Push clobbered neon registers on the stack

According to the calling convention, the registers q4-q7 should be preserved by functions. The caller (generated by the compiler) could be using those registers anywhere for any intermediate data. Functions that use more than 12 of the qX registers must push the clobbered registers on the stack in order to be able to restore them afterwards. In functions that don't use all 16 registers, but clobber some of the callee saved registers q4-q7, one or more of them are remapped to reduce the number of registers that have to be saved/restored. This incurs a very small (around 0.5%) slowdown in the decoder and encoder.

Martin Storsjö committed 12 years ago

Commit c0118901 parent 811c6475 files · +241 −194Browse files

Showing the changes to codec/processing/src (1 of 5 files). Show all

Unified · Split

codec/processing/src/arm/vaa_calc_neon.S+29−19

 WELS_ASM_FUNC_BEGIN VAACalcSadSsdBgd_neon
 	stmdb sp!, {r0-r12, r14}
+	vpush {q4-q7}
-	ldr r4, [sp, #56] //r4 keeps the pic_stride
+	ldr r4, [sp, #120] //r4 keeps the pic_stride
 	sub r5, r4, #1
 	lsl r5, r5, #4 //r5 keeps the little step
 	sub r6, r2, r6	//r6 keeps the big step
-	ldr r8, [sp, #64]//psad8x8
+	ldr r8, [sp, #128]//psad8x8
-	ldr r9, [sp, #68]//psum16x16
+	ldr r9, [sp, #132]//psum16x16
-	ldr r10, [sp, #72]//psqsum16x16
+	ldr r10, [sp, #136]//psqsum16x16
-	ldr r11, [sp, #76]//psqdiff16x16
+	ldr r11, [sp, #140]//psqdiff16x16
-	ldr r12, [sp, #80]//p_sd8x8
+	ldr r12, [sp, #144]//p_sd8x8
-	ldr r14, [sp, #84]//p_mad8x8
+	ldr r14, [sp, #148]//p_mad8x8
 	vmov.i8 q12, #0
 bne vaa_calc_sad_ssd_bgd_height_loop
 	//psadframe
-	ldr r7, [sp, #60]//psadframe
+	ldr r7, [sp, #124]//psadframe
 	vadd.i32 d24, d24, d25
 	vst1.32 {d24[0]}, [r7]
+	vpop {q4-q7}
 	ldmia sp!, {r0-r12, r14}
 WELS_ASM_FUNC_END
 WELS_ASM_FUNC_BEGIN VAACalcSadVar_neon
 	stmdb sp!, {r4-r11}
+	vpush {q4}
+	vpush {q6-q7}
-	ldr r4, [sp, #32] //r4 keeps the pic_stride
+	ldr r4, [sp, #80] //r4 keeps the pic_stride
 	sub r5, r4, #1
 	lsl r5, r5, #4 //r5 keeps the little step
 	lsl r6, r4, #4
 	sub r6, r2, r6	//r6 keeps the big step
-	ldr r7,		[sp, #36]	//psadframe
+	ldr r7,		[sp, #84]	//psadframe
-	ldr r8,		[sp, #40]	//psad8x8
+	ldr r8,		[sp, #88]	//psad8x8
-	ldr r9,		[sp, #44]	//psum16x16
+	ldr r9,		[sp, #92]	//psum16x16
-	ldr r10,	[sp, #48]	//psqsum16x16
+	ldr r10,	[sp, #96]	//psqsum16x16
 	vmov.i8 q12, #0
 vaa_calc_sad_var_height_loop:
 	vadd.i32 d24, d24, d25
 	vst1.32 {d24[0]}, [r7]
+	vpop {q6-q7}
+	vpop {q4}
 	ldmia sp!, {r4-r11}
 WELS_ASM_FUNC_END
 WELS_ASM_FUNC_BEGIN VAACalcSadSsd_neon
 	stmdb sp!, {r4-r12}
+	vpush {q4}
+	vpush {q6-q7}
-	ldr r4, [sp, #36] //r4 keeps the pic_stride
+	ldr r4, [sp, #84] //r4 keeps the pic_stride
 	sub r5, r4, #1
 	lsl r5, r5, #4 //r5 keeps the little step
 	lsl r6, r4, #4
 	sub r6, r2, r6	//r6 keeps the big step
-	ldr r7,		[sp, #40]	//psadframe
+	ldr r7,		[sp, #88]	//psadframe
-	ldr r8,		[sp, #44]	//psad8x8
+	ldr r8,		[sp, #92]	//psad8x8
-	ldr r9,		[sp, #48]	//psum16x16
+	ldr r9,		[sp, #96]	//psum16x16
-	ldr r10,	[sp, #52]	//psqsum16x16
+	ldr r10,	[sp, #100]	//psqsum16x16
-	ldr r11,	[sp, #56]	//psqdiff16x16
+	ldr r11,	[sp, #104]	//psqdiff16x16
 	vmov.i8 q12, #0
 vaa_calc_sad_ssd_height_loop:
 	vadd.i32 d24, d24, d25
 	vst1.32 {d24[0]}, [r7]
+	vpop {q6-q7}
+	vpop {q4}
 	ldmia sp!, {r4-r12}
 WELS_ASM_FUNC_END

Push clobbered neon registers on the stack

Keyboard shortcuts

General

Issues and pull requests

Issues, pull requests and Inbox

Repository code