diff --git a/basis/cpu/ppc/bootstrap.factor b/basis/cpu/ppc/bootstrap.factor
index 6a00dec12f..b09938f4b9 100644
--- a/basis/cpu/ppc/bootstrap.factor
+++ b/basis/cpu/ppc/bootstrap.factor
@@ -21,43 +21,48 @@ CONSTANT: rs-reg 14
 : xt-save ( -- n ) stack-frame 2 bootstrap-cells - ;
 
 [
-    0 6 LOAD32 rc-absolute-ppc-2/2 rt-immediate jit-rel
-    11 6 profile-count-offset LWZ
+    0 3 LOAD32 rc-absolute-ppc-2/2 rt-immediate jit-rel
+    11 3 profile-count-offset LWZ
     11 11 1 tag-fixnum ADDI
-    11 6 profile-count-offset STW
-    11 6 word-code-offset LWZ
+    11 3 profile-count-offset STW
+    11 3 word-code-offset LWZ
     11 11 compiled-header-size ADDI
     11 MTCTR
     BCTR
 ] jit-profiling jit-define
 
 [
-    0 6 LOAD32 rc-absolute-ppc-2/2 rt-this jit-rel
+    0 3 LOAD32 rc-absolute-ppc-2/2 rt-this jit-rel
     0 MFLR
     1 1 stack-frame SUBI
-    6 1 xt-save STW
-    stack-frame 6 LI
-    6 1 next-save STW
+    3 1 xt-save STW
+    stack-frame 3 LI
+    3 1 next-save STW
     0 1 lr-save stack-frame + STW
 ] jit-prolog jit-define
 
 [
-    0 6 LOAD32 rc-absolute-ppc-2/2 rt-immediate jit-rel
-    6 ds-reg 4 STWU
+    0 3 LOAD32 rc-absolute-ppc-2/2 rt-immediate jit-rel
+    3 ds-reg 4 STWU
 ] jit-push-immediate jit-define
 
 [
-    0 6 LOAD32 rc-absolute-ppc-2/2 rt-stack-chain jit-rel
-    7 6 0 LWZ
-    1 7 0 STW
-    0 6 LOAD32 rc-absolute-ppc-2/2 rt-primitive jit-rel
-    6 MTCTR
+    0 3 LOAD32 rc-absolute-ppc-2/2 rt-stack-chain jit-rel
+    4 3 0 LWZ
+    1 4 0 STW
+    0 3 LOAD32 rc-absolute-ppc-2/2 rt-primitive jit-rel
+    3 MTCTR
     BCTR
 ] jit-primitive jit-define
 
 [ 0 BL rc-relative-ppc-3 rt-xt-pic jit-rel ] jit-word-call jit-define
 
-[ 0 B rc-relative-ppc-3 rt-xt jit-rel ] jit-word-jump jit-define
+[
+    0 6 LOAD32 rc-absolute-ppc-2/2 rt-here jit-rel
+    0 B rc-relative-ppc-3 rt-xt-pic-tail jit-rel
+] jit-word-jump jit-define
+
+[ 0 B rc-relative-ppc-3 rt-xt jit-rel ] jit-word-special jit-define
 
 [
     3 ds-reg 0 LWZ
@@ -152,6 +157,9 @@ CONSTANT: rs-reg 14
 
 ! ! ! Polymorphic inline caches
 
+! Don't touch r6 here; it's used to pass the tail call site
+! address for tail PICs
+
 ! Load a value from a stack position
 [
     4 ds-reg 0 LWZ rc-absolute-ppc-2 rt-untagged jit-rel
@@ -225,7 +233,7 @@ CONSTANT: rs-reg 14
     ! if(get(cache) == class)
     6 3 0 LWZ
     6 0 4 CMP
-    5 BNE
+    10 BNE
     ! megamorphic_cache_hits++
     0 4 LOAD32 rc-absolute-ppc-2/2 rt-megamorphic-cache-hits jit-rel
     5 4 0 LWZ
diff --git a/basis/cpu/ppc/ppc.factor b/basis/cpu/ppc/ppc.factor
index a11b0daa86..beee48e5ea 100644
--- a/basis/cpu/ppc/ppc.factor
+++ b/basis/cpu/ppc/ppc.factor
@@ -3,9 +3,10 @@
 USING: accessors assocs sequences kernel combinators make math
 math.order math.ranges system namespaces locals layouts words
 alien alien.c-types literals cpu.architecture cpu.ppc.assembler
-literals compiler.cfg.registers compiler.cfg.instructions
-compiler.constants compiler.codegen compiler.codegen.fixup
-compiler.cfg.intrinsics compiler.cfg.stack-frame ;
+cpu.ppc.assembler.backend literals compiler.cfg.registers
+compiler.cfg.instructions compiler.constants compiler.codegen
+compiler.codegen.fixup compiler.cfg.intrinsics
+compiler.cfg.stack-frame ;
 IN: cpu.ppc
 
 ! PowerPC register assignments:
@@ -116,7 +117,7 @@ M: ppc stack-frame-size ( stack-frame -- i )
 M: ppc %call ( word -- ) 0 BL rc-relative-ppc-3 rel-word-pic ;
 
 M: ppc %jump ( word -- )
-    0 3 LOAD32 rc-absolute-ppc-2/2 rel-here
+    0 6 LOAD32 8 rc-absolute-ppc-2/2 rel-here
     0 B rc-relative-ppc-3 rel-word-pic-tail ;
 
 M: ppc %jump-label ( label -- ) B ;
@@ -130,7 +131,7 @@ M:: ppc %dispatch ( src temp offset -- )
     BCTR ;
 
 M: ppc %dispatch-label ( word -- )
-    0 , rc-absolute-cell rel-word ;
+    B{ 0 0 0 0 } % rc-absolute-cell rel-word ;
 
 :: (%slot) ( obj slot tag temp -- reg offset )
     temp slot obj ADD
diff --git a/basis/cpu/x86/bootstrap.factor b/basis/cpu/x86/bootstrap.factor
index 994591adcf..474ce2ea46 100644
--- a/basis/cpu/x86/bootstrap.factor
+++ b/basis/cpu/x86/bootstrap.factor
@@ -233,7 +233,7 @@ big-endian off
     temp0 temp2 ADD
     ! if(get(cache) == class)
     temp0 [] temp1 CMP
-    bootstrap-cell 4 = 14 18 ? JNE ! Yuck!
+    bootstrap-cell 4 = 14 22 ? JNE ! Yuck!
     ! megamorphic_cache_hits++
     temp1 0 MOV rc-absolute-cell rt-megamorphic-cache-hits jit-rel
     temp1 [] 1 ADD
diff --git a/vm/cpu-ppc.S b/vm/cpu-ppc.S
index f8dad4b2b2..a372b2b1f5 100755
--- a/vm/cpu-ppc.S
+++ b/vm/cpu-ppc.S
@@ -236,8 +236,10 @@ DEF(void,flush_icache,(void *start, int len)):
 	blr
 
 DEF(void,primitive_inline_cache_miss,(void)):
-    mflr r3
+    mflr r6
+DEF(void,primitive_inline_cache_miss_tail,(void)):
     PROLOGUE
+    mr r3,r6
     bl MANGLE(inline_cache_miss)
     EPILOGUE
     mtctr r3
diff --git a/vm/cpu-ppc.hpp b/vm/cpu-ppc.hpp
index d393223d8d..ae7f93ebf7 100755
--- a/vm/cpu-ppc.hpp
+++ b/vm/cpu-ppc.hpp
@@ -7,11 +7,22 @@ namespace factor
 register cell ds asm("r13");
 register cell rs asm("r14");
 
+/* In the instruction sequence:
+
+   LOAD32 r3,...
+   B blah
+
+   the offset from the immediate operand to LOAD32 to the instruction after
+   the branch is two instructions. */
+static const fixnum xt_tail_pic_offset = 4 * 2;
+
 inline static void check_call_site(cell return_address)
 {
 #ifdef FACTOR_DEBUG
 	cell insn = *(cell *)return_address;
-	assert((insn & 0x3) == 0x1);
+	/* Check that absolute bit is 0 */
+	assert((insn & 0x2) == 0x0);
+	/* Check that instruction is branch */
 	assert((insn >> 26) == 0x12);
 #endif
 }
@@ -21,8 +32,8 @@ inline static void check_call_site(cell return_address)
 inline static void *get_call_target(cell return_address)
 {
 	return_address -= sizeof(cell);
-
 	check_call_site(return_address);
+
 	cell insn = *(cell *)return_address;
 	cell unsigned_addr = (insn & B_MASK);
 	fixnum signed_addr = (fixnum)(unsigned_addr << 6) >> 6;
@@ -32,19 +43,25 @@ inline static void *get_call_target(cell return_address)
 inline static void set_call_target(cell return_address, void *target)
 {
 	return_address -= sizeof(cell);
-
-#ifdef FACTOR_DEBUG
-	assert((return_address & ~B_MASK) == 0);
 	check_call_site(return_address);
-#endif
+
 	cell insn = *(cell *)return_address;
-	insn = ((insn & ~B_MASK) | (((cell)target - return_address) & B_MASK));
+
+	fixnum relative_address = ((cell)target - return_address);
+	insn = ((insn & ~B_MASK) | (relative_address & B_MASK));
 	*(cell *)return_address = insn;
 
 	/* Flush the cache line containing the call we just patched */
 	__asm__ __volatile__ ("icbi 0, %0\n" "sync\n"::"r" (return_address):);
 }
 
+inline static bool tail_call_site_p(cell return_address)
+{
+	return_address -= sizeof(cell);
+	cell insn = *(cell *)return_address;
+	return (insn & 0x1) == 0;
+}
+
 /* Defined in assembly */
 VM_ASM_API void c_to_factor(cell quot);
 VM_ASM_API void throw_impl(cell quot, stack_frame *rewind);