[x264-devel] [PATCH] aarch64: Use ldurb/sturb for loads/stores with negative offset

[Martin Storsjö]

The assembler (both gas and clang/llvm) automatically fixes this,
armasm64 doesn't. We can fix it in gas-preprocessor, but we should
also be using the right instruction form.
---
 common/aarch64/cabac-a.S   |  4 ++--
 common/aarch64/predict-a.S | 18 +++++++++++-------
 2 files changed, 13 insertions(+), 9 deletions(-)

diff --git a/common/aarch64/cabac-a.S b/common/aarch64/cabac-a.S
index c05f963..70e112f 100644
--- a/common/aarch64/cabac-a.S
+++ b/common/aarch64/cabac-a.S
@@ -82,10 +82,10 @@ cabac_putbyte:
 1:
     ldr         x7,  [x0, #CABAC_P]
     asr         w5,  w4,  #8            // carry
-    ldrb        w8,  [x7, #-1]
+    ldurb       w8,  [x7, #-1]
     add         w8,  w8,  w5
     sub         w5,  w5,  #1
-    strb        w8,  [x7, #-1]
+    sturb       w8,  [x7, #-1]
     cbz         w6,  3f
 2:
     subs        w6,  w6,  #1
diff --git a/common/aarch64/predict-a.S b/common/aarch64/predict-a.S
index d287b12..c64d6d1 100644
--- a/common/aarch64/predict-a.S
+++ b/common/aarch64/predict-a.S
@@ -63,7 +63,7 @@ endconst
 
 
 function predict_4x4_h_aarch64, export=1
-    ldrb    w1,  [x0, #0*FDEC_STRIDE-1]
+    ldurb   w1,  [x0, #0*FDEC_STRIDE-1]
     mov     w5,  #0x01010101
     ldrb    w2,  [x0, #1*FDEC_STRIDE-1]
     ldrb    w3,  [x0, #2*FDEC_STRIDE-1]
@@ -80,7 +80,7 @@ function predict_4x4_h_aarch64, export=1
 endfunc
 
 function predict_4x4_v_aarch64, export=1
-    ldr     w1,  [x0, #0 - 1 * FDEC_STRIDE]
+    ldur    w1,  [x0, #0 - 1 * FDEC_STRIDE]
     str     w1,  [x0, #0 + 0 * FDEC_STRIDE]
     str     w1,  [x0, #0 + 1 * FDEC_STRIDE]
     str     w1,  [x0, #0 + 2 * FDEC_STRIDE]
@@ -90,7 +90,7 @@ endfunc
 
 function predict_4x4_dc_neon, export=1
     sub         x1,  x0,  #FDEC_STRIDE
-    ldrb        w4,  [x0, #-1 + 0 * FDEC_STRIDE]
+    ldurb       w4,  [x0, #-1 + 0 * FDEC_STRIDE]
     ldrb        w5,  [x0, #-1 + 1 * FDEC_STRIDE]
     ldrb        w6,  [x0, #-1 + 2 * FDEC_STRIDE]
     ldrb        w7,  [x0, #-1 + 3 * FDEC_STRIDE]
@@ -430,7 +430,7 @@ function predict_8x8c_dc_top_neon, export=1
 endfunc
 
 function predict_8x8c_dc_left_neon, export=1
-    ldrb        w2,  [x0, #0 * FDEC_STRIDE - 1]
+    ldurb       w2,  [x0, #0 * FDEC_STRIDE - 1]
     ldrb        w3,  [x0, #1 * FDEC_STRIDE - 1]
     ldrb        w4,  [x0, #2 * FDEC_STRIDE - 1]
     ldrb        w5,  [x0, #3 * FDEC_STRIDE - 1]
@@ -455,7 +455,7 @@ endfunc
 function predict_8x8c_dc_neon, export=1
     mov         x1,  #FDEC_STRIDE
     sub         x2,  x0,  #FDEC_STRIDE
-    ldrb        w10, [x0, #0 * FDEC_STRIDE - 1]
+    ldurb       w10, [x0, #0 * FDEC_STRIDE - 1]
     ldrb        w11, [x0, #1 * FDEC_STRIDE - 1]
     ldrb        w12, [x0, #2 * FDEC_STRIDE - 1]
     ldrb        w13, [x0, #3 * FDEC_STRIDE - 1]
@@ -511,7 +511,7 @@ function predict_8x8c_h_neon, export=1
 endfunc
 
 function predict_8x8c_v_aarch64, export=1
-    ldr         x1,  [x0, #-FDEC_STRIDE]
+    ldur        x1,  [x0, #-FDEC_STRIDE]
 .irp c, 0,1,2,3,4,5,6,7
     str         x1,  [x0, #\c * FDEC_STRIDE]
 .endr
@@ -568,7 +568,11 @@ endfunc
 
 
 .macro loadsum4 wd, t1, t2, t3, x, idx
+  .if \idx == 0
+    ldurb       \wd,  [\x, #(\idx + 0) * FDEC_STRIDE - 1]
+  .else
     ldrb        \wd,  [\x, #(\idx + 0) * FDEC_STRIDE - 1]
+  .endif
     ldrb        \t1,  [\x, #(\idx + 1) * FDEC_STRIDE - 1]
     ldrb        \t2,  [\x, #(\idx + 2) * FDEC_STRIDE - 1]
     ldrb        \t3,  [\x, #(\idx + 3) * FDEC_STRIDE - 1]
@@ -720,7 +724,7 @@ endfunc
 
 function predict_8x16c_dc_left_neon, export=1
     mov         x1,  #FDEC_STRIDE
-    ldrb        w2,  [x0, # 0 * FDEC_STRIDE - 1]
+    ldurb       w2,  [x0, # 0 * FDEC_STRIDE - 1]
     ldrb        w3,  [x0, # 1 * FDEC_STRIDE - 1]
     ldrb        w4,  [x0, # 2 * FDEC_STRIDE - 1]
     ldrb        w5,  [x0, # 3 * FDEC_STRIDE - 1]
-- 
2.7.4