From d5c0d28527e7141d6b078ab741e84f2a7bd69944 Mon Sep 17 00:00:00 2001
From: Klaus Post <klauspost@gmail.com>
Date: Tue, 8 Mar 2022 03:40:07 -0800
Subject: [PATCH] huff0: Skip unneeded exhaust check (#517)

When not filling there is no reason to check if exhausted.

```
BenchmarkDecompress4XNoTable/gettysburg/10000-32            11867         11550         -2.67%
BenchmarkDecompress4XNoTable/gettysburg/262143-32           336139        331035        -1.52%
BenchmarkDecompress4XNoTable/twain/10000-32                 11922         11561         -3.03%
BenchmarkDecompress4XNoTable/twain/262143-32                387406        374386        -3.36%
BenchmarkDecompress4XNoTable/pngdata.001/10000-32           12262         11961         -2.45%
BenchmarkDecompress4XNoTable/pngdata.001/262143-32          319131        308746        -3.25%
```
---
 huff0/decompress_amd64.s    | 24 ++++++++++++------------
 huff0/decompress_amd64.s.in |  4 ++--
 2 files changed, 14 insertions(+), 14 deletions(-)

diff --git a/huff0/decompress_amd64.s b/huff0/decompress_amd64.s
index 6d883c60dc..6c2b01db44 100644
--- a/huff0/decompress_amd64.s
+++ b/huff0/decompress_amd64.s
@@ -66,14 +66,14 @@ main_loop:
 	SHLQ CL, AX
 	ORQ  AX, br_value
 
-	// }
-skip_fill0:
-
 	// exhausted = exhausted || (br0.off < 4)
 	CMPQ  br_offset, $4
 	SETLT DL
 	ORB   DL, DH
 
+	// }
+skip_fill0:
+
 	// val0 := br0.peekTopBits(peekBits)
 	MOVQ br_value, AX
 	MOVQ peek_bits, CX
@@ -136,14 +136,14 @@ skip_fill0:
 	SHLQ CL, AX
 	ORQ  AX, br_value
 
-	// }
-skip_fill1:
-
 	// exhausted = exhausted || (br1.off < 4)
 	CMPQ  br_offset, $4
 	SETLT DL
 	ORB   DL, DH
 
+	// }
+skip_fill1:
+
 	// val0 := br1.peekTopBits(peekBits)
 	MOVQ br_value, AX
 	MOVQ peek_bits, CX
@@ -206,14 +206,14 @@ skip_fill1:
 	SHLQ CL, AX
 	ORQ  AX, br_value
 
-	// }
-skip_fill2:
-
 	// exhausted = exhausted || (br2.off < 4)
 	CMPQ  br_offset, $4
 	SETLT DL
 	ORB   DL, DH
 
+	// }
+skip_fill2:
+
 	// val0 := br2.peekTopBits(peekBits)
 	MOVQ br_value, AX
 	MOVQ peek_bits, CX
@@ -276,14 +276,14 @@ skip_fill2:
 	SHLQ CL, AX
 	ORQ  AX, br_value
 
-	// }
-skip_fill3:
-
 	// exhausted = exhausted || (br3.off < 4)
 	CMPQ  br_offset, $4
 	SETLT DL
 	ORB   DL, DH
 
+	// }
+skip_fill3:
+
 	// val0 := br3.peekTopBits(peekBits)
 	MOVQ br_value, AX
 	MOVQ peek_bits, CX
diff --git a/huff0/decompress_amd64.s.in b/huff0/decompress_amd64.s.in
index 34bb290b63..d28655489f 100644
--- a/huff0/decompress_amd64.s.in
+++ b/huff0/decompress_amd64.s.in
@@ -66,13 +66,13 @@ main_loop:
     MOVQ    br_bits_read, CX
     SHLQ    CL, AX
     ORQ     AX, br_value
-    // }
-skip_fill{{ var "id" }}:
 
     // exhausted = exhausted || (br{{ var "id"}}.off < 4)
     CMPQ    br_offset, $4
     SETLT   DL
     ORB     DL, DH
+    // }
+skip_fill{{ var "id" }}:
 
     // val0 := br{{ var "id"}}.peekTopBits(peekBits)
     MOVQ    br_value, AX