diff --git a/src/stub/amd_d_nrv2b.S b/src/stub/amd_d_nrv2b.S
index 07982429..2ebdf5c8 100644
--- a/src/stub/amd_d_nrv2b.S
+++ b/src/stub/amd_d_nrv2b.S
@@ -35,11 +35,11 @@ lit_n2b:
         incq %rdi
 top_n2b:
         movb (%rsi),%dl  # speculate: literal, or bottom 8 bits of offset
-        jnextb1y lit_n2b
+        jnextb1yp lit_n2b
         lea 1(lenq),off  # [len= 0] off= 1
 offmore_n2b:
-        getnextb(off)
-        jnextb0n offmore_n2b
+        getnextbp(off)
+        jnextb0np offmore_n2b
 
         subl $ 3,off; jc len_n2b  # use previous offset
         shll $ 8,off; movzbl %dl,%edx
diff --git a/src/stub/amd_d_nrv2e.S b/src/stub/amd_d_nrv2e.S
index 3cfcd6e5..8bcc29f4 100644
--- a/src/stub/amd_d_nrv2e.S
+++ b/src/stub/amd_d_nrv2e.S
@@ -35,16 +35,16 @@ lit_n2e:
         incq %rdi
 top_n2e:
         movb (%rsi),%dl  # speculate: literal, or bottom 8 bits of offset
-        jnextb1y lit_n2e
+        jnextb1yp lit_n2e
         lea 1(lenq),off  # [len= 0] off= 1
         jmp getoff_n2e
 
 off_n2e:
         dec off
-        getnextb(off)
+        getnextbp(off)
 getoff_n2e:
-        getnextb(off)
-        jnextb0n off_n2e
+        getnextbp(off)
+        jnextb0np off_n2e
 
         subl $ 3,off; jc offprev_n2e
         shll $ 8,off; movzbl %dl,%edx
diff --git a/src/stub/l_lx_elf64amd.S b/src/stub/l_lx_elf64amd.S
index e57adb68..dd418fcb 100644
--- a/src/stub/l_lx_elf64amd.S
+++ b/src/stub/l_lx_elf64amd.S
@@ -102,6 +102,17 @@ ra_setup:
 */
 /* jump on next bit {0,1} with prediction {y==>likely, n==>unlikely} */
 /* Prediction omitted for now. */
+/* On refill: prefetch next byte, for latency reduction on literals and offsets. */
+#define jnextb0np jnextb0yp
+#define jnextb0yp GETBITp; jnc
+#define jnextb1np jnextb1yp
+#define jnextb1yp GETBITp; jc
+#define GETBITp \
+        addl bits,bits; jnz 0f; \
+        movl (%rsi),bits; subq $-4,%rsi; \
+        adcl bits,bits; movb (%rsi),%dl; \
+0:
+/* Same, but without prefetch (not useful for length of match.) */
 #define jnextb0n jnextb0y
 #define jnextb0y GETBIT; jnc
 #define jnextb1n jnextb1y
@@ -109,11 +120,12 @@ ra_setup:
 #define GETBIT \
         addl bits,bits; jnz 0f; \
         movl (%rsi),bits; subq $-4,%rsi; \
-        adcl bits,bits; movb (%rsi),%dl; \
+        adcl bits,bits; \
 0:
 
 /* rotate next bit into bottom bit of reg */
-#define getnextb(reg) call *%r11; adcl reg,reg
+#define getnextbp(reg) call *%r11; adcl reg,reg
+#define getnextb(reg)  getnextbp(reg)
 
         ALIGN(1<<3)
 getbit: