diff --git a/results/MultiSource/Applications/JM/ldecod/CMakeFiles/ldecod.dir/block.s b/results/MultiSource/Applications/JM/ldecod/CMakeFiles/ldecod.dir/block.s
index 3f22d57..f3bffcb 100644
--- a/results/MultiSource/Applications/JM/ldecod/CMakeFiles/ldecod.dir/block.s
+++ b/results/MultiSource/Applications/JM/ldecod/CMakeFiles/ldecod.dir/block.s
@@ -7635,7 +7635,6 @@ itrans_sp_chroma:                       # @itrans_sp_chroma
 # %bb.8:                                # %.split301.us
 	slli.d	$a7, $a1, 8
 	add.d	$a7, $a0, $a7
-	vneg.w	$vr1, $vr0
 	ori	$t8, $zero, 2664
 	ldx.w	$t8, $a7, $t8
 	ori	$fp, $zero, 2728
@@ -7644,15 +7643,14 @@ itrans_sp_chroma:                       # @itrans_sp_chroma
 	ldx.w	$s0, $a7, $s0
 	ori	$s1, $zero, 3496
 	ldx.w	$a7, $a7, $s1
-	vldrepl.w	$vr2, $t7, 0
-	vmax.w	$vr1, $vr0, $vr1
+	vldrepl.w	$vr1, $t7, 0
+	vsigncov.w	$vr2, $vr0, $vr0
 	vslti.w	$vr0, $vr0, 0
 	vreplgr2vr.w	$vr3, $t6
-	vmadd.w	$vr3, $vr1, $vr2
+	vmadd.w	$vr3, $vr2, $vr1
 	vreplgr2vr.w	$vr1, $t5
 	vsra.w	$vr1, $vr3, $vr1
-	vneg.w	$vr2, $vr1
-	vmax.w	$vr1, $vr1, $vr2
+	vsigncov.w	$vr1, $vr1, $vr1
 	vneg.w	$vr2, $vr1
 	vbitsel.v	$vr0, $vr1, $vr2, $vr0
 	vinsgr2vr.w	$vr1, $a7, 0
@@ -7661,7 +7659,6 @@ itrans_sp_chroma:                       # @itrans_sp_chroma
 	vinsgr2vr.w	$vr1, $s0, 2
 	vinsgr2vr.w	$vr1, $t8, 3
 	vadd.w	$vr0, $vr0, $vr1
-	vmul.w	$vr0, $vr0, $vr2
 	b	.LBB9_10
 .LBB9_9:                                # %.split301
 	ldx.w	$a7, $t2, $a6
@@ -7685,22 +7682,20 @@ itrans_sp_chroma:                       # @itrans_sp_chroma
 	vreplgr2vr.w	$vr2, $t0
 	vsll.w	$vr1, $vr1, $vr2
 	vsrai.w	$vr1, $vr1, 5
+	vldrepl.w	$vr2, $t7, 0
 	vadd.w	$vr0, $vr1, $vr0
-	vldrepl.w	$vr1, $t7, 0
-	vneg.w	$vr2, $vr0
-	vmax.w	$vr2, $vr0, $vr2
+	vsigncov.w	$vr1, $vr0, $vr0
 	vreplgr2vr.w	$vr3, $t6
-	vmadd.w	$vr3, $vr2, $vr1
+	vmadd.w	$vr3, $vr1, $vr2
 	vreplgr2vr.w	$vr1, $t5
 	vsra.w	$vr1, $vr3, $vr1
 	vslti.w	$vr0, $vr0, 0
-	vneg.w	$vr2, $vr1
-	vldrepl.w	$vr3, $t4, 0
-	vmax.w	$vr1, $vr1, $vr2
-	vneg.w	$vr2, $vr1
-	vbitsel.v	$vr0, $vr1, $vr2, $vr0
-	vmul.w	$vr0, $vr0, $vr3
+	vldrepl.w	$vr2, $t4, 0
+	vsigncov.w	$vr1, $vr1, $vr1
+	vneg.w	$vr3, $vr1
+	vbitsel.v	$vr0, $vr1, $vr3, $vr0
 .LBB9_10:                               # %.preheader278
+	vmul.w	$vr0, $vr0, $vr2
 	add.d	$a6, $t2, $a6
 	addi.d	$t1, $t1, 12
 	add.d	$a7, $t2, $t1
diff --git a/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/macroblock.s b/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/macroblock.s
index af292b3..f2ba686 100644
--- a/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/macroblock.s
+++ b/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/macroblock.s
@@ -16985,9 +16985,9 @@ find_sad_16x16:                         # @find_sad_16x16
 	vadd.w	$vr15, $vr14, $vr13
 	vsub.w	$vr14, $vr13, $vr14
 	vsub.w	$vr8, $vr8, $vr10
-	vadd.w	$vr16, $vr15, $vr11
+	vadd.w	$vr17, $vr15, $vr11
 	vsub.w	$vr13, $vr11, $vr15
-	vadd.w	$vr18, $vr14, $vr8
+	vadd.w	$vr16, $vr14, $vr8
 	vsub.w	$vr8, $vr8, $vr14
 	ld.w	$a0, $a2, -960
 	ld.w	$a4, $a2, -896
@@ -17018,21 +17018,21 @@ find_sad_16x16:                         # @find_sad_16x16
 	ld.w	$a4, $a2, -384
 	ld.w	$a5, $a2, -320
 	ld.w	$t0, $a2, -256
-	vinsgr2vr.w	$vr17, $a0, 0
-	vinsgr2vr.w	$vr17, $a4, 1
-	vinsgr2vr.w	$vr17, $a5, 2
-	vinsgr2vr.w	$vr17, $t0, 3
-	vadd.w	$vr19, $vr17, $vr15
-	vsub.w	$vr15, $vr15, $vr17
+	vinsgr2vr.w	$vr18, $a0, 0
+	vinsgr2vr.w	$vr18, $a4, 1
+	vinsgr2vr.w	$vr18, $a5, 2
+	vinsgr2vr.w	$vr18, $t0, 3
+	vadd.w	$vr19, $vr18, $vr15
+	vsub.w	$vr15, $vr15, $vr18
 	vsub.w	$vr10, $vr10, $vr11
-	vadd.w	$vr17, $vr19, $vr14
+	vadd.w	$vr18, $vr19, $vr14
 	vsub.w	$vr19, $vr14, $vr19
 	vadd.w	$vr20, $vr15, $vr10
 	vsub.w	$vr15, $vr10, $vr15
-	vadd.w	$vr10, $vr17, $vr7
-	vadd.w	$vr11, $vr16, $vr12
-	vsub.w	$vr12, $vr12, $vr16
-	vsub.w	$vr7, $vr7, $vr17
+	vadd.w	$vr10, $vr18, $vr7
+	vadd.w	$vr11, $vr17, $vr12
+	vsub.w	$vr12, $vr12, $vr17
+	vsub.w	$vr7, $vr7, $vr18
 	vadd.w	$vr14, $vr11, $vr10
 	vpickve2gr.w	$a0, $vr14, 0
 	st.w	$a0, $a2, -1008
@@ -17060,20 +17060,20 @@ find_sad_16x16:                         # @find_sad_16x16
 	st.w	$a0, $a2, -864
 	vpickve2gr.w	$a0, $vr14, 3
 	st.w	$a0, $a2, -800
-	vsub.w	$vr16, $vr7, $vr12
-	vpickve2gr.w	$a0, $vr16, 0
+	vsub.w	$vr17, $vr7, $vr12
+	vpickve2gr.w	$a0, $vr17, 0
 	st.w	$a0, $a2, -960
-	vpickve2gr.w	$a0, $vr16, 1
+	vpickve2gr.w	$a0, $vr17, 1
 	st.w	$a0, $a2, -896
-	vpickve2gr.w	$a0, $vr16, 2
+	vpickve2gr.w	$a0, $vr17, 2
 	st.w	$a0, $a2, -832
-	vpickve2gr.w	$a0, $vr16, 3
+	vpickve2gr.w	$a0, $vr17, 3
 	st.w	$a0, $a2, -768
-	vadd.w	$vr16, $vr20, $vr5
-	vadd.w	$vr17, $vr18, $vr9
-	vsub.w	$vr9, $vr9, $vr18
-	vsub.w	$vr18, $vr5, $vr20
-	vadd.w	$vr5, $vr17, $vr16
+	vadd.w	$vr17, $vr20, $vr5
+	vadd.w	$vr18, $vr16, $vr9
+	vsub.w	$vr9, $vr9, $vr16
+	vsub.w	$vr16, $vr5, $vr20
+	vadd.w	$vr5, $vr18, $vr17
 	vpickve2gr.w	$a0, $vr5, 0
 	st.w	$a0, $a2, -752
 	vpickve2gr.w	$a0, $vr5, 1
@@ -17082,7 +17082,7 @@ find_sad_16x16:                         # @find_sad_16x16
 	st.w	$a0, $a2, -624
 	vpickve2gr.w	$a0, $vr5, 3
 	st.w	$a0, $a2, -560
-	vsub.w	$vr20, $vr16, $vr17
+	vsub.w	$vr20, $vr17, $vr18
 	vpickve2gr.w	$a0, $vr20, 0
 	st.w	$a0, $a2, -720
 	vpickve2gr.w	$a0, $vr20, 1
@@ -17091,7 +17091,7 @@ find_sad_16x16:                         # @find_sad_16x16
 	st.w	$a0, $a2, -592
 	vpickve2gr.w	$a0, $vr20, 3
 	st.w	$a0, $a2, -528
-	vadd.w	$vr20, $vr9, $vr18
+	vadd.w	$vr20, $vr9, $vr16
 	vpickve2gr.w	$a0, $vr20, 0
 	st.w	$a0, $a2, -736
 	vpickve2gr.w	$a0, $vr20, 1
@@ -17100,7 +17100,7 @@ find_sad_16x16:                         # @find_sad_16x16
 	st.w	$a0, $a2, -608
 	vpickve2gr.w	$a0, $vr20, 3
 	st.w	$a0, $a2, -544
-	vsub.w	$vr21, $vr18, $vr9
+	vsub.w	$vr21, $vr16, $vr9
 	vpickve2gr.w	$a0, $vr21, 0
 	st.w	$a0, $a2, -704
 	vpickve2gr.w	$a0, $vr21, 1
@@ -17158,26 +17158,19 @@ find_sad_16x16:                         # @find_sad_16x16
 	vstelm.w	$vr25, $a2, 0, 3
 	vabsd.w	$vr2, $vr2, $vr4
 	vabsd.w	$vr4, $vr23, $vr24
-	vneg.w	$vr23, $vr15
-	vmax.w	$vr15, $vr15, $vr23
-	vneg.w	$vr23, $vr8
-	vmax.w	$vr8, $vr8, $vr23
+	vsigncov.w	$vr15, $vr15, $vr15
+	vsigncov.w	$vr8, $vr8, $vr8
 	vabsd.w	$vr3, $vr3, $vr6
 	vabsd.w	$vr6, $vr21, $vr22
-	vneg.w	$vr21, $vr19
-	vmax.w	$vr19, $vr19, $vr21
-	vneg.w	$vr21, $vr13
-	vmax.w	$vr13, $vr13, $vr21
-	vabsd.w	$vr9, $vr18, $vr9
-	vabsd.w	$vr16, $vr16, $vr17
-	vneg.w	$vr17, $vr20
-	vmax.w	$vr17, $vr20, $vr17
-	vneg.w	$vr18, $vr5
-	vmax.w	$vr5, $vr5, $vr18
+	vsigncov.w	$vr19, $vr19, $vr19
+	vsigncov.w	$vr13, $vr13, $vr13
+	vabsd.w	$vr9, $vr16, $vr9
+	vabsd.w	$vr16, $vr17, $vr18
+	vsigncov.w	$vr17, $vr20, $vr20
+	vsigncov.w	$vr5, $vr5, $vr5
 	vabsd.w	$vr7, $vr7, $vr12
 	vabsd.w	$vr10, $vr10, $vr11
-	vneg.w	$vr11, $vr14
-	vmax.w	$vr11, $vr14, $vr11
+	vsigncov.w	$vr11, $vr14, $vr14
 	vadd.w	$vr1, $vr11, $vr1
 	vadd.w	$vr1, $vr10, $vr1
 	vadd.w	$vr1, $vr7, $vr1
diff --git a/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/me_distortion.s b/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/me_distortion.s
index 2750aa4..7439699 100644
--- a/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/me_distortion.s
+++ b/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/me_distortion.s
@@ -753,18 +753,12 @@ HadamardSAD8x8:                         # @HadamardSAD8x8
 	xvinsgr2vr.w	$xr0, $a3, 5
 	xvinsgr2vr.w	$xr0, $a2, 6
 	ld.w	$a0, $a0, 224
-	xvneg.w	$xr8, $xr1
-	xvmax.w	$xr1, $xr1, $xr8
-	xvneg.w	$xr8, $xr2
-	xvmax.w	$xr2, $xr2, $xr8
-	xvneg.w	$xr8, $xr3
-	xvmax.w	$xr3, $xr3, $xr8
-	xvneg.w	$xr8, $xr4
-	xvmax.w	$xr4, $xr4, $xr8
-	xvneg.w	$xr8, $xr5
-	xvmax.w	$xr5, $xr5, $xr8
-	xvneg.w	$xr8, $xr6
-	xvmax.w	$xr6, $xr6, $xr8
+	xvsigncov.w	$xr1, $xr1, $xr1
+	xvsigncov.w	$xr2, $xr2, $xr2
+	xvsigncov.w	$xr3, $xr3, $xr3
+	xvsigncov.w	$xr4, $xr4, $xr4
+	xvsigncov.w	$xr5, $xr5, $xr5
+	xvsigncov.w	$xr6, $xr6, $xr6
 	xvinsgr2vr.w	$xr0, $a0, 7
 	xvadda.w	$xr0, $xr7, $xr0
 	xvadd.w	$xr0, $xr6, $xr0
diff --git a/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/me_epzs.s b/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/me_epzs.s
index 1ed2c41..7486c0f 100644
--- a/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/me_epzs.s
+++ b/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/me_epzs.s
@@ -1814,8 +1814,7 @@ EPZSSliceInit:                          # @EPZSSliceInit
 	vsrli.b	$vr12, $vr11, 7
 	vadd.b	$vr11, $vr11, $vr12
 	vsrai.b	$vr11, $vr11, 1
-	vneg.b	$vr12, $vr11
-	vmax.b	$vr11, $vr11, $vr12
+	vsigncov.b	$vr11, $vr11, $vr11
 	vilvl.b	$vr11, $vr5, $vr11
 	vbitseti.h	$vr11, $vr11, 14
 	vpickev.h	$vr9, $vr9, $vr9
diff --git a/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/ratectl.s b/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/ratectl.s
index 46c34fa..ac66918 100644
--- a/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/ratectl.s
+++ b/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/ratectl.s
@@ -53,8 +53,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 524
 	ld.w	$a3, $a2, 588
 	ld.w	$a4, $a2, 652
@@ -71,8 +70,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 528
 	ld.w	$a3, $a2, 592
 	ld.w	$a4, $a2, 656
@@ -89,8 +87,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 532
 	ld.w	$a3, $a2, 596
 	ld.w	$a4, $a2, 660
@@ -107,8 +104,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 536
 	ld.w	$a3, $a2, 600
 	ld.w	$a4, $a2, 664
@@ -125,8 +121,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 540
 	ld.w	$a3, $a2, 604
 	ld.w	$a4, $a2, 668
@@ -143,8 +138,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 544
 	ld.w	$a3, $a2, 608
 	ld.w	$a4, $a2, 672
@@ -161,8 +155,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 548
 	ld.w	$a3, $a2, 612
 	ld.w	$a4, $a2, 676
@@ -179,8 +172,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 552
 	ld.w	$a3, $a2, 616
 	ld.w	$a4, $a2, 680
@@ -197,8 +189,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 556
 	ld.w	$a3, $a2, 620
 	ld.w	$a4, $a2, 684
@@ -215,8 +206,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 560
 	ld.w	$a3, $a2, 624
 	ld.w	$a4, $a2, 688
@@ -233,8 +223,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 564
 	ld.w	$a3, $a2, 628
 	ld.w	$a4, $a2, 692
@@ -251,8 +240,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 568
 	ld.w	$a3, $a2, 632
 	ld.w	$a4, $a2, 696
@@ -269,8 +257,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr1, $xr2
-	xvmax.w	$xr1, $xr2, $xr1
+	xvsigncov.w	$xr1, $xr2, $xr2
 	xvld	$xr2, $a2, 572
 	ld.w	$a3, $a2, 636
 	ld.w	$a4, $a2, 700
@@ -287,11 +274,10 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr2, $a4, 5
 	xvinsgr2vr.w	$xr2, $a5, 6
 	xvinsgr2vr.w	$xr2, $a6, 7
-	xvneg.w	$xr3, $xr2
 	xvld	$xr1, $a2, 0
 	ld.w	$a3, $a2, 64
 	ld.w	$a4, $a2, 128
-	xvmax.w	$xr2, $xr2, $xr3
+	xvsigncov.w	$xr2, $xr2, $xr2
 	xvadd.w	$xr0, $xr0, $xr2
 	xvinsgr2vr.w	$xr1, $a3, 1
 	xvinsgr2vr.w	$xr1, $a4, 2
@@ -336,8 +322,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvld	$xr3, $a2, 12
 	ld.w	$a3, $a2, 76
 	ld.w	$a4, $a2, 140
@@ -354,8 +339,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvld	$xr3, $a2, 16
 	ld.w	$a3, $a2, 80
 	ld.w	$a4, $a2, 144
@@ -372,8 +356,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvld	$xr3, $a2, 20
 	ld.w	$a3, $a2, 84
 	ld.w	$a4, $a2, 148
@@ -390,8 +373,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvld	$xr3, $a2, 24
 	ld.w	$a3, $a2, 88
 	ld.w	$a4, $a2, 152
@@ -408,8 +390,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvld	$xr3, $a2, 28
 	ld.w	$a3, $a2, 92
 	ld.w	$a4, $a2, 156
@@ -426,8 +407,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvld	$xr3, $a2, 32
 	ld.w	$a3, $a2, 96
 	ld.w	$a4, $a2, 160
@@ -444,8 +424,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvld	$xr3, $a2, 36
 	ld.w	$a3, $a2, 100
 	ld.w	$a4, $a2, 164
@@ -462,8 +441,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvld	$xr3, $a2, 40
 	ld.w	$a3, $a2, 104
 	ld.w	$a4, $a2, 168
@@ -480,8 +458,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvld	$xr3, $a2, 44
 	ld.w	$a3, $a2, 108
 	ld.w	$a4, $a2, 172
@@ -498,8 +475,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvld	$xr3, $a2, 48
 	ld.w	$a3, $a2, 112
 	ld.w	$a4, $a2, 176
@@ -516,35 +492,33 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
-	xvld	$xr3, $a2, 52
+	xvsigncov.w	$xr2, $xr3, $xr3
+	xvadd.w	$xr1, $xr1, $xr2
+	xvld	$xr2, $a2, 52
 	ld.w	$a3, $a2, 116
 	ld.w	$a4, $a2, 180
 	ld.w	$a5, $a2, 244
-	xvadd.w	$xr1, $xr1, $xr2
-	xvinsgr2vr.w	$xr3, $a3, 1
-	xvinsgr2vr.w	$xr3, $a4, 2
-	xvinsgr2vr.w	$xr3, $a5, 3
-	ld.w	$a3, $a2, 308
-	ld.w	$a4, $a2, 372
-	ld.w	$a5, $a2, 436
-	ld.w	$a6, $a2, 500
-	xvinsgr2vr.w	$xr3, $a3, 4
-	xvinsgr2vr.w	$xr3, $a4, 5
-	xvinsgr2vr.w	$xr3, $a5, 6
-	xvinsgr2vr.w	$xr3, $a6, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	ld.w	$a6, $a2, 308
+	xvinsgr2vr.w	$xr2, $a3, 1
+	xvinsgr2vr.w	$xr2, $a4, 2
+	xvinsgr2vr.w	$xr2, $a5, 3
+	xvinsgr2vr.w	$xr2, $a6, 4
+	ld.w	$a3, $a2, 372
+	ld.w	$a4, $a2, 436
+	ld.w	$a5, $a2, 500
 	xvld	$xr3, $a2, 56
-	ld.w	$a3, $a2, 120
-	ld.w	$a4, $a2, 184
-	ld.w	$a5, $a2, 248
-	ld.w	$a6, $a2, 312
-	xvinsgr2vr.w	$xr3, $a3, 1
-	xvinsgr2vr.w	$xr3, $a4, 2
-	xvinsgr2vr.w	$xr3, $a5, 3
-	xvinsgr2vr.w	$xr3, $a6, 4
+	ld.w	$a6, $a2, 120
+	xvinsgr2vr.w	$xr2, $a3, 5
+	xvinsgr2vr.w	$xr2, $a4, 6
+	xvinsgr2vr.w	$xr2, $a5, 7
+	xvinsgr2vr.w	$xr3, $a6, 1
+	ld.w	$a3, $a2, 184
+	ld.w	$a4, $a2, 248
+	ld.w	$a5, $a2, 312
+	xvsigncov.w	$xr2, $xr2, $xr2
+	xvinsgr2vr.w	$xr3, $a3, 2
+	xvinsgr2vr.w	$xr3, $a4, 3
+	xvinsgr2vr.w	$xr3, $a5, 4
 	ld.w	$a3, $a2, 376
 	ld.w	$a4, $a2, 440
 	ld.w	$a5, $a2, 504
@@ -552,8 +526,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a3, 5
 	xvinsgr2vr.w	$xr3, $a4, 6
 	xvinsgr2vr.w	$xr3, $a5, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvld	$xr3, $a2, 60
 	ld.w	$a3, $a2, 124
 	ld.w	$a4, $a2, 188
@@ -570,8 +543,7 @@ update_rc:                              # @update_rc
 	xvinsgr2vr.w	$xr3, $a4, 5
 	xvinsgr2vr.w	$xr3, $a5, 6
 	xvinsgr2vr.w	$xr3, $a2, 7
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
+	xvsigncov.w	$xr2, $xr3, $xr3
 	xvadd.w	$xr1, $xr1, $xr2
 	pcalau12i	$a2, %pc_hi20(generic_RC)
 	ld.d	$a4, $a2, %pc_lo12(generic_RC)
@@ -655,573 +627,554 @@ update_rc:                              # @update_rc
 	.type	calc_MAD,@function
 calc_MAD:                               # @calc_MAD
 # %bb.0:                                # %vector.ph
+	addi.d	$sp, $sp, -80
+	fst.d	$fs0, $sp, 72                   # 8-byte Folded Spill
+	fst.d	$fs1, $sp, 64                   # 8-byte Folded Spill
+	fst.d	$fs2, $sp, 56                   # 8-byte Folded Spill
+	fst.d	$fs3, $sp, 48                   # 8-byte Folded Spill
+	fst.d	$fs4, $sp, 40                   # 8-byte Folded Spill
+	fst.d	$fs5, $sp, 32                   # 8-byte Folded Spill
+	fst.d	$fs6, $sp, 24                   # 8-byte Folded Spill
+	fst.d	$fs7, $sp, 16                   # 8-byte Folded Spill
 	pcalau12i	$a0, %pc_hi20(diffy)
 	addi.d	$a0, $a0, %pc_lo12(diffy)
 	xvld	$xr0, $a0, 572
-	ld.w	$a3, $a0, 636
-	ld.w	$a2, $a0, 700
-	ld.w	$a1, $a0, 764
-	xvinsgr2vr.w	$xr0, $a3, 1
-	ld.w	$a3, $a0, 828
-	xvinsgr2vr.w	$xr0, $a2, 2
-	ld.w	$a2, $a0, 892
-	xvinsgr2vr.w	$xr0, $a1, 3
-	ld.w	$a1, $a0, 956
-	xvinsgr2vr.w	$xr0, $a3, 4
-	ld.w	$a3, $a0, 1020
+	ld.w	$a4, $a0, 636
+	ld.w	$a3, $a0, 700
+	ld.w	$a2, $a0, 764
+	ld.w	$a1, $a0, 828
+	xvinsgr2vr.w	$xr0, $a4, 1
+	ld.w	$a4, $a0, 892
+	xvinsgr2vr.w	$xr0, $a3, 2
+	ld.w	$a3, $a0, 956
+	xvinsgr2vr.w	$xr0, $a2, 3
+	ld.w	$a2, $a0, 1020
 	xvld	$xr1, $a0, 568
-	xvinsgr2vr.w	$xr0, $a2, 5
-	ld.w	$a2, $a0, 632
-	xvinsgr2vr.w	$xr0, $a1, 6
-	ld.w	$a1, $a0, 696
-	xvinsgr2vr.w	$xr0, $a3, 7
+	xvinsgr2vr.w	$xr0, $a1, 4
+	ld.w	$a1, $a0, 632
+	xvinsgr2vr.w	$xr0, $a4, 5
+	ld.w	$a4, $a0, 696
+	xvinsgr2vr.w	$xr0, $a3, 6
 	ld.w	$a3, $a0, 760
-	xvinsgr2vr.w	$xr1, $a2, 1
+	xvinsgr2vr.w	$xr0, $a2, 7
 	ld.w	$a2, $a0, 824
-	xvinsgr2vr.w	$xr1, $a1, 2
+	xvinsgr2vr.w	$xr1, $a1, 1
 	ld.w	$a1, $a0, 888
+	xvinsgr2vr.w	$xr1, $a4, 2
+	ld.w	$a4, $a0, 952
 	xvinsgr2vr.w	$xr1, $a3, 3
-	ld.w	$a3, $a0, 952
+	ld.w	$a3, $a0, 1016
+	xvld	$xr2, $a0, 564
 	xvinsgr2vr.w	$xr1, $a2, 4
-	ld.w	$a2, $a0, 1016
-	xvld	$xr3, $a0, 564
+	ld.w	$a2, $a0, 628
 	xvinsgr2vr.w	$xr1, $a1, 5
-	ld.w	$a1, $a0, 628
-	xvinsgr2vr.w	$xr1, $a3, 6
-	ld.w	$a3, $a0, 692
-	xvinsgr2vr.w	$xr1, $a2, 7
-	ld.w	$a2, $a0, 756
-	xvinsgr2vr.w	$xr3, $a1, 1
-	ld.w	$a1, $a0, 820
-	xvinsgr2vr.w	$xr3, $a3, 2
-	ld.w	$a3, $a0, 884
-	xvinsgr2vr.w	$xr3, $a2, 3
-	ld.w	$a2, $a0, 948
-	xvinsgr2vr.w	$xr3, $a1, 4
-	ld.w	$a1, $a0, 1012
-	xvld	$xr4, $a0, 560
-	xvinsgr2vr.w	$xr3, $a3, 5
+	ld.w	$a1, $a0, 692
+	xvinsgr2vr.w	$xr1, $a4, 6
+	ld.w	$a4, $a0, 756
+	xvinsgr2vr.w	$xr1, $a3, 7
+	ld.w	$a3, $a0, 820
+	xvinsgr2vr.w	$xr2, $a2, 1
+	ld.w	$a2, $a0, 884
+	xvinsgr2vr.w	$xr2, $a1, 2
+	ld.w	$a1, $a0, 948
+	xvinsgr2vr.w	$xr2, $a4, 3
+	ld.w	$a4, $a0, 1012
+	xvld	$xr3, $a0, 560
+	xvinsgr2vr.w	$xr2, $a3, 4
 	ld.w	$a3, $a0, 624
-	xvinsgr2vr.w	$xr3, $a2, 6
+	xvinsgr2vr.w	$xr2, $a2, 5
 	ld.w	$a2, $a0, 688
-	xvinsgr2vr.w	$xr3, $a1, 7
+	xvinsgr2vr.w	$xr2, $a1, 6
 	ld.w	$a1, $a0, 752
-	xvinsgr2vr.w	$xr4, $a3, 1
-	ld.w	$a3, $a0, 816
-	xvinsgr2vr.w	$xr4, $a2, 2
-	ld.w	$a2, $a0, 880
-	xvinsgr2vr.w	$xr4, $a1, 3
-	ld.w	$a1, $a0, 944
-	xvinsgr2vr.w	$xr4, $a3, 4
-	ld.w	$a3, $a0, 1008
-	xvld	$xr6, $a0, 556
-	xvinsgr2vr.w	$xr4, $a2, 5
-	ld.w	$a2, $a0, 620
-	xvinsgr2vr.w	$xr4, $a1, 6
-	ld.w	$a1, $a0, 684
-	xvinsgr2vr.w	$xr4, $a3, 7
-	ld.w	$a3, $a0, 748
-	xvinsgr2vr.w	$xr6, $a2, 1
-	ld.w	$a2, $a0, 812
-	xvinsgr2vr.w	$xr6, $a1, 2
-	ld.w	$a1, $a0, 876
-	xvinsgr2vr.w	$xr6, $a3, 3
+	xvinsgr2vr.w	$xr2, $a4, 7
+	ld.w	$a4, $a0, 816
+	xvinsgr2vr.w	$xr3, $a3, 1
+	ld.w	$a3, $a0, 880
+	xvinsgr2vr.w	$xr3, $a2, 2
+	ld.w	$a2, $a0, 944
+	xvinsgr2vr.w	$xr3, $a1, 3
+	ld.w	$a1, $a0, 1008
+	xvld	$xr4, $a0, 556
+	xvinsgr2vr.w	$xr3, $a4, 4
+	ld.w	$a4, $a0, 620
+	xvinsgr2vr.w	$xr3, $a3, 5
+	ld.w	$a3, $a0, 684
+	xvinsgr2vr.w	$xr3, $a2, 6
+	ld.w	$a2, $a0, 748
+	xvinsgr2vr.w	$xr3, $a1, 7
+	ld.w	$a1, $a0, 812
+	xvinsgr2vr.w	$xr4, $a4, 1
+	ld.w	$a4, $a0, 876
+	xvinsgr2vr.w	$xr4, $a3, 2
 	ld.w	$a3, $a0, 940
-	xvinsgr2vr.w	$xr6, $a2, 4
+	xvinsgr2vr.w	$xr4, $a2, 3
 	ld.w	$a2, $a0, 1004
-	xvld	$xr7, $a0, 552
-	xvinsgr2vr.w	$xr6, $a1, 5
+	xvld	$xr5, $a0, 552
+	xvinsgr2vr.w	$xr4, $a1, 4
 	ld.w	$a1, $a0, 616
-	xvinsgr2vr.w	$xr6, $a3, 6
-	ld.w	$a3, $a0, 680
-	xvinsgr2vr.w	$xr6, $a2, 7
-	ld.w	$a2, $a0, 744
-	xvinsgr2vr.w	$xr7, $a1, 1
-	ld.w	$a1, $a0, 808
-	xvinsgr2vr.w	$xr7, $a3, 2
-	ld.w	$a3, $a0, 872
-	xvinsgr2vr.w	$xr7, $a2, 3
-	ld.w	$a2, $a0, 936
-	xvinsgr2vr.w	$xr7, $a1, 4
-	ld.w	$a1, $a0, 1000
-	xvld	$xr9, $a0, 548
-	xvinsgr2vr.w	$xr7, $a3, 5
-	ld.w	$a3, $a0, 612
-	xvinsgr2vr.w	$xr7, $a2, 6
-	ld.w	$a2, $a0, 676
-	xvinsgr2vr.w	$xr7, $a1, 7
-	ld.w	$a1, $a0, 740
-	xvinsgr2vr.w	$xr9, $a3, 1
+	xvinsgr2vr.w	$xr4, $a4, 5
+	ld.w	$a4, $a0, 680
+	xvinsgr2vr.w	$xr4, $a3, 6
+	ld.w	$a3, $a0, 744
+	xvinsgr2vr.w	$xr4, $a2, 7
+	ld.w	$a2, $a0, 808
+	xvinsgr2vr.w	$xr5, $a1, 1
+	ld.w	$a1, $a0, 872
+	xvinsgr2vr.w	$xr5, $a4, 2
+	ld.w	$a4, $a0, 936
+	xvinsgr2vr.w	$xr5, $a3, 3
+	ld.w	$a3, $a0, 1000
+	xvld	$xr6, $a0, 548
+	xvinsgr2vr.w	$xr5, $a2, 4
+	ld.w	$a2, $a0, 612
+	xvinsgr2vr.w	$xr5, $a1, 5
+	ld.w	$a1, $a0, 676
+	xvinsgr2vr.w	$xr5, $a4, 6
+	ld.w	$a4, $a0, 740
+	xvinsgr2vr.w	$xr5, $a3, 7
 	ld.w	$a3, $a0, 804
-	xvinsgr2vr.w	$xr9, $a2, 2
+	xvinsgr2vr.w	$xr6, $a2, 1
 	ld.w	$a2, $a0, 868
-	xvinsgr2vr.w	$xr9, $a1, 3
+	xvinsgr2vr.w	$xr6, $a1, 2
 	ld.w	$a1, $a0, 932
-	xvinsgr2vr.w	$xr9, $a3, 4
-	ld.w	$a3, $a0, 996
-	xvld	$xr11, $a0, 544
-	xvinsgr2vr.w	$xr9, $a2, 5
-	ld.w	$a2, $a0, 608
-	xvinsgr2vr.w	$xr9, $a1, 6
-	ld.w	$a1, $a0, 672
-	xvinsgr2vr.w	$xr9, $a3, 7
-	ld.w	$a3, $a0, 736
-	xvinsgr2vr.w	$xr11, $a2, 1
-	ld.w	$a2, $a0, 800
-	xvinsgr2vr.w	$xr11, $a1, 2
-	ld.w	$a1, $a0, 864
-	xvinsgr2vr.w	$xr11, $a3, 3
-	ld.w	$a3, $a0, 928
-	xvinsgr2vr.w	$xr11, $a2, 4
-	ld.w	$a2, $a0, 992
-	xvld	$xr13, $a0, 540
-	xvinsgr2vr.w	$xr11, $a1, 5
-	ld.w	$a1, $a0, 604
-	xvinsgr2vr.w	$xr11, $a3, 6
+	xvinsgr2vr.w	$xr6, $a4, 3
+	ld.w	$a4, $a0, 996
+	xvld	$xr7, $a0, 544
+	xvinsgr2vr.w	$xr6, $a3, 4
+	ld.w	$a3, $a0, 608
+	xvinsgr2vr.w	$xr6, $a2, 5
+	ld.w	$a2, $a0, 672
+	xvinsgr2vr.w	$xr6, $a1, 6
+	ld.w	$a1, $a0, 736
+	xvinsgr2vr.w	$xr6, $a4, 7
+	ld.w	$a4, $a0, 800
+	xvinsgr2vr.w	$xr7, $a3, 1
+	ld.w	$a3, $a0, 864
+	xvinsgr2vr.w	$xr7, $a2, 2
+	ld.w	$a2, $a0, 928
+	xvinsgr2vr.w	$xr7, $a1, 3
+	ld.w	$a1, $a0, 992
+	xvld	$xr8, $a0, 540
+	xvinsgr2vr.w	$xr7, $a4, 4
+	ld.w	$a4, $a0, 604
+	xvinsgr2vr.w	$xr7, $a3, 5
 	ld.w	$a3, $a0, 668
-	xvinsgr2vr.w	$xr11, $a2, 7
+	xvinsgr2vr.w	$xr7, $a2, 6
 	ld.w	$a2, $a0, 732
-	xvinsgr2vr.w	$xr13, $a1, 1
+	xvinsgr2vr.w	$xr7, $a1, 7
 	ld.w	$a1, $a0, 796
-	xvinsgr2vr.w	$xr13, $a3, 2
-	ld.w	$a3, $a0, 860
-	xvinsgr2vr.w	$xr13, $a2, 3
-	ld.w	$a2, $a0, 924
-	xvinsgr2vr.w	$xr13, $a1, 4
-	ld.w	$a1, $a0, 988
-	xvld	$xr16, $a0, 536
-	xvinsgr2vr.w	$xr13, $a3, 5
-	ld.w	$a3, $a0, 600
-	xvinsgr2vr.w	$xr13, $a2, 6
-	ld.w	$a2, $a0, 664
-	xvinsgr2vr.w	$xr13, $a1, 7
-	ld.w	$a1, $a0, 728
-	xvinsgr2vr.w	$xr16, $a3, 1
-	ld.w	$a3, $a0, 792
-	xvinsgr2vr.w	$xr16, $a2, 2
-	ld.w	$a2, $a0, 856
-	xvinsgr2vr.w	$xr16, $a1, 3
-	ld.w	$a1, $a0, 920
-	xvinsgr2vr.w	$xr16, $a3, 4
+	xvinsgr2vr.w	$xr8, $a4, 1
+	ld.w	$a4, $a0, 860
+	xvinsgr2vr.w	$xr8, $a3, 2
+	ld.w	$a3, $a0, 924
+	xvinsgr2vr.w	$xr8, $a2, 3
+	ld.w	$a2, $a0, 988
+	xvld	$xr9, $a0, 536
+	xvinsgr2vr.w	$xr8, $a1, 4
+	ld.w	$a1, $a0, 600
+	xvinsgr2vr.w	$xr8, $a4, 5
+	ld.w	$a4, $a0, 664
+	xvinsgr2vr.w	$xr8, $a3, 6
+	ld.w	$a3, $a0, 728
+	xvinsgr2vr.w	$xr8, $a2, 7
+	ld.w	$a2, $a0, 792
+	xvinsgr2vr.w	$xr9, $a1, 1
+	ld.w	$a1, $a0, 856
+	xvinsgr2vr.w	$xr9, $a4, 2
+	ld.w	$a4, $a0, 920
+	xvinsgr2vr.w	$xr9, $a3, 3
 	ld.w	$a3, $a0, 984
-	xvld	$xr17, $a0, 532
-	xvinsgr2vr.w	$xr16, $a2, 5
+	xvld	$xr10, $a0, 532
+	xvinsgr2vr.w	$xr9, $a2, 4
 	ld.w	$a2, $a0, 596
-	xvinsgr2vr.w	$xr16, $a1, 6
+	xvinsgr2vr.w	$xr9, $a1, 5
 	ld.w	$a1, $a0, 660
-	xvinsgr2vr.w	$xr16, $a3, 7
-	ld.w	$a3, $a0, 724
-	xvinsgr2vr.w	$xr17, $a2, 1
-	ld.w	$a2, $a0, 788
-	xvinsgr2vr.w	$xr17, $a1, 2
-	ld.w	$a1, $a0, 852
-	xvinsgr2vr.w	$xr17, $a3, 3
-	ld.w	$a3, $a0, 916
-	xvinsgr2vr.w	$xr17, $a2, 4
-	ld.w	$a2, $a0, 980
-	xvld	$xr18, $a0, 528
-	xvinsgr2vr.w	$xr17, $a1, 5
-	ld.w	$a1, $a0, 592
-	xvinsgr2vr.w	$xr17, $a3, 6
-	ld.w	$a3, $a0, 656
-	xvinsgr2vr.w	$xr17, $a2, 7
-	ld.w	$a2, $a0, 720
-	xvinsgr2vr.w	$xr18, $a1, 1
-	ld.w	$a1, $a0, 784
-	xvinsgr2vr.w	$xr18, $a3, 2
+	xvinsgr2vr.w	$xr9, $a4, 6
+	ld.w	$a4, $a0, 724
+	xvinsgr2vr.w	$xr9, $a3, 7
+	ld.w	$a3, $a0, 788
+	xvinsgr2vr.w	$xr10, $a2, 1
+	ld.w	$a2, $a0, 852
+	xvinsgr2vr.w	$xr10, $a1, 2
+	ld.w	$a1, $a0, 916
+	xvinsgr2vr.w	$xr10, $a4, 3
+	ld.w	$a4, $a0, 980
+	xvld	$xr11, $a0, 528
+	xvinsgr2vr.w	$xr10, $a3, 4
+	ld.w	$a3, $a0, 592
+	xvinsgr2vr.w	$xr10, $a2, 5
+	ld.w	$a2, $a0, 656
+	xvinsgr2vr.w	$xr10, $a1, 6
+	ld.w	$a1, $a0, 720
+	xvinsgr2vr.w	$xr10, $a4, 7
+	ld.w	$a4, $a0, 784
+	xvinsgr2vr.w	$xr11, $a3, 1
 	ld.w	$a3, $a0, 848
-	xvinsgr2vr.w	$xr18, $a2, 3
+	xvinsgr2vr.w	$xr11, $a2, 2
 	ld.w	$a2, $a0, 912
-	xvinsgr2vr.w	$xr18, $a1, 4
+	xvinsgr2vr.w	$xr11, $a1, 3
 	ld.w	$a1, $a0, 976
-	xvld	$xr19, $a0, 524
-	xvinsgr2vr.w	$xr18, $a3, 5
-	ld.w	$a3, $a0, 588
-	xvinsgr2vr.w	$xr18, $a2, 6
-	ld.w	$a2, $a0, 652
-	xvinsgr2vr.w	$xr18, $a1, 7
-	ld.w	$a1, $a0, 716
-	xvinsgr2vr.w	$xr19, $a3, 1
-	ld.w	$a3, $a0, 780
-	xvinsgr2vr.w	$xr19, $a2, 2
-	ld.w	$a2, $a0, 844
-	xvinsgr2vr.w	$xr19, $a1, 3
-	ld.w	$a1, $a0, 908
-	xvinsgr2vr.w	$xr19, $a3, 4
-	ld.w	$a3, $a0, 972
-	xvld	$xr21, $a0, 520
-	xvinsgr2vr.w	$xr19, $a2, 5
-	ld.w	$a2, $a0, 584
-	xvinsgr2vr.w	$xr19, $a1, 6
-	ld.w	$a1, $a0, 648
-	xvinsgr2vr.w	$xr19, $a3, 7
+	xvld	$xr12, $a0, 524
+	xvinsgr2vr.w	$xr11, $a4, 4
+	ld.w	$a4, $a0, 588
+	xvinsgr2vr.w	$xr11, $a3, 5
+	ld.w	$a3, $a0, 652
+	xvinsgr2vr.w	$xr11, $a2, 6
+	ld.w	$a2, $a0, 716
+	xvinsgr2vr.w	$xr11, $a1, 7
+	ld.w	$a1, $a0, 780
+	xvinsgr2vr.w	$xr12, $a4, 1
+	ld.w	$a4, $a0, 844
+	xvinsgr2vr.w	$xr12, $a3, 2
+	ld.w	$a3, $a0, 908
+	xvinsgr2vr.w	$xr12, $a2, 3
+	ld.w	$a2, $a0, 972
+	xvld	$xr13, $a0, 520
+	xvinsgr2vr.w	$xr12, $a1, 4
+	ld.w	$a1, $a0, 584
+	xvinsgr2vr.w	$xr12, $a4, 5
+	ld.w	$a4, $a0, 648
+	xvinsgr2vr.w	$xr12, $a3, 6
 	ld.w	$a3, $a0, 712
-	xvinsgr2vr.w	$xr21, $a2, 1
+	xvinsgr2vr.w	$xr12, $a2, 7
 	ld.w	$a2, $a0, 776
-	xvinsgr2vr.w	$xr21, $a1, 2
+	xvinsgr2vr.w	$xr13, $a1, 1
 	ld.w	$a1, $a0, 840
-	xvinsgr2vr.w	$xr21, $a3, 3
-	ld.w	$a3, $a0, 904
-	xvinsgr2vr.w	$xr21, $a2, 4
-	ld.w	$a2, $a0, 968
-	xvld	$xr20, $a0, 516
-	xvinsgr2vr.w	$xr21, $a1, 5
-	ld.w	$a1, $a0, 580
-	xvinsgr2vr.w	$xr21, $a3, 6
-	ld.w	$a3, $a0, 644
-	xvinsgr2vr.w	$xr21, $a2, 7
-	ld.w	$a2, $a0, 708
-	xvinsgr2vr.w	$xr20, $a1, 1
-	ld.w	$a1, $a0, 772
-	xvinsgr2vr.w	$xr20, $a3, 2
-	ld.w	$a3, $a0, 836
-	xvinsgr2vr.w	$xr20, $a2, 3
-	ld.w	$a2, $a0, 900
-	xvinsgr2vr.w	$xr20, $a1, 4
-	ld.w	$a1, $a0, 964
-	xvld	$xr22, $a0, 512
-	xvinsgr2vr.w	$xr20, $a3, 5
+	xvinsgr2vr.w	$xr13, $a4, 2
+	ld.w	$a4, $a0, 904
+	xvinsgr2vr.w	$xr13, $a3, 3
+	ld.w	$a3, $a0, 968
+	xvld	$xr14, $a0, 516
+	xvinsgr2vr.w	$xr13, $a2, 4
+	ld.w	$a2, $a0, 580
+	xvinsgr2vr.w	$xr13, $a1, 5
+	ld.w	$a1, $a0, 644
+	xvinsgr2vr.w	$xr13, $a4, 6
+	ld.w	$a4, $a0, 708
+	xvinsgr2vr.w	$xr13, $a3, 7
+	ld.w	$a3, $a0, 772
+	xvinsgr2vr.w	$xr14, $a2, 1
+	ld.w	$a2, $a0, 836
+	xvinsgr2vr.w	$xr14, $a1, 2
+	ld.w	$a1, $a0, 900
+	xvinsgr2vr.w	$xr14, $a4, 3
+	ld.w	$a4, $a0, 964
+	xvld	$xr15, $a0, 512
+	xvinsgr2vr.w	$xr14, $a3, 4
 	ld.w	$a3, $a0, 576
-	xvinsgr2vr.w	$xr20, $a2, 6
+	xvinsgr2vr.w	$xr14, $a2, 5
 	ld.w	$a2, $a0, 640
-	xvinsgr2vr.w	$xr20, $a1, 7
+	xvinsgr2vr.w	$xr14, $a1, 6
 	ld.w	$a1, $a0, 704
-	xvinsgr2vr.w	$xr22, $a3, 1
-	ld.w	$a3, $a0, 768
-	xvinsgr2vr.w	$xr22, $a2, 2
-	ld.w	$a2, $a0, 832
-	xvinsgr2vr.w	$xr22, $a1, 3
-	ld.w	$a1, $a0, 896
-	xvinsgr2vr.w	$xr22, $a3, 4
-	ld.w	$a3, $a0, 960
-	xvneg.w	$xr2, $xr0
-	xvmax.w	$xr0, $xr0, $xr2
-	xvneg.w	$xr2, $xr1
-	xvmax.w	$xr2, $xr1, $xr2
-	xvneg.w	$xr1, $xr3
-	xvmax.w	$xr3, $xr3, $xr1
-	xvneg.w	$xr1, $xr4
-	xvmax.w	$xr5, $xr4, $xr1
-	xvneg.w	$xr1, $xr6
-	xvmax.w	$xr8, $xr6, $xr1
-	xvneg.w	$xr1, $xr7
-	xvmax.w	$xr10, $xr7, $xr1
-	xvneg.w	$xr1, $xr9
-	xvmax.w	$xr12, $xr9, $xr1
-	xvneg.w	$xr1, $xr11
-	xvmax.w	$xr14, $xr11, $xr1
-	xvneg.w	$xr1, $xr13
-	xvmax.w	$xr15, $xr13, $xr1
-	xvneg.w	$xr1, $xr16
-	xvmax.w	$xr13, $xr16, $xr1
-	xvneg.w	$xr1, $xr17
-	xvmax.w	$xr11, $xr17, $xr1
-	xvneg.w	$xr1, $xr18
-	xvmax.w	$xr9, $xr18, $xr1
-	xvneg.w	$xr1, $xr19
-	xvmax.w	$xr7, $xr19, $xr1
-	xvneg.w	$xr1, $xr21
-	xvmax.w	$xr6, $xr21, $xr1
-	xvld	$xr1, $a0, 60
-	xvinsgr2vr.w	$xr22, $a2, 5
-	ld.w	$a2, $a0, 124
-	xvinsgr2vr.w	$xr22, $a1, 6
-	ld.w	$a1, $a0, 188
-	xvinsgr2vr.w	$xr22, $a3, 7
-	ld.w	$a3, $a0, 252
-	xvinsgr2vr.w	$xr1, $a2, 1
-	ld.w	$a2, $a0, 316
-	xvinsgr2vr.w	$xr1, $a1, 2
-	ld.w	$a1, $a0, 380
-	xvinsgr2vr.w	$xr1, $a3, 3
+	xvinsgr2vr.w	$xr14, $a4, 7
+	ld.w	$a4, $a0, 768
+	xvinsgr2vr.w	$xr15, $a3, 1
+	ld.w	$a3, $a0, 832
+	xvinsgr2vr.w	$xr15, $a2, 2
+	ld.w	$a2, $a0, 896
+	xvinsgr2vr.w	$xr15, $a1, 3
+	ld.w	$a1, $a0, 960
+	xvld	$xr16, $a0, 60
+	xvinsgr2vr.w	$xr15, $a4, 4
+	ld.w	$a4, $a0, 124
+	xvinsgr2vr.w	$xr15, $a3, 5
+	ld.w	$a3, $a0, 188
+	xvinsgr2vr.w	$xr15, $a2, 6
+	ld.w	$a2, $a0, 252
+	xvinsgr2vr.w	$xr15, $a1, 7
+	ld.w	$a1, $a0, 316
+	xvinsgr2vr.w	$xr16, $a4, 1
+	ld.w	$a4, $a0, 380
+	xvinsgr2vr.w	$xr16, $a3, 2
 	ld.w	$a3, $a0, 444
-	xvinsgr2vr.w	$xr1, $a2, 4
+	xvinsgr2vr.w	$xr16, $a2, 3
 	ld.w	$a2, $a0, 508
-	xvadda.w	$xr16, $xr20, $xr22
-	xvld	$xr4, $a0, 56
-	xvinsgr2vr.w	$xr1, $a1, 5
+	xvld	$xr17, $a0, 56
+	xvinsgr2vr.w	$xr16, $a1, 4
 	ld.w	$a1, $a0, 120
-	xvinsgr2vr.w	$xr1, $a3, 6
-	ld.w	$a3, $a0, 184
-	xvinsgr2vr.w	$xr1, $a2, 7
-	ld.w	$a2, $a0, 248
-	xvinsgr2vr.w	$xr4, $a1, 1
-	ld.w	$a1, $a0, 312
-	xvinsgr2vr.w	$xr4, $a3, 2
-	ld.w	$a3, $a0, 376
-	xvinsgr2vr.w	$xr4, $a2, 3
-	ld.w	$a2, $a0, 440
-	xvinsgr2vr.w	$xr4, $a1, 4
-	ld.w	$a1, $a0, 504
-	xvadd.w	$xr16, $xr6, $xr16
-	xvld	$xr6, $a0, 52
-	xvinsgr2vr.w	$xr4, $a3, 5
-	ld.w	$a3, $a0, 116
-	xvinsgr2vr.w	$xr4, $a2, 6
-	ld.w	$a2, $a0, 180
-	xvinsgr2vr.w	$xr4, $a1, 7
-	ld.w	$a1, $a0, 244
-	xvinsgr2vr.w	$xr6, $a3, 1
+	xvinsgr2vr.w	$xr16, $a4, 5
+	ld.w	$a4, $a0, 184
+	xvinsgr2vr.w	$xr16, $a3, 6
+	ld.w	$a3, $a0, 248
+	xvinsgr2vr.w	$xr16, $a2, 7
+	ld.w	$a2, $a0, 312
+	xvinsgr2vr.w	$xr17, $a1, 1
+	ld.w	$a1, $a0, 376
+	xvinsgr2vr.w	$xr17, $a4, 2
+	ld.w	$a4, $a0, 440
+	xvinsgr2vr.w	$xr17, $a3, 3
+	ld.w	$a3, $a0, 504
+	xvld	$xr18, $a0, 52
+	xvinsgr2vr.w	$xr17, $a2, 4
+	ld.w	$a2, $a0, 116
+	xvinsgr2vr.w	$xr17, $a1, 5
+	ld.w	$a1, $a0, 180
+	xvinsgr2vr.w	$xr17, $a4, 6
+	ld.w	$a4, $a0, 244
+	xvinsgr2vr.w	$xr17, $a3, 7
 	ld.w	$a3, $a0, 308
-	xvinsgr2vr.w	$xr6, $a2, 2
+	xvinsgr2vr.w	$xr18, $a2, 1
 	ld.w	$a2, $a0, 372
-	xvinsgr2vr.w	$xr6, $a1, 3
+	xvinsgr2vr.w	$xr18, $a1, 2
 	ld.w	$a1, $a0, 436
-	xvinsgr2vr.w	$xr6, $a3, 4
-	ld.w	$a3, $a0, 500
-	xvadd.w	$xr16, $xr7, $xr16
-	xvld	$xr7, $a0, 48
-	xvinsgr2vr.w	$xr6, $a2, 5
-	ld.w	$a2, $a0, 112
-	xvinsgr2vr.w	$xr6, $a1, 6
-	ld.w	$a1, $a0, 176
-	xvinsgr2vr.w	$xr6, $a3, 7
-	ld.w	$a3, $a0, 240
-	xvinsgr2vr.w	$xr7, $a2, 1
-	ld.w	$a2, $a0, 304
-	xvinsgr2vr.w	$xr7, $a1, 2
-	ld.w	$a1, $a0, 368
-	xvinsgr2vr.w	$xr7, $a3, 3
-	ld.w	$a3, $a0, 432
-	xvinsgr2vr.w	$xr7, $a2, 4
-	ld.w	$a2, $a0, 496
-	xvadd.w	$xr16, $xr9, $xr16
-	xvld	$xr9, $a0, 44
-	xvinsgr2vr.w	$xr7, $a1, 5
-	ld.w	$a1, $a0, 108
-	xvinsgr2vr.w	$xr7, $a3, 6
+	xvinsgr2vr.w	$xr18, $a4, 3
+	ld.w	$a4, $a0, 500
+	xvld	$xr19, $a0, 48
+	xvinsgr2vr.w	$xr18, $a3, 4
+	ld.w	$a3, $a0, 112
+	xvinsgr2vr.w	$xr18, $a2, 5
+	ld.w	$a2, $a0, 176
+	xvinsgr2vr.w	$xr18, $a1, 6
+	ld.w	$a1, $a0, 240
+	xvinsgr2vr.w	$xr18, $a4, 7
+	ld.w	$a4, $a0, 304
+	xvinsgr2vr.w	$xr19, $a3, 1
+	ld.w	$a3, $a0, 368
+	xvinsgr2vr.w	$xr19, $a2, 2
+	ld.w	$a2, $a0, 432
+	xvinsgr2vr.w	$xr19, $a1, 3
+	ld.w	$a1, $a0, 496
+	xvld	$xr20, $a0, 44
+	xvinsgr2vr.w	$xr19, $a4, 4
+	ld.w	$a4, $a0, 108
+	xvinsgr2vr.w	$xr19, $a3, 5
 	ld.w	$a3, $a0, 172
-	xvinsgr2vr.w	$xr7, $a2, 7
+	xvinsgr2vr.w	$xr19, $a2, 6
 	ld.w	$a2, $a0, 236
-	xvinsgr2vr.w	$xr9, $a1, 1
+	xvinsgr2vr.w	$xr19, $a1, 7
 	ld.w	$a1, $a0, 300
-	xvinsgr2vr.w	$xr9, $a3, 2
-	ld.w	$a3, $a0, 364
-	xvinsgr2vr.w	$xr9, $a2, 3
-	ld.w	$a2, $a0, 428
-	xvinsgr2vr.w	$xr9, $a1, 4
-	ld.w	$a1, $a0, 492
-	xvadd.w	$xr16, $xr11, $xr16
-	xvld	$xr11, $a0, 40
-	xvinsgr2vr.w	$xr9, $a3, 5
-	ld.w	$a3, $a0, 104
-	xvinsgr2vr.w	$xr9, $a2, 6
-	ld.w	$a2, $a0, 168
-	xvinsgr2vr.w	$xr9, $a1, 7
-	ld.w	$a1, $a0, 232
-	xvinsgr2vr.w	$xr11, $a3, 1
-	ld.w	$a3, $a0, 296
-	xvinsgr2vr.w	$xr11, $a2, 2
-	ld.w	$a2, $a0, 360
-	xvinsgr2vr.w	$xr11, $a1, 3
-	ld.w	$a1, $a0, 424
-	xvinsgr2vr.w	$xr11, $a3, 4
+	xvinsgr2vr.w	$xr20, $a4, 1
+	ld.w	$a4, $a0, 364
+	xvinsgr2vr.w	$xr20, $a3, 2
+	ld.w	$a3, $a0, 428
+	xvinsgr2vr.w	$xr20, $a2, 3
+	ld.w	$a2, $a0, 492
+	xvld	$xr21, $a0, 40
+	xvinsgr2vr.w	$xr20, $a1, 4
+	ld.w	$a1, $a0, 104
+	xvinsgr2vr.w	$xr20, $a4, 5
+	ld.w	$a4, $a0, 168
+	xvinsgr2vr.w	$xr20, $a3, 6
+	ld.w	$a3, $a0, 232
+	xvinsgr2vr.w	$xr20, $a2, 7
+	ld.w	$a2, $a0, 296
+	xvinsgr2vr.w	$xr21, $a1, 1
+	ld.w	$a1, $a0, 360
+	xvinsgr2vr.w	$xr21, $a4, 2
+	ld.w	$a4, $a0, 424
+	xvinsgr2vr.w	$xr21, $a3, 3
 	ld.w	$a3, $a0, 488
-	xvadd.w	$xr16, $xr13, $xr16
-	xvld	$xr13, $a0, 36
-	xvinsgr2vr.w	$xr11, $a2, 5
+	xvld	$xr22, $a0, 36
+	xvinsgr2vr.w	$xr21, $a2, 4
 	ld.w	$a2, $a0, 100
-	xvinsgr2vr.w	$xr11, $a1, 6
+	xvinsgr2vr.w	$xr21, $a1, 5
 	ld.w	$a1, $a0, 164
-	xvinsgr2vr.w	$xr11, $a3, 7
-	ld.w	$a3, $a0, 228
-	xvinsgr2vr.w	$xr13, $a2, 1
-	ld.w	$a2, $a0, 292
-	xvinsgr2vr.w	$xr13, $a1, 2
-	ld.w	$a1, $a0, 356
-	xvinsgr2vr.w	$xr13, $a3, 3
-	ld.w	$a3, $a0, 420
-	xvinsgr2vr.w	$xr13, $a2, 4
-	ld.w	$a2, $a0, 484
-	xvadd.w	$xr16, $xr15, $xr16
-	xvld	$xr15, $a0, 32
-	xvinsgr2vr.w	$xr13, $a1, 5
-	ld.w	$a1, $a0, 96
-	xvinsgr2vr.w	$xr13, $a3, 6
-	ld.w	$a3, $a0, 160
-	xvinsgr2vr.w	$xr13, $a2, 7
-	ld.w	$a2, $a0, 224
-	xvinsgr2vr.w	$xr15, $a1, 1
-	ld.w	$a1, $a0, 288
-	xvinsgr2vr.w	$xr15, $a3, 2
+	xvinsgr2vr.w	$xr21, $a4, 6
+	ld.w	$a4, $a0, 228
+	xvinsgr2vr.w	$xr21, $a3, 7
+	ld.w	$a3, $a0, 292
+	xvinsgr2vr.w	$xr22, $a2, 1
+	ld.w	$a2, $a0, 356
+	xvinsgr2vr.w	$xr22, $a1, 2
+	ld.w	$a1, $a0, 420
+	xvinsgr2vr.w	$xr22, $a4, 3
+	ld.w	$a4, $a0, 484
+	xvld	$xr23, $a0, 32
+	xvinsgr2vr.w	$xr22, $a3, 4
+	ld.w	$a3, $a0, 96
+	xvinsgr2vr.w	$xr22, $a2, 5
+	ld.w	$a2, $a0, 160
+	xvinsgr2vr.w	$xr22, $a1, 6
+	ld.w	$a1, $a0, 224
+	xvinsgr2vr.w	$xr22, $a4, 7
+	ld.w	$a4, $a0, 288
+	xvinsgr2vr.w	$xr23, $a3, 1
 	ld.w	$a3, $a0, 352
-	xvinsgr2vr.w	$xr15, $a2, 3
+	xvinsgr2vr.w	$xr23, $a2, 2
 	ld.w	$a2, $a0, 416
-	xvinsgr2vr.w	$xr15, $a1, 4
+	xvinsgr2vr.w	$xr23, $a1, 3
 	ld.w	$a1, $a0, 480
-	xvadd.w	$xr16, $xr14, $xr16
-	xvld	$xr14, $a0, 28
-	xvinsgr2vr.w	$xr15, $a3, 5
-	ld.w	$a3, $a0, 92
-	xvinsgr2vr.w	$xr15, $a2, 6
-	ld.w	$a2, $a0, 156
-	xvinsgr2vr.w	$xr15, $a1, 7
-	ld.w	$a1, $a0, 220
-	xvinsgr2vr.w	$xr14, $a3, 1
-	ld.w	$a3, $a0, 284
-	xvinsgr2vr.w	$xr14, $a2, 2
-	ld.w	$a2, $a0, 348
-	xvinsgr2vr.w	$xr14, $a1, 3
-	ld.w	$a1, $a0, 412
-	xvinsgr2vr.w	$xr14, $a3, 4
-	ld.w	$a3, $a0, 476
-	xvadd.w	$xr16, $xr12, $xr16
-	xvld	$xr12, $a0, 24
-	xvinsgr2vr.w	$xr14, $a2, 5
-	ld.w	$a2, $a0, 88
-	xvinsgr2vr.w	$xr14, $a1, 6
-	ld.w	$a1, $a0, 152
-	xvinsgr2vr.w	$xr14, $a3, 7
+	xvld	$xr24, $a0, 28
+	xvinsgr2vr.w	$xr23, $a4, 4
+	ld.w	$a4, $a0, 92
+	xvinsgr2vr.w	$xr23, $a3, 5
+	ld.w	$a3, $a0, 156
+	xvinsgr2vr.w	$xr23, $a2, 6
+	ld.w	$a2, $a0, 220
+	xvinsgr2vr.w	$xr23, $a1, 7
+	ld.w	$a1, $a0, 284
+	xvinsgr2vr.w	$xr24, $a4, 1
+	ld.w	$a4, $a0, 348
+	xvinsgr2vr.w	$xr24, $a3, 2
+	ld.w	$a3, $a0, 412
+	xvinsgr2vr.w	$xr24, $a2, 3
+	ld.w	$a2, $a0, 476
+	xvld	$xr25, $a0, 24
+	xvinsgr2vr.w	$xr24, $a1, 4
+	ld.w	$a1, $a0, 88
+	xvinsgr2vr.w	$xr24, $a4, 5
+	ld.w	$a4, $a0, 152
+	xvinsgr2vr.w	$xr24, $a3, 6
 	ld.w	$a3, $a0, 216
-	xvinsgr2vr.w	$xr12, $a2, 1
+	xvinsgr2vr.w	$xr24, $a2, 7
 	ld.w	$a2, $a0, 280
-	xvinsgr2vr.w	$xr12, $a1, 2
+	xvinsgr2vr.w	$xr25, $a1, 1
 	ld.w	$a1, $a0, 344
-	xvinsgr2vr.w	$xr12, $a3, 3
-	ld.w	$a3, $a0, 408
-	xvinsgr2vr.w	$xr12, $a2, 4
-	ld.w	$a2, $a0, 472
-	xvadd.w	$xr16, $xr10, $xr16
-	xvld	$xr10, $a0, 20
-	xvinsgr2vr.w	$xr12, $a1, 5
-	ld.w	$a1, $a0, 84
-	xvinsgr2vr.w	$xr12, $a3, 6
-	ld.w	$a3, $a0, 148
-	xvinsgr2vr.w	$xr12, $a2, 7
-	ld.w	$a2, $a0, 212
-	xvinsgr2vr.w	$xr10, $a1, 1
-	ld.w	$a1, $a0, 276
-	xvinsgr2vr.w	$xr10, $a3, 2
-	ld.w	$a3, $a0, 340
-	xvinsgr2vr.w	$xr10, $a2, 3
-	ld.w	$a2, $a0, 404
-	xvinsgr2vr.w	$xr10, $a1, 4
-	ld.w	$a1, $a0, 468
-	xvadd.w	$xr16, $xr8, $xr16
-	xvld	$xr8, $a0, 16
-	xvinsgr2vr.w	$xr10, $a3, 5
+	xvinsgr2vr.w	$xr25, $a4, 2
+	ld.w	$a4, $a0, 408
+	xvinsgr2vr.w	$xr25, $a3, 3
+	ld.w	$a3, $a0, 472
+	xvld	$xr26, $a0, 20
+	xvinsgr2vr.w	$xr25, $a2, 4
+	ld.w	$a2, $a0, 84
+	xvinsgr2vr.w	$xr25, $a1, 5
+	ld.w	$a1, $a0, 148
+	xvinsgr2vr.w	$xr25, $a4, 6
+	ld.w	$a4, $a0, 212
+	xvinsgr2vr.w	$xr25, $a3, 7
+	ld.w	$a3, $a0, 276
+	xvinsgr2vr.w	$xr26, $a2, 1
+	ld.w	$a2, $a0, 340
+	xvinsgr2vr.w	$xr26, $a1, 2
+	ld.w	$a1, $a0, 404
+	xvinsgr2vr.w	$xr26, $a4, 3
+	ld.w	$a4, $a0, 468
+	xvld	$xr27, $a0, 16
+	xvinsgr2vr.w	$xr26, $a3, 4
 	ld.w	$a3, $a0, 80
-	xvinsgr2vr.w	$xr10, $a2, 6
+	xvinsgr2vr.w	$xr26, $a2, 5
 	ld.w	$a2, $a0, 144
-	xvinsgr2vr.w	$xr10, $a1, 7
+	xvinsgr2vr.w	$xr26, $a1, 6
 	ld.w	$a1, $a0, 208
-	xvinsgr2vr.w	$xr8, $a3, 1
-	ld.w	$a3, $a0, 272
-	xvinsgr2vr.w	$xr8, $a2, 2
-	ld.w	$a2, $a0, 336
-	xvinsgr2vr.w	$xr8, $a1, 3
-	ld.w	$a1, $a0, 400
-	xvinsgr2vr.w	$xr8, $a3, 4
-	ld.w	$a3, $a0, 464
-	xvadd.w	$xr16, $xr5, $xr16
-	xvld	$xr5, $a0, 12
-	xvinsgr2vr.w	$xr8, $a2, 5
-	ld.w	$a2, $a0, 76
-	xvinsgr2vr.w	$xr8, $a1, 6
-	ld.w	$a1, $a0, 140
-	xvinsgr2vr.w	$xr8, $a3, 7
-	ld.w	$a3, $a0, 204
-	xvinsgr2vr.w	$xr5, $a2, 1
-	ld.w	$a2, $a0, 268
-	xvinsgr2vr.w	$xr5, $a1, 2
-	ld.w	$a1, $a0, 332
-	xvinsgr2vr.w	$xr5, $a3, 3
+	xvinsgr2vr.w	$xr26, $a4, 7
+	ld.w	$a4, $a0, 272
+	xvinsgr2vr.w	$xr27, $a3, 1
+	ld.w	$a3, $a0, 336
+	xvinsgr2vr.w	$xr27, $a2, 2
+	ld.w	$a2, $a0, 400
+	xvinsgr2vr.w	$xr27, $a1, 3
+	ld.w	$a1, $a0, 464
+	xvld	$xr28, $a0, 12
+	xvinsgr2vr.w	$xr27, $a4, 4
+	ld.w	$a4, $a0, 76
+	xvinsgr2vr.w	$xr27, $a3, 5
+	ld.w	$a3, $a0, 140
+	xvinsgr2vr.w	$xr27, $a2, 6
+	ld.w	$a2, $a0, 204
+	xvinsgr2vr.w	$xr27, $a1, 7
+	ld.w	$a1, $a0, 268
+	xvinsgr2vr.w	$xr28, $a4, 1
+	ld.w	$a4, $a0, 332
+	xvinsgr2vr.w	$xr28, $a3, 2
 	ld.w	$a3, $a0, 396
-	xvinsgr2vr.w	$xr5, $a2, 4
+	xvinsgr2vr.w	$xr28, $a2, 3
 	ld.w	$a2, $a0, 460
-	xvadd.w	$xr16, $xr3, $xr16
-	xvld	$xr3, $a0, 8
-	xvinsgr2vr.w	$xr5, $a1, 5
+	xvld	$xr29, $a0, 8
+	xvinsgr2vr.w	$xr28, $a1, 4
 	ld.w	$a1, $a0, 72
-	xvinsgr2vr.w	$xr5, $a3, 6
-	ld.w	$a3, $a0, 136
-	xvinsgr2vr.w	$xr5, $a2, 7
-	ld.w	$a2, $a0, 200
-	xvinsgr2vr.w	$xr3, $a1, 1
-	ld.w	$a1, $a0, 264
-	xvinsgr2vr.w	$xr3, $a3, 2
-	ld.w	$a3, $a0, 328
-	xvinsgr2vr.w	$xr3, $a2, 3
-	ld.w	$a2, $a0, 392
-	xvinsgr2vr.w	$xr3, $a1, 4
-	ld.w	$a1, $a0, 456
-	xvadd.w	$xr16, $xr2, $xr16
-	xvld	$xr2, $a0, 4
-	xvinsgr2vr.w	$xr3, $a3, 5
-	ld.w	$a3, $a0, 68
-	xvinsgr2vr.w	$xr3, $a2, 6
-	ld.w	$a2, $a0, 132
-	xvinsgr2vr.w	$xr3, $a1, 7
-	ld.w	$a1, $a0, 196
-	xvinsgr2vr.w	$xr2, $a3, 1
+	xvinsgr2vr.w	$xr28, $a4, 5
+	ld.w	$a4, $a0, 136
+	xvinsgr2vr.w	$xr28, $a3, 6
+	ld.w	$a3, $a0, 200
+	xvinsgr2vr.w	$xr28, $a2, 7
+	ld.w	$a2, $a0, 264
+	xvinsgr2vr.w	$xr29, $a1, 1
+	ld.w	$a1, $a0, 328
+	xvinsgr2vr.w	$xr29, $a4, 2
+	ld.w	$a4, $a0, 392
+	xvinsgr2vr.w	$xr29, $a3, 3
+	ld.w	$a3, $a0, 456
+	xvld	$xr30, $a0, 4
+	xvinsgr2vr.w	$xr29, $a2, 4
+	ld.w	$a2, $a0, 68
+	xvinsgr2vr.w	$xr29, $a1, 5
+	ld.w	$a1, $a0, 132
+	xvinsgr2vr.w	$xr29, $a4, 6
+	ld.w	$a4, $a0, 196
+	xvinsgr2vr.w	$xr29, $a3, 7
 	ld.w	$a3, $a0, 260
-	xvinsgr2vr.w	$xr2, $a2, 2
+	xvinsgr2vr.w	$xr30, $a2, 1
 	ld.w	$a2, $a0, 324
-	xvinsgr2vr.w	$xr2, $a1, 3
+	xvinsgr2vr.w	$xr30, $a1, 2
 	ld.w	$a1, $a0, 388
-	xvinsgr2vr.w	$xr2, $a3, 4
-	ld.w	$a3, $a0, 452
-	xvadd.w	$xr0, $xr0, $xr16
-	xvneg.w	$xr16, $xr1
-	xvmax.w	$xr1, $xr1, $xr16
-	xvneg.w	$xr16, $xr4
-	xvmax.w	$xr4, $xr4, $xr16
-	xvneg.w	$xr16, $xr6
-	xvmax.w	$xr6, $xr6, $xr16
-	xvneg.w	$xr16, $xr7
-	xvmax.w	$xr7, $xr7, $xr16
-	xvneg.w	$xr16, $xr9
-	xvmax.w	$xr9, $xr9, $xr16
-	xvneg.w	$xr16, $xr11
-	xvmax.w	$xr11, $xr11, $xr16
-	xvneg.w	$xr16, $xr13
-	xvmax.w	$xr13, $xr13, $xr16
-	xvneg.w	$xr16, $xr15
-	xvmax.w	$xr15, $xr15, $xr16
-	xvneg.w	$xr16, $xr14
-	xvmax.w	$xr14, $xr14, $xr16
-	xvneg.w	$xr16, $xr12
-	xvmax.w	$xr12, $xr12, $xr16
-	xvneg.w	$xr16, $xr10
-	xvmax.w	$xr10, $xr10, $xr16
-	xvneg.w	$xr16, $xr8
-	xvmax.w	$xr8, $xr8, $xr16
-	xvneg.w	$xr16, $xr5
-	xvmax.w	$xr5, $xr5, $xr16
-	xvneg.w	$xr16, $xr3
-	xvmax.w	$xr3, $xr3, $xr16
-	xvld	$xr16, $a0, 0
-	xvinsgr2vr.w	$xr2, $a2, 5
-	ld.w	$a2, $a0, 64
-	xvinsgr2vr.w	$xr2, $a1, 6
-	ld.w	$a1, $a0, 128
-	xvinsgr2vr.w	$xr2, $a3, 7
-	ld.w	$a3, $a0, 192
-	xvinsgr2vr.w	$xr16, $a2, 1
-	ld.w	$a2, $a0, 256
-	xvinsgr2vr.w	$xr16, $a1, 2
-	ld.w	$a1, $a0, 320
-	xvinsgr2vr.w	$xr16, $a3, 3
-	ld.w	$a3, $a0, 384
+	xvinsgr2vr.w	$xr30, $a4, 3
+	ld.w	$a4, $a0, 452
+	xvld	$xr31, $a0, 0
+	xvinsgr2vr.w	$xr30, $a3, 4
+	ld.w	$a3, $a0, 64
+	xvinsgr2vr.w	$xr30, $a2, 5
+	ld.w	$a2, $a0, 128
+	xvinsgr2vr.w	$xr30, $a1, 6
+	ld.w	$a1, $a0, 192
+	xvinsgr2vr.w	$xr30, $a4, 7
+	ld.w	$a4, $a0, 256
+	xvinsgr2vr.w	$xr31, $a3, 1
+	ld.w	$a3, $a0, 320
+	xvinsgr2vr.w	$xr31, $a2, 2
+	ld.w	$a2, $a0, 384
+	xvinsgr2vr.w	$xr31, $a1, 3
+	xvinsgr2vr.w	$xr31, $a4, 4
+	xvinsgr2vr.w	$xr31, $a3, 5
+	xvinsgr2vr.w	$xr31, $a2, 6
 	ld.w	$a0, $a0, 448
-	xvinsgr2vr.w	$xr16, $a2, 4
-	xvinsgr2vr.w	$xr16, $a1, 5
-	xvinsgr2vr.w	$xr16, $a3, 6
-	xvinsgr2vr.w	$xr16, $a0, 7
-	xvadda.w	$xr2, $xr2, $xr16
-	xvadd.w	$xr2, $xr3, $xr2
-	xvadd.w	$xr2, $xr5, $xr2
-	xvadd.w	$xr2, $xr8, $xr2
-	xvadd.w	$xr2, $xr10, $xr2
-	xvadd.w	$xr2, $xr12, $xr2
-	xvadd.w	$xr2, $xr14, $xr2
-	xvadd.w	$xr2, $xr15, $xr2
-	xvadd.w	$xr2, $xr13, $xr2
-	xvadd.w	$xr2, $xr11, $xr2
-	xvadd.w	$xr2, $xr9, $xr2
-	xvadd.w	$xr2, $xr7, $xr2
-	xvadd.w	$xr2, $xr6, $xr2
-	xvadd.w	$xr2, $xr4, $xr2
+	xvsigncov.w	$xr0, $xr0, $xr0
+	xvsigncov.w	$xr1, $xr1, $xr1
+	xvsigncov.w	$xr2, $xr2, $xr2
+	xvsigncov.w	$xr3, $xr3, $xr3
+	xvsigncov.w	$xr4, $xr4, $xr4
+	xvsigncov.w	$xr5, $xr5, $xr5
+	xvsigncov.w	$xr6, $xr6, $xr6
+	xvsigncov.w	$xr7, $xr7, $xr7
+	xvsigncov.w	$xr8, $xr8, $xr8
+	xvsigncov.w	$xr9, $xr9, $xr9
+	xvsigncov.w	$xr10, $xr10, $xr10
+	xvsigncov.w	$xr11, $xr11, $xr11
+	xvsigncov.w	$xr12, $xr12, $xr12
+	xvsigncov.w	$xr13, $xr13, $xr13
+	xvadda.w	$xr14, $xr14, $xr15
+	xvadd.w	$xr13, $xr13, $xr14
+	xvadd.w	$xr12, $xr12, $xr13
+	xvadd.w	$xr11, $xr11, $xr12
+	xvadd.w	$xr10, $xr10, $xr11
+	xvadd.w	$xr9, $xr9, $xr10
+	xvadd.w	$xr8, $xr8, $xr9
+	xvadd.w	$xr7, $xr7, $xr8
+	xvadd.w	$xr6, $xr6, $xr7
+	xvadd.w	$xr5, $xr5, $xr6
+	xvadd.w	$xr4, $xr4, $xr5
+	xvadd.w	$xr3, $xr3, $xr4
+	xvadd.w	$xr2, $xr2, $xr3
+	xvadd.w	$xr1, $xr1, $xr2
+	xvadd.w	$xr0, $xr0, $xr1
+	xvsigncov.w	$xr1, $xr16, $xr16
+	xvsigncov.w	$xr2, $xr17, $xr17
+	xvsigncov.w	$xr3, $xr18, $xr18
+	xvsigncov.w	$xr4, $xr19, $xr19
+	xvsigncov.w	$xr5, $xr20, $xr20
+	xvsigncov.w	$xr6, $xr21, $xr21
+	xvsigncov.w	$xr7, $xr22, $xr22
+	xvsigncov.w	$xr8, $xr23, $xr23
+	xvsigncov.w	$xr9, $xr24, $xr24
+	xvsigncov.w	$xr10, $xr25, $xr25
+	xvsigncov.w	$xr11, $xr26, $xr26
+	xvsigncov.w	$xr12, $xr27, $xr27
+	xvsigncov.w	$xr13, $xr28, $xr28
+	xvsigncov.w	$xr14, $xr29, $xr29
+	xvinsgr2vr.w	$xr31, $a0, 7
+	xvadda.w	$xr15, $xr30, $xr31
+	xvadd.w	$xr14, $xr14, $xr15
+	xvadd.w	$xr13, $xr13, $xr14
+	xvadd.w	$xr12, $xr12, $xr13
+	xvadd.w	$xr11, $xr11, $xr12
+	xvadd.w	$xr10, $xr10, $xr11
+	xvadd.w	$xr9, $xr9, $xr10
+	xvadd.w	$xr8, $xr8, $xr9
+	xvadd.w	$xr7, $xr7, $xr8
+	xvadd.w	$xr6, $xr6, $xr7
+	xvadd.w	$xr5, $xr5, $xr6
+	xvadd.w	$xr4, $xr4, $xr5
+	xvadd.w	$xr3, $xr3, $xr4
+	xvadd.w	$xr2, $xr2, $xr3
 	xvadd.w	$xr1, $xr1, $xr2
 	xvadd.w	$xr0, $xr0, $xr1
 	xvhaddw.d.w	$xr0, $xr0, $xr0
@@ -1230,6 +1183,15 @@ calc_MAD:                               # @calc_MAD
 	xvadd.d	$xr0, $xr1, $xr0
 	xvpickve2gr.d	$a0, $xr0, 0
 	addi.w	$a0, $a0, 0
+	fld.d	$fs7, $sp, 16                   # 8-byte Folded Reload
+	fld.d	$fs6, $sp, 24                   # 8-byte Folded Reload
+	fld.d	$fs5, $sp, 32                   # 8-byte Folded Reload
+	fld.d	$fs4, $sp, 40                   # 8-byte Folded Reload
+	fld.d	$fs3, $sp, 48                   # 8-byte Folded Reload
+	fld.d	$fs2, $sp, 56                   # 8-byte Folded Reload
+	fld.d	$fs1, $sp, 64                   # 8-byte Folded Reload
+	fld.d	$fs0, $sp, 72                   # 8-byte Folded Reload
+	addi.d	$sp, $sp, 80
 	ret
 .Lfunc_end1:
 	.size	calc_MAD, .Lfunc_end1-calc_MAD
diff --git a/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/weighted_prediction.s b/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/weighted_prediction.s
index 7a864e8..3b4d2d6 100644
--- a/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/weighted_prediction.s
+++ b/results/MultiSource/Applications/JM/lencod/CMakeFiles/lencod.dir/weighted_prediction.s
@@ -2170,8 +2170,7 @@ test_wp_B_slice:                        # @test_wp_B_slice
 	vsrli.b	$vr16, $vr15, 7
 	vadd.b	$vr15, $vr15, $vr16
 	vsrai.b	$vr15, $vr15, 1
-	vneg.b	$vr16, $vr15
-	vmax.b	$vr15, $vr15, $vr16
+	vsigncov.b	$vr15, $vr15, $vr15
 	vilvl.b	$vr15, $vr3, $vr15
 	vbitseti.h	$vr15, $vr15, 14
 	vpickev.h	$vr12, $vr12, $vr12
diff --git a/results/MultiSource/Applications/SIBsim4/CMakeFiles/SIBsim4.dir/sim4b1.s b/results/MultiSource/Applications/SIBsim4/CMakeFiles/SIBsim4.dir/sim4b1.s
index 03a0e2d..8ca85f3 100644
--- a/results/MultiSource/Applications/SIBsim4/CMakeFiles/SIBsim4.dir/sim4b1.s
+++ b/results/MultiSource/Applications/SIBsim4/CMakeFiles/SIBsim4.dir/sim4b1.s
@@ -8423,8 +8423,7 @@ link_msps:                              # @link_msps
 	ld.d	$t6, $a7, 0
 	vinsgr2vr.d	$vr1, $t6, 0
 	vsub.w	$vr0, $vr0, $vr1
-	vneg.w	$vr1, $vr0
-	vmax.w	$vr0, $vr0, $vr1
+	vsigncov.w	$vr0, $vr0, $vr0
 	vsrli.w	$vr0, $vr0, 15
 	vpickve2gr.w	$t6, $vr0, 0
 	vpickve2gr.w	$t7, $vr0, 1
diff --git a/results/MultiSource/Benchmarks/ASCI_Purple/SMG2000/CMakeFiles/smg2000.dir/struct_stencil.s b/results/MultiSource/Benchmarks/ASCI_Purple/SMG2000/CMakeFiles/smg2000.dir/struct_stencil.s
index 2c86d2d..11d5ef0 100644
--- a/results/MultiSource/Benchmarks/ASCI_Purple/SMG2000/CMakeFiles/smg2000.dir/struct_stencil.s
+++ b/results/MultiSource/Benchmarks/ASCI_Purple/SMG2000/CMakeFiles/smg2000.dir/struct_stencil.s
@@ -84,10 +84,8 @@ hypre_StructStencilCreate:              # @hypre_StructStencilCreate
 	vinsgr2vr.w	$vr4, $a7, 2
 	vinsgr2vr.w	$vr4, $t0, 3
 	xvpermi.q	$xr4, $xr2, 2
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
-	xvneg.w	$xr3, $xr4
-	xvmax.w	$xr3, $xr4, $xr3
+	xvsigncov.w	$xr2, $xr3, $xr3
+	xvsigncov.w	$xr3, $xr4, $xr4
 	xvmax.wu	$xr0, $xr2, $xr0
 	xvmax.wu	$xr1, $xr3, $xr1
 	ld.w	$a5, $a3, -92
@@ -124,10 +122,8 @@ hypre_StructStencilCreate:              # @hypre_StructStencilCreate
 	vinsgr2vr.w	$vr4, $a7, 2
 	vinsgr2vr.w	$vr4, $t0, 3
 	xvpermi.q	$xr4, $xr2, 2
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
-	xvneg.w	$xr3, $xr4
-	xvmax.w	$xr3, $xr4, $xr3
+	xvsigncov.w	$xr2, $xr3, $xr3
+	xvsigncov.w	$xr3, $xr4, $xr4
 	xvmax.wu	$xr0, $xr2, $xr0
 	xvmax.wu	$xr1, $xr3, $xr1
 	ld.w	$a5, $a3, -88
@@ -164,10 +160,8 @@ hypre_StructStencilCreate:              # @hypre_StructStencilCreate
 	vinsgr2vr.w	$vr4, $a7, 2
 	vinsgr2vr.w	$vr4, $t0, 3
 	xvpermi.q	$xr4, $xr2, 2
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
-	xvneg.w	$xr3, $xr4
-	xvmax.w	$xr3, $xr4, $xr3
+	xvsigncov.w	$xr2, $xr3, $xr3
+	xvsigncov.w	$xr3, $xr4, $xr4
 	xvmax.wu	$xr0, $xr2, $xr0
 	xvmax.wu	$xr1, $xr3, $xr1
 	addi.d	$a4, $a4, -16
@@ -206,8 +200,7 @@ hypre_StructStencilCreate:              # @hypre_StructStencilCreate
 	vinsgr2vr.w	$vr1, $a5, 1
 	vinsgr2vr.w	$vr1, $a6, 2
 	vinsgr2vr.w	$vr1, $a7, 3
-	vneg.w	$vr2, $vr1
-	vmax.w	$vr1, $vr1, $vr2
+	vsigncov.w	$vr1, $vr1, $vr1
 	vmax.wu	$vr0, $vr1, $vr0
 	ld.w	$a4, $a3, -20
 	ld.w	$a5, $a3, -8
@@ -217,8 +210,7 @@ hypre_StructStencilCreate:              # @hypre_StructStencilCreate
 	vinsgr2vr.w	$vr1, $a5, 1
 	vinsgr2vr.w	$vr1, $a6, 2
 	vinsgr2vr.w	$vr1, $a7, 3
-	vneg.w	$vr2, $vr1
-	vmax.w	$vr1, $vr1, $vr2
+	vsigncov.w	$vr1, $vr1, $vr1
 	vmax.wu	$vr0, $vr1, $vr0
 	ld.w	$a4, $a3, -16
 	ld.w	$a5, $a3, -4
@@ -228,8 +220,7 @@ hypre_StructStencilCreate:              # @hypre_StructStencilCreate
 	vinsgr2vr.w	$vr1, $a5, 1
 	vinsgr2vr.w	$vr1, $a6, 2
 	vinsgr2vr.w	$vr1, $a7, 3
-	vneg.w	$vr2, $vr1
-	vmax.w	$vr1, $vr1, $vr2
+	vsigncov.w	$vr1, $vr1, $vr1
 	vmax.wu	$vr0, $vr1, $vr0
 	addi.d	$a2, $a2, 4
 	addi.d	$a3, $a3, 48
@@ -594,10 +585,8 @@ hypre_StructStencilSymmetrize:          # @hypre_StructStencilSymmetrize
 	vinsgr2vr.w	$vr4, $a7, 2
 	vinsgr2vr.w	$vr4, $t0, 3
 	xvpermi.q	$xr4, $xr2, 2
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
-	xvneg.w	$xr3, $xr4
-	xvmax.w	$xr3, $xr4, $xr3
+	xvsigncov.w	$xr2, $xr3, $xr3
+	xvsigncov.w	$xr3, $xr4, $xr4
 	xvmax.wu	$xr0, $xr2, $xr0
 	xvmax.wu	$xr1, $xr3, $xr1
 	ld.w	$a5, $a3, -92
@@ -634,10 +623,8 @@ hypre_StructStencilSymmetrize:          # @hypre_StructStencilSymmetrize
 	vinsgr2vr.w	$vr4, $a7, 2
 	vinsgr2vr.w	$vr4, $t0, 3
 	xvpermi.q	$xr4, $xr2, 2
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
-	xvneg.w	$xr3, $xr4
-	xvmax.w	$xr3, $xr4, $xr3
+	xvsigncov.w	$xr2, $xr3, $xr3
+	xvsigncov.w	$xr3, $xr4, $xr4
 	xvmax.wu	$xr0, $xr2, $xr0
 	xvmax.wu	$xr1, $xr3, $xr1
 	ld.w	$a5, $a3, -88
@@ -674,10 +661,8 @@ hypre_StructStencilSymmetrize:          # @hypre_StructStencilSymmetrize
 	vinsgr2vr.w	$vr4, $a7, 2
 	vinsgr2vr.w	$vr4, $t0, 3
 	xvpermi.q	$xr4, $xr2, 2
-	xvneg.w	$xr2, $xr3
-	xvmax.w	$xr2, $xr3, $xr2
-	xvneg.w	$xr3, $xr4
-	xvmax.w	$xr3, $xr4, $xr3
+	xvsigncov.w	$xr2, $xr3, $xr3
+	xvsigncov.w	$xr3, $xr4, $xr4
 	xvmax.wu	$xr0, $xr2, $xr0
 	xvmax.wu	$xr1, $xr3, $xr1
 	addi.d	$a4, $a4, -16
@@ -716,8 +701,7 @@ hypre_StructStencilSymmetrize:          # @hypre_StructStencilSymmetrize
 	vinsgr2vr.w	$vr1, $a5, 1
 	vinsgr2vr.w	$vr1, $a6, 2
 	vinsgr2vr.w	$vr1, $a7, 3
-	vneg.w	$vr2, $vr1
-	vmax.w	$vr1, $vr1, $vr2
+	vsigncov.w	$vr1, $vr1, $vr1
 	vmax.wu	$vr0, $vr1, $vr0
 	ld.w	$a4, $a3, -20
 	ld.w	$a5, $a3, -8
@@ -727,8 +711,7 @@ hypre_StructStencilSymmetrize:          # @hypre_StructStencilSymmetrize
 	vinsgr2vr.w	$vr1, $a5, 1
 	vinsgr2vr.w	$vr1, $a6, 2
 	vinsgr2vr.w	$vr1, $a7, 3
-	vneg.w	$vr2, $vr1
-	vmax.w	$vr1, $vr1, $vr2
+	vsigncov.w	$vr1, $vr1, $vr1
 	vmax.wu	$vr0, $vr1, $vr0
 	ld.w	$a4, $a3, -16
 	ld.w	$a5, $a3, -4
@@ -738,8 +721,7 @@ hypre_StructStencilSymmetrize:          # @hypre_StructStencilSymmetrize
 	vinsgr2vr.w	$vr1, $a5, 1
 	vinsgr2vr.w	$vr1, $a6, 2
 	vinsgr2vr.w	$vr1, $a7, 3
-	vneg.w	$vr2, $vr1
-	vmax.w	$vr1, $vr1, $vr2
+	vsigncov.w	$vr1, $vr1, $vr1
 	vmax.wu	$vr0, $vr1, $vr0
 	addi.d	$a2, $a2, 4
 	addi.d	$a3, $a3, 48
diff --git a/results/MultiSource/Benchmarks/MiBench/consumer-lame/CMakeFiles/consumer-lame.dir/l3bitstream.s b/results/MultiSource/Benchmarks/MiBench/consumer-lame/CMakeFiles/consumer-lame.dir/l3bitstream.s
index acc0aa1..0b4d277 100644
--- a/results/MultiSource/Benchmarks/MiBench/consumer-lame/CMakeFiles/consumer-lame.dir/l3bitstream.s
+++ b/results/MultiSource/Benchmarks/MiBench/consumer-lame/CMakeFiles/consumer-lame.dir/l3bitstream.s
@@ -4028,8 +4028,7 @@ L3_huffman_coder_count1:                # @L3_huffman_coder_count1
 	sub.d	$a0, $a2, $a0
 	vinsgr2vr.w	$vr0, $a3, 0
 	vinsgr2vr.w	$vr0, $a4, 1
-	vneg.w	$vr1, $vr0
-	vmax.w	$vr0, $vr0, $vr1
+	vsigncov.w	$vr0, $vr0, $vr0
 	srai.d	$a2, $a5, 31
 	xor	$a3, $a5, $a2
 	sub.d	$a2, $a3, $a2
@@ -4631,53 +4630,50 @@ Huffmancodebits:                        # @Huffmancodebits
 	.p2align	4, , 16
 .LBB6_41:                               # =>This Inner Loop Header: Depth=1
 	st.d	$a5, $sp, 80                    # 8-byte Folded Spill
-	ld.wu	$s7, $s3, 0
-	ld.w	$a1, $a2, 72
-	addi.w	$a3, $s7, 0
-	st.d	$a3, $sp, 88                    # 8-byte Folded Spill
-	ld.w	$s2, $s3, 12
-	addi.d	$a1, $a1, 32
-	bstrpick.d	$a1, $a1, 31, 0
-	slli.d	$a2, $a1, 4
-	alsl.d	$a1, $a1, $a2, 3
-	ld.d	$a2, $sp, 56                    # 8-byte Folded Reload
-	add.d	$a1, $a2, $a1
-	srai.d	$a2, $a3, 31
-	xor	$a3, $s7, $a2
-	sub.d	$a2, $a3, $a2
-	slti	$a3, $s2, 1
-	st.d	$a3, $sp, 96                    # 8-byte Folded Spill
-	ld.d	$a3, $s3, 4
-	srai.d	$a4, $s2, 31
-	xor	$a5, $s2, $a4
-	sub.d	$a4, $a5, $a4
-	vinsgr2vr.d	$vr0, $a3, 0
-	vpickve2gr.w	$s4, $vr0, 0
-	vneg.w	$vr1, $vr0
-	vmax.w	$vr1, $vr0, $vr1
+	ld.d	$a1, $s3, 4
+	ld.wu	$s2, $s3, 0
+	ld.w	$a2, $a2, 72
+	vinsgr2vr.d	$vr0, $a1, 0
+	vsigncov.w	$vr1, $vr0, $vr0
 	vld	$vr2, $sp, 32                   # 16-byte Folded Reload
 	vsll.w	$vr1, $vr1, $vr2
-	vpickve2gr.w	$a3, $vr1, 0
-	alsl.d	$a2, $a2, $a3, 3
+	vpickve2gr.w	$a1, $vr1, 0
 	vpickve2gr.w	$a3, $vr1, 1
-	add.d	$a2, $a2, $a3
-	add.d	$a2, $a2, $a4
-	ld.d	$a3, $a1, 8
-	ld.d	$a4, $a1, 16
+	addi.w	$s6, $s2, 0
+	ld.w	$s7, $s3, 12
+	addi.d	$a2, $a2, 32
 	bstrpick.d	$a2, $a2, 31, 0
-	slli.d	$a1, $a2, 3
+	slli.d	$a4, $a2, 4
+	alsl.d	$a2, $a2, $a4, 3
+	ld.d	$a4, $sp, 56                    # 8-byte Folded Reload
+	add.d	$a2, $a4, $a2
+	srai.d	$a4, $s6, 31
+	xor	$a5, $s2, $a4
+	sub.d	$a4, $a5, $a4
+	slti	$a5, $s7, 1
+	st.d	$a5, $sp, 96                    # 8-byte Folded Spill
+	srai.d	$a5, $s7, 31
+	xor	$a6, $s7, $a5
+	sub.d	$a5, $a6, $a5
+	vpickve2gr.w	$s4, $vr0, 0
+	alsl.d	$a1, $a4, $a1, 3
+	add.d	$a1, $a1, $a3
+	add.d	$a1, $a1, $a5
+	ld.d	$a3, $a2, 8
+	ld.d	$a2, $a2, 16
+	bstrpick.d	$a4, $a1, 31, 0
+	slli.d	$a1, $a4, 3
 	ldx.w	$a1, $a3, $a1
-	ldx.bu	$s1, $a4, $a2
+	ldx.bu	$a2, $a2, $a4
+	st.d	$a2, $sp, 88                    # 8-byte Folded Spill
 	slti	$s0, $s4, 1
-	vpickve2gr.w	$s6, $vr0, 1
-	slti	$s8, $s6, 1
-	move	$a2, $s1
+	vpickve2gr.w	$s8, $vr0, 1
+	slti	$s1, $s8, 1
 	pcaddu18i	$ra, %call36(BF_addEntry)
 	jirl	$ra, $ra, 0
 	st.d	$a0, $fp, 0
-	ld.d	$a1, $sp, 88                    # 8-byte Folded Reload
-	sltu	$a1, $zero, $a1
-	srli.d	$a2, $s7, 31
+	sltu	$a1, $zero, $s6
+	srli.d	$a2, $s2, 31
 	sltui	$a3, $s4, 1
 	slli.d	$a4, $a2, 1
 	or	$a4, $a4, $s0
@@ -4688,14 +4684,14 @@ Huffmancodebits:                        # @Huffmancodebits
 	masknez	$a4, $a5, $a3
 	maskeqz	$a1, $a1, $a3
 	or	$a1, $a1, $a4
-	sltu	$a3, $zero, $s6
+	sltu	$a3, $zero, $s8
 	slli.d	$a4, $a2, 1
-	or	$a4, $a4, $s8
+	or	$a4, $a4, $s1
 	masknez	$a2, $a2, $a3
 	maskeqz	$a4, $a4, $a3
 	or	$a2, $a4, $a2
 	add.d	$a1, $a1, $a3
-	sltu	$a3, $zero, $s2
+	sltu	$a3, $zero, $s7
 	slli.d	$a4, $a2, 1
 	ld.d	$a5, $sp, 96                    # 8-byte Folded Reload
 	or	$a4, $a4, $a5
@@ -4712,7 +4708,8 @@ Huffmancodebits:                        # @Huffmancodebits
 	ld.d	$a2, $sp, 64                    # 8-byte Folded Reload
 	st.d	$a0, $fp, 0
 	add.d	$a1, $s2, $s5
-	add.d	$s5, $a1, $s1
+	ld.d	$a4, $sp, 88                    # 8-byte Folded Reload
+	add.d	$s5, $a1, $a4
 	addi.d	$a5, $a5, 4
 	addi.d	$s3, $s3, 16
 	blt	$a5, $a3, .LBB6_41
diff --git a/results/SingleSource/Regression/C/gcc-c-torture/execute/CMakeFiles/GCC-C-execute-pr33142.dir/pr33142.s b/results/SingleSource/Regression/C/gcc-c-torture/execute/CMakeFiles/GCC-C-execute-pr33142.dir/pr33142.s
index bcec72f..9987843 100644
--- a/results/SingleSource/Regression/C/gcc-c-torture/execute/CMakeFiles/GCC-C-execute-pr33142.dir/pr33142.s
+++ b/results/SingleSource/Regression/C/gcc-c-torture/execute/CMakeFiles/GCC-C-execute-pr33142.dir/pr33142.s
@@ -13,8 +13,7 @@ lisp_atan2:                             # @lisp_atan2
 # %bb.2:
 	vinsgr2vr.w	$vr0, $a1, 0
 	vinsgr2vr.w	$vr0, $a2, 1
-	vneg.w	$vr1, $vr0
-	vmax.w	$vr0, $vr0, $vr1
+	vsigncov.w	$vr0, $vr0, $vr0
 	vpickve2gr.w	$a0, $vr0, 0
 	vpickve2gr.w	$a1, $vr0, 1
 	sltu	$a0, $a1, $a0