diff options
Diffstat (limited to 'pixman/pixman/pixman-arm-neon-asm.h')
| -rw-r--r-- | pixman/pixman/pixman-arm-neon-asm.h | 45 |
1 files changed, 32 insertions, 13 deletions
diff --git a/pixman/pixman/pixman-arm-neon-asm.h b/pixman/pixman/pixman-arm-neon-asm.h index 97adc6a87..1673b080f 100644 --- a/pixman/pixman/pixman-arm-neon-asm.h +++ b/pixman/pixman/pixman-arm-neon-asm.h @@ -212,27 +212,39 @@ .macro pixld1_s elem_size, reg1, mem_operand .if elem_size == 16 mov TMP1, VX, asr #16 - add VX, VX, UNIT_X + adds VX, VX, UNIT_X +5: subpls VX, VX, SRC_WIDTH_FIXED + bpl 5b add TMP1, mem_operand, TMP1, asl #1 mov TMP2, VX, asr #16 - add VX, VX, UNIT_X + adds VX, VX, UNIT_X +5: subpls VX, VX, SRC_WIDTH_FIXED + bpl 5b add TMP2, mem_operand, TMP2, asl #1 vld1.16 {d®1&[0]}, [TMP1, :16] mov TMP1, VX, asr #16 - add VX, VX, UNIT_X + adds VX, VX, UNIT_X +5: subpls VX, VX, SRC_WIDTH_FIXED + bpl 5b add TMP1, mem_operand, TMP1, asl #1 vld1.16 {d®1&[1]}, [TMP2, :16] mov TMP2, VX, asr #16 - add VX, VX, UNIT_X + adds VX, VX, UNIT_X +5: subpls VX, VX, SRC_WIDTH_FIXED + bpl 5b add TMP2, mem_operand, TMP2, asl #1 vld1.16 {d®1&[2]}, [TMP1, :16] vld1.16 {d®1&[3]}, [TMP2, :16] .elseif elem_size == 32 mov TMP1, VX, asr #16 - add VX, VX, UNIT_X + adds VX, VX, UNIT_X +5: subpls VX, VX, SRC_WIDTH_FIXED + bpl 5b add TMP1, mem_operand, TMP1, asl #2 mov TMP2, VX, asr #16 - add VX, VX, UNIT_X + adds VX, VX, UNIT_X +5: subpls VX, VX, SRC_WIDTH_FIXED + bpl 5b add TMP2, mem_operand, TMP2, asl #2 vld1.32 {d®1&[0]}, [TMP1, :32] vld1.32 {d®1&[1]}, [TMP2, :32] @@ -242,7 +254,7 @@ .endm .macro pixld2_s elem_size, reg1, reg2, mem_operand -.if elem_size == 32 +.if 0 /* elem_size == 32 */ mov TMP1, VX, asr #16 add VX, VX, UNIT_X, asl #1 add TMP1, mem_operand, TMP1, asl #2 @@ -268,12 +280,16 @@ .macro pixld0_s elem_size, reg1, idx, mem_operand .if elem_size == 16 mov TMP1, VX, asr #16 - add VX, VX, UNIT_X + adds VX, VX, UNIT_X +5: subpls VX, VX, SRC_WIDTH_FIXED + bpl 5b add TMP1, mem_operand, TMP1, asl #1 vld1.16 {d®1&[idx]}, [TMP1, :16] .elseif elem_size == 32 mov TMP1, VX, asr #16 - add VX, VX, UNIT_X + adds VX, VX, UNIT_X +5: subpls VX, VX, SRC_WIDTH_FIXED + bpl 5b add TMP1, mem_operand, TMP1, asl #2 vld1.32 {d®1&[idx]}, [TMP1, :32] .endif @@ -964,15 +980,17 @@ fname: TMP1 .req r4 TMP2 .req r5 DST_R .req r6 + SRC_WIDTH_FIXED .req r7 .macro pixld_src x:vararg pixld_s x .endm ldr UNIT_X, [sp] - push {r4-r6, lr} + push {r4-r8, lr} + ldr SRC_WIDTH_FIXED, [sp, #(24 + 4)] .if mask_bpp != 0 - ldr MASK, [sp, #(16 + 4)] + ldr MASK, [sp, #(24 + 8)] .endif .else /* @@ -1044,7 +1062,7 @@ fname: cleanup .if use_nearest_scaling != 0 - pop {r4-r6, pc} /* exit */ + pop {r4-r8, pc} /* exit */ .else bx lr /* exit */ .endif @@ -1058,7 +1076,7 @@ fname: cleanup .if use_nearest_scaling != 0 - pop {r4-r6, pc} /* exit */ + pop {r4-r8, pc} /* exit */ .unreq DST_R .unreq SRC @@ -1069,6 +1087,7 @@ fname: .unreq TMP2 .unreq DST_W .unreq MASK + .unreq SRC_WIDTH_FIXED .else bx lr /* exit */ |