diff --git a/src/alloy/backend/ivm/ivm_intcode.cc b/src/alloy/backend/ivm/ivm_intcode.cc
index 555a904bf..f38e0f2b8 100644
--- a/src/alloy/backend/ivm/ivm_intcode.cc
+++ b/src/alloy/backend/ivm/ivm_intcode.cc
@@ -3071,12 +3071,10 @@ uint32_t IntCode_PERMUTE_V128_BY_INT32(IntCodeState& ics, const IntCode* i) {
   const vec128_t& src3 = ics.rf[i->src3_reg].v128;
   vec128_t& dest = ics.rf[i->dest_reg].v128;
   for (size_t i = 0; i < 4; i++) {
-#define SWAP_INLINE(x) (((x) & ~0x3) + (3 - ((x) % 4)))
-    size_t m = SWAP_INLINE(i);
-    size_t b = (src1 >> (m * 8)) & 0x3;
-    dest.i4[m] = b < 4 ?
-        src2.i4[SWAP_INLINE(b)] :
-        src3.i4[SWAP_INLINE(b - 4)];
+    size_t b = (src1 >> ((3 - i) * 8)) & 0x7;
+    dest.i4[i] = b < 4 ?
+        src2.i4[b] :
+        src3.i4[b - 4];
   }
   return IA_NEXT;
 }
@@ -3086,12 +3084,10 @@ uint32_t IntCode_PERMUTE_V128_BY_V128(IntCodeState& ics, const IntCode* i) {
   const vec128_t& src3 = ics.rf[i->src3_reg].v128;
   vec128_t& dest = ics.rf[i->dest_reg].v128;
   for (size_t i = 0; i < 16; i++) {
-#define SWAP_INLINE(x) (((x) & ~0x3) + (3 - ((x) % 4)))
-    size_t m = SWAP_INLINE(i);
-    size_t b = src1.b16[m] & 0x1F;
-    dest.b16[m] = b < 16 ?
-        src2.b16[SWAP_INLINE(b)] :
-        src3.b16[SWAP_INLINE(b - 16)];
+    size_t b = src1.b16[i] & 0x1F;
+    dest.b16[i] = b < 16 ?
+        src2.b16[b] :
+        src3.b16[b - 16];
   }
   return IA_NEXT;
 }
diff --git a/src/alloy/frontend/ppc/ppc_emit_altivec.cc b/src/alloy/frontend/ppc/ppc_emit_altivec.cc
index 018f518b7..742d981c6 100644
--- a/src/alloy/frontend/ppc/ppc_emit_altivec.cc
+++ b/src/alloy/frontend/ppc/ppc_emit_altivec.cc
@@ -801,7 +801,7 @@ int InstrEmit_vmrghw_(PPCFunctionBuilder& f, uint32_t vd, uint32_t va, uint32_t
   // (VD.z) = (VA.y)
   // (VD.w) = (VB.y)
   Value* v = f.Permute(
-      f.LoadConstant(0x05010400),
+      f.LoadConstant(0x00040105),
       f.LoadVR(va),
       f.LoadVR(vb),
       INT32_TYPE);
@@ -831,7 +831,7 @@ int InstrEmit_vmrglw_(PPCFunctionBuilder& f, uint32_t vd, uint32_t va, uint32_t
   // (VD.z) = (VA.w)
   // (VD.w) = (VB.w)
   Value* v = f.Permute(
-      f.LoadConstant(0x07030602),
+      f.LoadConstant(0x02060307),
       f.LoadVR(va),
       f.LoadVR(vb),
       INT32_TYPE);
@@ -1164,7 +1164,11 @@ XEEMITTER(vrlw128,        VX128(6, 80),     VX128  )(PPCFunctionBuilder& f, Inst
 XEEMITTER(vrlimi128,      VX128_4(6, 1808), VX128_4)(PPCFunctionBuilder& f, InstrData& i) {
   const uint32_t vd = i.VX128_4.VD128l | (i.VX128_4.VD128h << 5);
   const uint32_t vb = i.VX128_4.VB128l | (i.VX128_4.VB128h << 5);
-  uint32_t blend_mask = i.VX128_4.IMM;
+  uint32_t blend_mask_src = i.VX128_4.IMM;
+  uint32_t blend_mask = 0;
+  for (int n = 0; n < 4; n++) {
+    blend_mask |= ((blend_mask_src >> n) ? n : (n + 4)) << ((3 - n) * 8);
+  }
   uint32_t rotate = i.VX128_4.z;
   // This is just a fancy permute.
   // X Y Z W, rotated left by 2 = Z W X Y
@@ -1193,8 +1197,10 @@ XEEMITTER(vrlimi128,      VX128_4(6, 1808), VX128_4)(PPCFunctionBuilder& f, Inst
   } else {
     v = f.LoadVR(vb);
   }
-  v = f.Permute(
-      f.LoadConstant(blend_mask), v, f.LoadVR(vd), FLOAT32_TYPE);
+  if (blend_mask != 0x00010203) {
+    v = f.Permute(
+        f.LoadConstant(blend_mask), v, f.LoadVR(vd), INT32_TYPE);
+  }
   f.StoreVR(vd, v);
   return 0;
 }
diff --git a/src/alloy/hir/opcodes.h b/src/alloy/hir/opcodes.h
index bbb93c68b..4e32ef80c 100644
--- a/src/alloy/hir/opcodes.h
+++ b/src/alloy/hir/opcodes.h
@@ -49,7 +49,7 @@ enum ArithmeticFlags {
   ARITHMETIC_SET_CARRY = (1 << 1),
 };
 enum Permutes {
-  PERMUTE_XY_ZW = 0x05040100,
+  PERMUTE_XY_ZW = 0x00010405,
 };
 enum Swizzles {
   SWIZZLE_XYZW_TO_XYZW = 0xE4,