;
     );
+
+    // There is no apfloat fallback for `powi` so we just skip these tests on these platforms
+    #[cfg(not(any(feature = "no-f16-f128", feature = "no-sys-f128")))]
+    {
+        #[cfg(any(target_arch = "powerpc", target_arch = "powerpc64"))]
+        use compiler_builtins::float::pow::__powikf2;
+        #[cfg(not(any(target_arch = "powerpc", target_arch = "powerpc64")))]
+        use compiler_builtins::float::pow::__powitf2;
+
+        // FIXME(f16_f128): we do this to block const eval which currently ICEs. Change this
+        // once it works correctly.
+        use core::hint::black_box;
+
+        pow!(f128, 1e-24, __powitf2;);
+    }
 }