[llvm] c1b6ed4 - [AArch64] Fix postinc operands for Neoverse-N2 scheduling

David Green via llvm-commits llvm-commits at lists.llvm.org
Wed Oct 11 09:57:49 PDT 2023


Author: David Green
Date: 2023-10-11T17:57:42+01:00
New Revision: c1b6ed42b75a17d1718aa377f76633b27e15a4e2

URL: https://github.com/llvm/llvm-project/commit/c1b6ed42b75a17d1718aa377f76633b27e15a4e2
DIFF: https://github.com/llvm/llvm-project/commit/c1b6ed42b75a17d1718aa377f76633b27e15a4e2.diff

LOG: [AArch64] Fix postinc operands for Neoverse-N2 scheduling

Similar to D159254, this fixes the order of WriteAdr operands on post/pre-inc
loads/stores in the Neoverse-N2 scheduling model.

Added: 
    

Modified: 
    llvm/lib/Target/AArch64/AArch64SchedNeoverseN2.td
    llvm/test/tools/llvm-mca/AArch64/Neoverse/N2-writeback.s

Removed: 
    


################################################################################
diff  --git a/llvm/lib/Target/AArch64/AArch64SchedNeoverseN2.td b/llvm/lib/Target/AArch64/AArch64SchedNeoverseN2.td
index a6e28e653e33c17..517d0da7f47f428 100644
--- a/llvm/lib/Target/AArch64/AArch64SchedNeoverseN2.td
+++ b/llvm/lib/Target/AArch64/AArch64SchedNeoverseN2.td
@@ -742,7 +742,7 @@ def : SchedAlias<WriteLDIdx, N2Write_4cyc_1I_1L>;
 // Load pair, signed immed offset, signed words
 def : InstRW<[N2Write_5cyc_1M0, WriteLDHi], (instrs LDPSWi)>;
 // Load pair, immed post-index or immed pre-index, signed words
-def : InstRW<[N2Write_5cyc_1M0, WriteLDHi, WriteAdr],
+def : InstRW<[WriteAdr, N2Write_5cyc_1M0, WriteLDHi],
              (instregex "^LDPSW(post|pre)$")>;
 
 // Store instructions
@@ -860,7 +860,7 @@ def : InstRW<[N2Write_6cyc_1L], (instregex "^LDR[SDQ]l$",
 // Load vector reg, immed post-index
 def : InstRW<[N2Write_6cyc_1I_1L, WriteI], (instregex "^LDR[BHSDQ]post$")>;
 // Load vector reg, immed pre-index
-def : InstRW<[N2Write_6cyc_1I_1L, WriteAdr], (instregex "^LDR[BHSDQ]pre$")>;
+def : InstRW<[WriteAdr, N2Write_6cyc_1I_1L], (instregex "^LDR[BHSDQ]pre$")>;
 
 // Load vector reg, unsigned immed
 def : InstRW<[N2Write_6cyc_1L], (instregex "^LDR[BHSDQ]ui$")>;
@@ -883,12 +883,12 @@ def : InstRW<[N2Write_6cyc_2L, WriteLDHi], (instrs LDPQi, LDNPQi)>;
 
 // Load vector pair, immed post-index, S/D-form
 // Load vector pair, immed pre-index, S/D-form
-def : InstRW<[N2Write_6cyc_1I_1L, WriteLDHi, WriteAdr],
+def : InstRW<[WriteAdr, N2Write_6cyc_1I_1L, WriteLDHi],
              (instregex "^LDP[SD](pre|post)$")>;
 
 // Load vector pair, immed post-index, Q-form
 // Load vector pair, immed pre-index, Q-form
-def : InstRW<[N2Write_6cyc_2I_2L, WriteLDHi, WriteAdr], (instrs LDPQpost,
+def : InstRW<[WriteAdr, N2Write_6cyc_2I_2L, WriteLDHi], (instrs LDPQpost,
                                                                 LDPQpre)>;
 
 // FP store instructions
@@ -1238,223 +1238,223 @@ def : InstRW<[N2Write_5cyc_1M0_1V], (instregex "^INSvi(8|16|32|64)gpr$")>;
 
 // ASIMD load, 1 element, multiple, 1 reg, D-form
 def : InstRW<[N2Write_6cyc_1L], (instregex "^LD1Onev(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_6cyc_1L, WriteAdr],
+def : InstRW<[WriteAdr, N2Write_6cyc_1L],
              (instregex "^LD1Onev(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD load, 1 element, multiple, 1 reg, Q-form
 def : InstRW<[N2Write_6cyc_1L], (instregex "^LD1Onev(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_6cyc_1L, WriteAdr],
+def : InstRW<[WriteAdr, N2Write_6cyc_1L],
              (instregex "^LD1Onev(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD load, 1 element, multiple, 2 reg, D-form
 def : InstRW<[N2Write_6cyc_2L], (instregex "^LD1Twov(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_6cyc_2L, WriteAdr],
+def : InstRW<[WriteAdr, N2Write_6cyc_2L],
              (instregex "^LD1Twov(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD load, 1 element, multiple, 2 reg, Q-form
 def : InstRW<[N2Write_6cyc_2L], (instregex "^LD1Twov(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_6cyc_2L, WriteAdr],
+def : InstRW<[WriteAdr, N2Write_6cyc_2L],
              (instregex "^LD1Twov(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD load, 1 element, multiple, 3 reg, D-form
 def : InstRW<[N2Write_6cyc_3L], (instregex "^LD1Threev(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_6cyc_3L, WriteAdr],
+def : InstRW<[WriteAdr, N2Write_6cyc_3L],
              (instregex "^LD1Threev(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD load, 1 element, multiple, 3 reg, Q-form
 def : InstRW<[N2Write_6cyc_3L], (instregex "^LD1Threev(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_6cyc_3L, WriteAdr],
+def : InstRW<[WriteAdr, N2Write_6cyc_3L],
              (instregex "^LD1Threev(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD load, 1 element, multiple, 4 reg, D-form
 def : InstRW<[N2Write_7cyc_4L], (instregex "^LD1Fourv(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_7cyc_4L, WriteAdr],
+def : InstRW<[WriteAdr, N2Write_7cyc_4L],
              (instregex "^LD1Fourv(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD load, 1 element, multiple, 4 reg, Q-form
 def : InstRW<[N2Write_7cyc_4L], (instregex "^LD1Fourv(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_7cyc_4L, WriteAdr],
+def : InstRW<[WriteAdr, N2Write_7cyc_4L],
              (instregex "^LD1Fourv(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD load, 1 element, one lane, B/H/S
 // ASIMD load, 1 element, one lane, D
 def : InstRW<[N2Write_8cyc_1L_1V],           (instregex "LD1i(8|16|32|64)$")>;
-def : InstRW<[N2Write_8cyc_1L_1V, WriteAdr], (instregex "LD1i(8|16|32|64)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_1L_1V], (instregex "LD1i(8|16|32|64)_POST$")>;
 
 // ASIMD load, 1 element, all lanes, D-form, B/H/S
 // ASIMD load, 1 element, all lanes, D-form, D
 def : InstRW<[N2Write_8cyc_1L_1V],           (instregex "LD1Rv(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_8cyc_1L_1V, WriteAdr], (instregex "LD1Rv(8b|4h|2s|1d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_1L_1V], (instregex "LD1Rv(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD load, 1 element, all lanes, Q-form
 def : InstRW<[N2Write_8cyc_1L_1V],           (instregex "LD1Rv(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_8cyc_1L_1V, WriteAdr], (instregex "LD1Rv(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_1L_1V], (instregex "LD1Rv(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD load, 2 element, multiple, D-form, B/H/S
 def : InstRW<[N2Write_8cyc_1L_2V],           (instregex "LD2Twov(8b|4h|2s)$")>;
-def : InstRW<[N2Write_8cyc_1L_2V, WriteAdr], (instregex "LD2Twov(8b|4h|2s)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_1L_2V], (instregex "LD2Twov(8b|4h|2s)_POST$")>;
 
 // ASIMD load, 2 element, multiple, Q-form, B/H/S
 // ASIMD load, 2 element, multiple, Q-form, D
 def : InstRW<[N2Write_8cyc_2L_2V],           (instregex "LD2Twov(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_8cyc_2L_2V, WriteAdr], (instregex "LD2Twov(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_2L_2V], (instregex "LD2Twov(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD load, 2 element, one lane, B/H
 // ASIMD load, 2 element, one lane, S
 // ASIMD load, 2 element, one lane, D
 def : InstRW<[N2Write_8cyc_1L_2V],           (instregex "LD2i(8|16|32|64)$")>;
-def : InstRW<[N2Write_8cyc_1L_2V, WriteAdr], (instregex "LD2i(8|16|32|64)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_1L_2V], (instregex "LD2i(8|16|32|64)_POST$")>;
 
 // ASIMD load, 2 element, all lanes, D-form, B/H/S
 // ASIMD load, 2 element, all lanes, D-form, D
 def : InstRW<[N2Write_8cyc_1L_2V],            (instregex "LD2Rv(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_8cyc_1L_2V, WriteAdr],  (instregex "LD2Rv(8b|4h|2s|1d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_1L_2V],  (instregex "LD2Rv(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD load, 2 element, all lanes, Q-form
 def : InstRW<[N2Write_8cyc_1L_2V],           (instregex "LD2Rv(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_8cyc_1L_2V, WriteAdr], (instregex "LD2Rv(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_1L_2V], (instregex "LD2Rv(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD load, 3 element, multiple, D-form, B/H/S
 def : InstRW<[N2Write_8cyc_2L_3V],           (instregex "LD3Threev(8b|4h|2s)$")>;
-def : InstRW<[N2Write_8cyc_2L_3V, WriteAdr], (instregex "LD3Threev(8b|4h|2s)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_2L_3V], (instregex "LD3Threev(8b|4h|2s)_POST$")>;
 
 // ASIMD load, 3 element, multiple, Q-form, B/H/S
 def : InstRW<[N2Write_8cyc_3L_3V],           (instregex "LD3Threev(16b|8h|4s)$")>;
-def : InstRW<[N2Write_8cyc_3L_3V, WriteAdr], (instregex "LD3Threev(16b|8h|4s)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_3L_3V], (instregex "LD3Threev(16b|8h|4s)_POST$")>;
 
 // ASIMD load, 3 element, multiple, Q-form, D
 def : InstRW<[N2Write_8cyc_3L_3V],           (instregex "LD3Threev(2d)$")>;
-def : InstRW<[N2Write_8cyc_3L_3V, WriteAdr], (instregex "LD3Threev(2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_3L_3V], (instregex "LD3Threev(2d)_POST$")>;
 
 // ASIMD load, 3 element, one lane, B/H
 // ASIMD load, 3 element, one lane, S
 // ASIMD load, 3 element, one lane, D
 def : InstRW<[N2Write_8cyc_2L_3V],           (instregex "LD3i(8|16|32|64)$")>;
-def : InstRW<[N2Write_8cyc_2L_3V, WriteAdr], (instregex "LD3i(8|16|32|64)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_2L_3V], (instregex "LD3i(8|16|32|64)_POST$")>;
 
 // ASIMD load, 3 element, all lanes, D-form, B/H/S
 // ASIMD load, 3 element, all lanes, D-form, D
 def : InstRW<[N2Write_8cyc_2L_3V],           (instregex "LD3Rv(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_8cyc_2L_3V, WriteAdr], (instregex "LD3Rv(8b|4h|2s|1d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_2L_3V], (instregex "LD3Rv(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD load, 3 element, all lanes, Q-form, B/H/S
 // ASIMD load, 3 element, all lanes, Q-form, D
 def : InstRW<[N2Write_8cyc_3L_3V],           (instregex "LD3Rv(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_8cyc_3L_3V, WriteAdr], (instregex "LD3Rv(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_3L_3V], (instregex "LD3Rv(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD load, 4 element, multiple, D-form, B/H/S
 def : InstRW<[N2Write_8cyc_3L_4V],           (instregex "LD4Fourv(8b|4h|2s)$")>;
-def : InstRW<[N2Write_8cyc_3L_4V, WriteAdr], (instregex "LD4Fourv(8b|4h|2s)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_3L_4V], (instregex "LD4Fourv(8b|4h|2s)_POST$")>;
 
 // ASIMD load, 4 element, multiple, Q-form, B/H/S
 // ASIMD load, 4 element, multiple, Q-form, D
 def : InstRW<[N2Write_9cyc_4L_4V],           (instregex "LD4Fourv(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_9cyc_4L_4V, WriteAdr], (instregex "LD4Fourv(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_9cyc_4L_4V], (instregex "LD4Fourv(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD load, 4 element, one lane, B/H
 // ASIMD load, 4 element, one lane, S
 // ASIMD load, 4 element, one lane, D
 def : InstRW<[N2Write_8cyc_3L_4V],           (instregex "LD4i(8|16|32|64)$")>;
-def : InstRW<[N2Write_8cyc_3L_4V, WriteAdr], (instregex "LD4i(8|16|32|64)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_3L_4V], (instregex "LD4i(8|16|32|64)_POST$")>;
 
 // ASIMD load, 4 element, all lanes, D-form, B/H/S
 // ASIMD load, 4 element, all lanes, D-form, D
 def : InstRW<[N2Write_8cyc_3L_4V],              (instregex "LD4Rv(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_8cyc_3L_4V, WriteAdr],    (instregex "LD4Rv(8b|4h|2s|1d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_3L_4V],    (instregex "LD4Rv(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD load, 4 element, all lanes, Q-form, B/H/S
 // ASIMD load, 4 element, all lanes, Q-form, D
 def : InstRW<[N2Write_8cyc_4L_4V],            (instregex "LD4Rv(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_8cyc_4L_4V, WriteAdr],  (instregex "LD4Rv(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_8cyc_4L_4V],  (instregex "LD4Rv(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD store instructions
 // -----------------------------------------------------------------------------
 
 // ASIMD store, 1 element, multiple, 1 reg, D-form
 def : InstRW<[N2Write_2cyc_1L01_1V],           (instregex "ST1Onev(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_2cyc_1L01_1V, WriteAdr], (instregex "ST1Onev(8b|4h|2s|1d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_2cyc_1L01_1V], (instregex "ST1Onev(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD store, 1 element, multiple, 1 reg, Q-form
 def : InstRW<[N2Write_2cyc_1L01_1V],           (instregex "ST1Onev(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_2cyc_1L01_1V, WriteAdr], (instregex "ST1Onev(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_2cyc_1L01_1V], (instregex "ST1Onev(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD store, 1 element, multiple, 2 reg, D-form
 def : InstRW<[N2Write_2cyc_1L01_1V],           (instregex "ST1Twov(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_2cyc_1L01_1V, WriteAdr], (instregex "ST1Twov(8b|4h|2s|1d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_2cyc_1L01_1V], (instregex "ST1Twov(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD store, 1 element, multiple, 2 reg, Q-form
 def : InstRW<[N2Write_2cyc_2L01_2V],           (instregex "ST1Twov(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_2cyc_2L01_2V, WriteAdr], (instregex "ST1Twov(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_2cyc_2L01_2V], (instregex "ST1Twov(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD store, 1 element, multiple, 3 reg, D-form
 def : InstRW<[N2Write_2cyc_2L01_2V],           (instregex "ST1Threev(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_2cyc_2L01_2V, WriteAdr], (instregex "ST1Threev(8b|4h|2s|1d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_2cyc_2L01_2V], (instregex "ST1Threev(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD store, 1 element, multiple, 3 reg, Q-form
 def : InstRW<[N2Write_2cyc_3L01_3V],           (instregex "ST1Threev(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_2cyc_3L01_3V, WriteAdr], (instregex "ST1Threev(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_2cyc_3L01_3V], (instregex "ST1Threev(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD store, 1 element, multiple, 4 reg, D-form
 def : InstRW<[N2Write_2cyc_2L01_2V],           (instregex "ST1Fourv(8b|4h|2s|1d)$")>;
-def : InstRW<[N2Write_2cyc_2L01_2V, WriteAdr], (instregex "ST1Fourv(8b|4h|2s|1d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_2cyc_2L01_2V], (instregex "ST1Fourv(8b|4h|2s|1d)_POST$")>;
 
 // ASIMD store, 1 element, multiple, 4 reg, Q-form
 def : InstRW<[N2Write_2cyc_4L01_4V],           (instregex "ST1Fourv(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_2cyc_4L01_4V, WriteAdr], (instregex "ST1Fourv(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_2cyc_4L01_4V], (instregex "ST1Fourv(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD store, 1 element, one lane, B/H/S
 // ASIMD store, 1 element, one lane, D
 def : InstRW<[N2Write_4cyc_1L01_1V],           (instregex "ST1i(8|16|32|64)$")>;
-def : InstRW<[N2Write_4cyc_1L01_1V, WriteAdr], (instregex "ST1i(8|16|32|64)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_4cyc_1L01_1V], (instregex "ST1i(8|16|32|64)_POST$")>;
 
 // ASIMD store, 2 element, multiple, D-form, B/H/S
 def : InstRW<[N2Write_4cyc_1L01_1V],           (instregex "ST2Twov(8b|4h|2s)$")>;
-def : InstRW<[N2Write_4cyc_1L01_1V, WriteAdr], (instregex "ST2Twov(8b|4h|2s)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_4cyc_1L01_1V], (instregex "ST2Twov(8b|4h|2s)_POST$")>;
 
 // ASIMD store, 2 element, multiple, Q-form, B/H/S
 // ASIMD store, 2 element, multiple, Q-form, D
 def : InstRW<[N2Write_4cyc_2L01_2V],           (instregex "ST2Twov(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_4cyc_2L01_2V, WriteAdr], (instregex "ST2Twov(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_4cyc_2L01_2V], (instregex "ST2Twov(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD store, 2 element, one lane, B/H/S
 // ASIMD store, 2 element, one lane, D
 def : InstRW<[N2Write_4cyc_1L01_1V],           (instregex "ST2i(8|16|32|64)$")>;
-def : InstRW<[N2Write_4cyc_1L01_1V, WriteAdr], (instregex "ST2i(8|16|32|64)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_4cyc_1L01_1V], (instregex "ST2i(8|16|32|64)_POST$")>;
 
 // ASIMD store, 3 element, multiple, D-form, B/H/S
 def : InstRW<[N2Write_5cyc_2L01_2V],           (instregex "ST3Threev(8b|4h|2s)$")>;
-def : InstRW<[N2Write_5cyc_2L01_2V, WriteAdr], (instregex "ST3Threev(8b|4h|2s)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_5cyc_2L01_2V], (instregex "ST3Threev(8b|4h|2s)_POST$")>;
 
 // ASIMD store, 3 element, multiple, Q-form, B/H/S
 // ASIMD store, 3 element, multiple, Q-form, D
 def : InstRW<[N2Write_6cyc_3L01_3V],           (instregex "ST3Threev(16b|8h|4s|2d)$")>;
-def : InstRW<[N2Write_6cyc_3L01_3V, WriteAdr], (instregex "ST3Threev(16b|8h|4s|2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_6cyc_3L01_3V], (instregex "ST3Threev(16b|8h|4s|2d)_POST$")>;
 
 // ASIMD store, 3 element, one lane, B/H
 // ASIMD store, 3 element, one lane, S
 // ASIMD store, 3 element, one lane, D
 def : InstRW<[N2Write_6cyc_3L01_3V],           (instregex "ST3i(8|16|32|64)$")>;
-def : InstRW<[N2Write_6cyc_3L01_3V, WriteAdr], (instregex "ST3i(8|16|32|64)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_6cyc_3L01_3V], (instregex "ST3i(8|16|32|64)_POST$")>;
 
 // ASIMD store, 4 element, multiple, D-form, B/H/S
 def : InstRW<[N2Write_6cyc_3L01_3V],           (instregex "ST4Fourv(8b|4h|2s)$")>;
-def : InstRW<[N2Write_6cyc_3L01_3V, WriteAdr], (instregex "ST4Fourv(8b|4h|2s)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_6cyc_3L01_3V], (instregex "ST4Fourv(8b|4h|2s)_POST$")>;
 
 // ASIMD store, 4 element, multiple, Q-form, B/H/S
 def : InstRW<[N2Write_7cyc_6L01_6V],           (instregex "ST4Fourv(16b|8h|4s)$")>;
-def : InstRW<[N2Write_7cyc_6L01_6V, WriteAdr], (instregex "ST4Fourv(16b|8h|4s)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_7cyc_6L01_6V], (instregex "ST4Fourv(16b|8h|4s)_POST$")>;
 
 // ASIMD store, 4 element, multiple, Q-form, D
 def : InstRW<[N2Write_5cyc_4L01_4V],           (instregex "ST4Fourv(2d)$")>;
-def : InstRW<[N2Write_5cyc_4L01_4V, WriteAdr], (instregex "ST4Fourv(2d)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_5cyc_4L01_4V], (instregex "ST4Fourv(2d)_POST$")>;
 
 // ASIMD store, 4 element, one lane, B/H/S
 def : InstRW<[N2Write_6cyc_3L01_3V],           (instregex "ST4i(8|16|32)$")>;
-def : InstRW<[N2Write_6cyc_3L01_3V, WriteAdr], (instregex "ST4i(8|16|32)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_6cyc_3L01_3V], (instregex "ST4i(8|16|32)_POST$")>;
 
 // ASIMD store, 4 element, one lane, D
 def : InstRW<[N2Write_4cyc_3L01_3V],            (instregex "ST4i(64)$")>;
-def : InstRW<[N2Write_4cyc_3L01_3V, WriteAdr],  (instregex "ST4i(64)_POST$")>;
+def : InstRW<[WriteAdr, N2Write_4cyc_3L01_3V],  (instregex "ST4i(64)_POST$")>;
 
 // Cryptography extensions
 // -----------------------------------------------------------------------------

diff  --git a/llvm/test/tools/llvm-mca/AArch64/Neoverse/N2-writeback.s b/llvm/test/tools/llvm-mca/AArch64/Neoverse/N2-writeback.s
index 7ee97a088ecba18..0c6ccc1face972f 100644
--- a/llvm/test/tools/llvm-mca/AArch64/Neoverse/N2-writeback.s
+++ b/llvm/test/tools/llvm-mca/AArch64/Neoverse/N2-writeback.s
@@ -1182,28 +1182,28 @@ add x0, x27, 1
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      508
 # CHECK-NEXT: Total uOps:        1500
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.50
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    2.95
+# CHECK-NEXT: IPC:               1.97
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     012
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   ld1	{ v1.1d }, [x27], #8
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D======eeeeeeER.    .    .    .  .   ld1	{ v1.2d }, [x27], #16
-# CHECK-NEXT: [0,3]     D============eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     D============eeeeeeER    .    .  .   ld1	{ v1.2s }, [x27], #8
-# CHECK-NEXT: [0,5]     D==================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=================eeeeeeER   .  .   ld1	{ v1.4h }, [x27], #8
-# CHECK-NEXT: [0,7]     .D=======================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .D=======================eeeeeeER.   ld1	{ v1.4s }, [x27], #16
-# CHECK-NEXT: [0,9]     .D=============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER . .   ld1	{ v1.1d }, [x27], #8
+# CHECK-NEXT: [0,1]     D=eE----R . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeER. .   ld1	{ v1.2d }, [x27], #16
+# CHECK-NEXT: [0,3]     D==eE----R. .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     D==eeeeeeER .   ld1	{ v1.2s }, [x27], #8
+# CHECK-NEXT: [0,5]     D===eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeER.   ld1	{ v1.4h }, [x27], #8
+# CHECK-NEXT: [0,7]     .D===eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .D===eeeeeeER   ld1	{ v1.4s }, [x27], #16
+# CHECK-NEXT: [0,9]     .D====eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1213,43 +1213,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.1d }, [x27], #8
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.2d }, [x27], #16
-# CHECK-NEXT: 3.     1     13.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     13.0   0.0    0.0       ld1	{ v1.2s }, [x27], #8
-# CHECK-NEXT: 5.     1     19.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     18.0   0.0    0.0       ld1	{ v1.4h }, [x27], #8
-# CHECK-NEXT: 7.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     24.0   0.0    0.0       ld1	{ v1.4s }, [x27], #16
-# CHECK-NEXT: 9.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     15.6   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1	{ v1.2d }, [x27], #16
+# CHECK-NEXT: 3.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    0.0    0.0       ld1	{ v1.2s }, [x27], #8
+# CHECK-NEXT: 5.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1	{ v1.4h }, [x27], #8
+# CHECK-NEXT: 7.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     4.0    0.0    0.0       ld1	{ v1.4s }, [x27], #16
+# CHECK-NEXT: 9.     1     5.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.1    0.1    2.0       <total>
 
 # CHECK:      [1] Code Region - G02
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      508
 # CHECK-NEXT: Total uOps:        1500
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.50
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    2.95
+# CHECK-NEXT: IPC:               1.97
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     012
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   ld1	{ v1.8b }, [x27], #8
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D======eeeeeeER.    .    .    .  .   ld1	{ v1.8h }, [x27], #16
-# CHECK-NEXT: [0,3]     D============eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     D============eeeeeeER    .    .  .   ld1	{ v1.16b }, [x27], #16
-# CHECK-NEXT: [0,5]     D==================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=================eeeeeeER   .  .   ld1	{ v1.1d }, [x27], x28
-# CHECK-NEXT: [0,7]     .D=======================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .D=======================eeeeeeER.   ld1	{ v1.2d }, [x27], x28
-# CHECK-NEXT: [0,9]     .D=============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER . .   ld1	{ v1.8b }, [x27], #8
+# CHECK-NEXT: [0,1]     D=eE----R . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeER. .   ld1	{ v1.8h }, [x27], #16
+# CHECK-NEXT: [0,3]     D==eE----R. .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     D==eeeeeeER .   ld1	{ v1.16b }, [x27], #16
+# CHECK-NEXT: [0,5]     D===eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeER.   ld1	{ v1.1d }, [x27], x28
+# CHECK-NEXT: [0,7]     .D===eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .D===eeeeeeER   ld1	{ v1.2d }, [x27], x28
+# CHECK-NEXT: [0,9]     .D====eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1259,43 +1259,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.8b }, [x27], #8
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.8h }, [x27], #16
-# CHECK-NEXT: 3.     1     13.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     13.0   0.0    0.0       ld1	{ v1.16b }, [x27], #16
-# CHECK-NEXT: 5.     1     19.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     18.0   0.0    0.0       ld1	{ v1.1d }, [x27], x28
-# CHECK-NEXT: 7.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     24.0   0.0    0.0       ld1	{ v1.2d }, [x27], x28
-# CHECK-NEXT: 9.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     15.6   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1	{ v1.8h }, [x27], #16
+# CHECK-NEXT: 3.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    0.0    0.0       ld1	{ v1.16b }, [x27], #16
+# CHECK-NEXT: 5.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1	{ v1.1d }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     4.0    0.0    0.0       ld1	{ v1.2d }, [x27], x28
+# CHECK-NEXT: 9.     1     5.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.1    0.1    2.0       <total>
 
 # CHECK:      [2] Code Region - G03
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      508
 # CHECK-NEXT: Total uOps:        1500
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.50
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    2.95
+# CHECK-NEXT: IPC:               1.97
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     012
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   ld1	{ v1.2s }, [x27], x28
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D======eeeeeeER.    .    .    .  .   ld1	{ v1.4h }, [x27], x28
-# CHECK-NEXT: [0,3]     D============eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     D============eeeeeeER    .    .  .   ld1	{ v1.4s }, [x27], x28
-# CHECK-NEXT: [0,5]     D==================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=================eeeeeeER   .  .   ld1	{ v1.8b }, [x27], x28
-# CHECK-NEXT: [0,7]     .D=======================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .D=======================eeeeeeER.   ld1	{ v1.8h }, [x27], x28
-# CHECK-NEXT: [0,9]     .D=============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER . .   ld1	{ v1.2s }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE----R . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeER. .   ld1	{ v1.4h }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eE----R. .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     D==eeeeeeER .   ld1	{ v1.4s }, [x27], x28
+# CHECK-NEXT: [0,5]     D===eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeER.   ld1	{ v1.8b }, [x27], x28
+# CHECK-NEXT: [0,7]     .D===eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .D===eeeeeeER   ld1	{ v1.8h }, [x27], x28
+# CHECK-NEXT: [0,9]     .D====eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1305,43 +1305,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.2s }, [x27], x28
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.4h }, [x27], x28
-# CHECK-NEXT: 3.     1     13.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     13.0   0.0    0.0       ld1	{ v1.4s }, [x27], x28
-# CHECK-NEXT: 5.     1     19.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     18.0   0.0    0.0       ld1	{ v1.8b }, [x27], x28
-# CHECK-NEXT: 7.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     24.0   0.0    0.0       ld1	{ v1.8h }, [x27], x28
-# CHECK-NEXT: 9.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     15.6   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1	{ v1.4h }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    0.0    0.0       ld1	{ v1.4s }, [x27], x28
+# CHECK-NEXT: 5.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1	{ v1.8b }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     4.0    0.0    0.0       ld1	{ v1.8h }, [x27], x28
+# CHECK-NEXT: 9.     1     5.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.1    0.1    2.0       <total>
 
 # CHECK:      [3] Code Region - G04
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      508
 # CHECK-NEXT: Total uOps:        1900
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.63
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    3.74
+# CHECK-NEXT: IPC:               1.97
 # CHECK-NEXT: Block RThroughput: 3.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     012
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   ld1	{ v1.16b }, [x27], x28
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D======eeeeeeER.    .    .    .  .   ld1	{ v1.1d, v2.1d }, [x27], #16
-# CHECK-NEXT: [0,3]     D============eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     D============eeeeeeER    .    .  .   ld1	{ v1.2d, v2.2d }, [x27], #32
-# CHECK-NEXT: [0,5]     .D=================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=================eeeeeeER   .  .   ld1	{ v1.2s, v2.2s }, [x27], #16
-# CHECK-NEXT: [0,7]     .D=======================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .D=======================eeeeeeER.   ld1	{ v1.4h, v2.4h }, [x27], #16
-# CHECK-NEXT: [0,9]     .D=============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER . .   ld1	{ v1.16b }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE----R . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeER. .   ld1	{ v1.1d, v2.1d }, [x27], #16
+# CHECK-NEXT: [0,3]     D==eE----R. .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     D==eeeeeeER .   ld1	{ v1.2d, v2.2d }, [x27], #32
+# CHECK-NEXT: [0,5]     .D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeER.   ld1	{ v1.2s, v2.2s }, [x27], #16
+# CHECK-NEXT: [0,7]     .D===eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .D===eeeeeeER   ld1	{ v1.4h, v2.4h }, [x27], #16
+# CHECK-NEXT: [0,9]     .D====eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1351,43 +1351,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.16b }, [x27], x28
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.1d, v2.1d }, [x27], #16
-# CHECK-NEXT: 3.     1     13.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     13.0   0.0    0.0       ld1	{ v1.2d, v2.2d }, [x27], #32
-# CHECK-NEXT: 5.     1     18.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     18.0   0.0    0.0       ld1	{ v1.2s, v2.2s }, [x27], #16
-# CHECK-NEXT: 7.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     24.0   0.0    0.0       ld1	{ v1.4h, v2.4h }, [x27], #16
-# CHECK-NEXT: 9.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     15.5   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1	{ v1.1d, v2.1d }, [x27], #16
+# CHECK-NEXT: 3.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    0.0    0.0       ld1	{ v1.2d, v2.2d }, [x27], #32
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1	{ v1.2s, v2.2s }, [x27], #16
+# CHECK-NEXT: 7.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     4.0    0.0    0.0       ld1	{ v1.4h, v2.4h }, [x27], #16
+# CHECK-NEXT: 9.     1     5.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.0    0.1    2.0       <total>
 
 # CHECK:      [4] Code Region - G05
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      508
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.67
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    3.94
+# CHECK-NEXT: IPC:               1.97
 # CHECK-NEXT: Block RThroughput: 3.3
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     012
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   ld1	{ v1.4s, v2.4s }, [x27], #32
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D======eeeeeeER.    .    .    .  .   ld1	{ v1.8b, v2.8b }, [x27], #16
-# CHECK-NEXT: [0,3]     D============eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===========eeeeeeER    .    .  .   ld1	{ v1.8h, v2.8h }, [x27], #32
-# CHECK-NEXT: [0,5]     .D=================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=================eeeeeeER   .  .   ld1	{ v1.16b, v2.16b }, [x27], #32
-# CHECK-NEXT: [0,7]     .D=======================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======================eeeeeeER.   ld1	{ v1.1d, v2.1d }, [x27], x28
-# CHECK-NEXT: [0,9]     . D============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER . .   ld1	{ v1.4s, v2.4s }, [x27], #32
+# CHECK-NEXT: [0,1]     D=eE----R . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeER. .   ld1	{ v1.8b, v2.8b }, [x27], #16
+# CHECK-NEXT: [0,3]     D==eE----R. .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeER .   ld1	{ v1.8h, v2.8h }, [x27], #32
+# CHECK-NEXT: [0,5]     .D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeER.   ld1	{ v1.16b, v2.16b }, [x27], #32
+# CHECK-NEXT: [0,7]     .D===eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeER   ld1	{ v1.1d, v2.1d }, [x27], x28
+# CHECK-NEXT: [0,9]     . D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1397,43 +1397,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.4s, v2.4s }, [x27], #32
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.8b, v2.8b }, [x27], #16
-# CHECK-NEXT: 3.     1     13.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     12.0   0.0    0.0       ld1	{ v1.8h, v2.8h }, [x27], #32
-# CHECK-NEXT: 5.     1     18.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     18.0   0.0    0.0       ld1	{ v1.16b, v2.16b }, [x27], #32
-# CHECK-NEXT: 7.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     23.0   0.0    0.0       ld1	{ v1.1d, v2.1d }, [x27], x28
-# CHECK-NEXT: 9.     1     29.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     15.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1	{ v1.8b, v2.8b }, [x27], #16
+# CHECK-NEXT: 3.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld1	{ v1.8h, v2.8h }, [x27], #32
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1	{ v1.16b, v2.16b }, [x27], #32
+# CHECK-NEXT: 7.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld1	{ v1.1d, v2.1d }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    2.0       <total>
 
 # CHECK:      [5] Code Region - G06
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      508
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.67
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    3.94
+# CHECK-NEXT: IPC:               1.97
 # CHECK-NEXT: Block RThroughput: 3.3
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     012
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   ld1	{ v1.2d, v2.2d }, [x27], x28
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D======eeeeeeER.    .    .    .  .   ld1	{ v1.2s, v2.2s }, [x27], x28
-# CHECK-NEXT: [0,3]     D============eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===========eeeeeeER    .    .  .   ld1	{ v1.4h, v2.4h }, [x27], x28
-# CHECK-NEXT: [0,5]     .D=================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=================eeeeeeER   .  .   ld1	{ v1.4s, v2.4s }, [x27], x28
-# CHECK-NEXT: [0,7]     .D=======================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======================eeeeeeER.   ld1	{ v1.8b, v2.8b }, [x27], x28
-# CHECK-NEXT: [0,9]     . D============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER . .   ld1	{ v1.2d, v2.2d }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE----R . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeER. .   ld1	{ v1.2s, v2.2s }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eE----R. .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeER .   ld1	{ v1.4h, v2.4h }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeER.   ld1	{ v1.4s, v2.4s }, [x27], x28
+# CHECK-NEXT: [0,7]     .D===eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeER   ld1	{ v1.8b, v2.8b }, [x27], x28
+# CHECK-NEXT: [0,9]     . D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1443,43 +1443,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.2d, v2.2d }, [x27], x28
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.2s, v2.2s }, [x27], x28
-# CHECK-NEXT: 3.     1     13.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     12.0   0.0    0.0       ld1	{ v1.4h, v2.4h }, [x27], x28
-# CHECK-NEXT: 5.     1     18.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     18.0   0.0    0.0       ld1	{ v1.4s, v2.4s }, [x27], x28
-# CHECK-NEXT: 7.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     23.0   0.0    0.0       ld1	{ v1.8b, v2.8b }, [x27], x28
-# CHECK-NEXT: 9.     1     29.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     15.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1	{ v1.2s, v2.2s }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld1	{ v1.4h, v2.4h }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1	{ v1.4s, v2.4s }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld1	{ v1.8b, v2.8b }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    2.0       <total>
 
 # CHECK:      [6] Code Region - G07
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      508
 # CHECK-NEXT: Total uOps:        2300
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.77
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    4.53
+# CHECK-NEXT: IPC:               1.97
 # CHECK-NEXT: Block RThroughput: 4.3
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     012
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   ld1	{ v1.8h, v2.8h }, [x27], x28
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D======eeeeeeER.    .    .    .  .   ld1	{ v1.16b, v2.16b }, [x27], x28
-# CHECK-NEXT: [0,3]     D============eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===========eeeeeeER    .    .  .   ld1	{ v1.1d, v2.1d, v3.1d }, [x27], #24
-# CHECK-NEXT: [0,5]     .D=================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=================eeeeeeER   .  .   ld1	{ v1.2d, v2.2d, v3.2d }, [x27], #48
-# CHECK-NEXT: [0,7]     .D=======================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======================eeeeeeER.   ld1	{ v1.2s, v2.2s, v3.2s }, [x27], #24
-# CHECK-NEXT: [0,9]     . D============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER . .   ld1	{ v1.8h, v2.8h }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE----R . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeER. .   ld1	{ v1.16b, v2.16b }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eE----R. .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeER .   ld1	{ v1.1d, v2.1d, v3.1d }, [x27], #24
+# CHECK-NEXT: [0,5]     .D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeER.   ld1	{ v1.2d, v2.2d, v3.2d }, [x27], #48
+# CHECK-NEXT: [0,7]     .D===eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeER   ld1	{ v1.2s, v2.2s, v3.2s }, [x27], #24
+# CHECK-NEXT: [0,9]     . D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1489,43 +1489,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.8h, v2.8h }, [x27], x28
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.16b, v2.16b }, [x27], x28
-# CHECK-NEXT: 3.     1     13.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     12.0   0.0    0.0       ld1	{ v1.1d, v2.1d, v3.1d }, [x27], #24
-# CHECK-NEXT: 5.     1     18.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     18.0   0.0    0.0       ld1	{ v1.2d, v2.2d, v3.2d }, [x27], #48
-# CHECK-NEXT: 7.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     23.0   0.0    0.0       ld1	{ v1.2s, v2.2s, v3.2s }, [x27], #24
-# CHECK-NEXT: 9.     1     29.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     15.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1	{ v1.16b, v2.16b }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld1	{ v1.1d, v2.1d, v3.1d }, [x27], #24
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1	{ v1.2d, v2.2d, v3.2d }, [x27], #48
+# CHECK-NEXT: 7.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld1	{ v1.2s, v2.2s, v3.2s }, [x27], #24
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    2.0       <total>
 
 # CHECK:      [7] Code Region - G08
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      508
 # CHECK-NEXT: Total uOps:        2500
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.83
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    4.92
+# CHECK-NEXT: IPC:               1.97
 # CHECK-NEXT: Block RThroughput: 5.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     012
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   ld1	{ v1.4h, v2.4h, v3.4h }, [x27], #24
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D======eeeeeeER.    .    .    .  .   ld1	{ v1.4s, v2.4s, v3.4s }, [x27], #48
-# CHECK-NEXT: [0,3]     D============eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===========eeeeeeER    .    .  .   ld1	{ v1.8b, v2.8b, v3.8b }, [x27], #24
-# CHECK-NEXT: [0,5]     .D=================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=================eeeeeeER   .  .   ld1	{ v1.8h, v2.8h, v3.8h }, [x27], #48
-# CHECK-NEXT: [0,7]     .D=======================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======================eeeeeeER.   ld1	{ v1.16b, v2.16b, v3.16b }, [x27], #48
-# CHECK-NEXT: [0,9]     . D============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER . .   ld1	{ v1.4h, v2.4h, v3.4h }, [x27], #24
+# CHECK-NEXT: [0,1]     D=eE----R . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeER. .   ld1	{ v1.4s, v2.4s, v3.4s }, [x27], #48
+# CHECK-NEXT: [0,3]     D==eE----R. .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeER .   ld1	{ v1.8b, v2.8b, v3.8b }, [x27], #24
+# CHECK-NEXT: [0,5]     .D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeER.   ld1	{ v1.8h, v2.8h, v3.8h }, [x27], #48
+# CHECK-NEXT: [0,7]     .D===eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeER   ld1	{ v1.16b, v2.16b, v3.16b }, [x27], #48
+# CHECK-NEXT: [0,9]     . D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1535,43 +1535,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.4h, v2.4h, v3.4h }, [x27], #24
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.4s, v2.4s, v3.4s }, [x27], #48
-# CHECK-NEXT: 3.     1     13.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     12.0   0.0    0.0       ld1	{ v1.8b, v2.8b, v3.8b }, [x27], #24
-# CHECK-NEXT: 5.     1     18.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     18.0   0.0    0.0       ld1	{ v1.8h, v2.8h, v3.8h }, [x27], #48
-# CHECK-NEXT: 7.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     23.0   0.0    0.0       ld1	{ v1.16b, v2.16b, v3.16b }, [x27], #48
-# CHECK-NEXT: 9.     1     29.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     15.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1	{ v1.4s, v2.4s, v3.4s }, [x27], #48
+# CHECK-NEXT: 3.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld1	{ v1.8b, v2.8b, v3.8b }, [x27], #24
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1	{ v1.8h, v2.8h, v3.8h }, [x27], #48
+# CHECK-NEXT: 7.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld1	{ v1.16b, v2.16b, v3.16b }, [x27], #48
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    2.0       <total>
 
 # CHECK:      [8] Code Region - G09
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      508
 # CHECK-NEXT: Total uOps:        2500
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.83
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    4.92
+# CHECK-NEXT: IPC:               1.97
 # CHECK-NEXT: Block RThroughput: 5.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     012
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   ld1	{ v1.1d, v2.1d, v3.1d }, [x27], x28
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D======eeeeeeER.    .    .    .  .   ld1	{ v1.2d, v2.2d, v3.2d }, [x27], x28
-# CHECK-NEXT: [0,3]     D============eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===========eeeeeeER    .    .  .   ld1	{ v1.2s, v2.2s, v3.2s }, [x27], x28
-# CHECK-NEXT: [0,5]     .D=================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=================eeeeeeER   .  .   ld1	{ v1.4h, v2.4h, v3.4h }, [x27], x28
-# CHECK-NEXT: [0,7]     .D=======================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======================eeeeeeER.   ld1	{ v1.4s, v2.4s, v3.4s }, [x27], x28
-# CHECK-NEXT: [0,9]     . D============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER . .   ld1	{ v1.1d, v2.1d, v3.1d }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE----R . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeER. .   ld1	{ v1.2d, v2.2d, v3.2d }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eE----R. .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeER .   ld1	{ v1.2s, v2.2s, v3.2s }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeER.   ld1	{ v1.4h, v2.4h, v3.4h }, [x27], x28
+# CHECK-NEXT: [0,7]     .D===eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeER   ld1	{ v1.4s, v2.4s, v3.4s }, [x27], x28
+# CHECK-NEXT: [0,9]     . D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1581,43 +1581,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.1d, v2.1d, v3.1d }, [x27], x28
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.2d, v2.2d, v3.2d }, [x27], x28
-# CHECK-NEXT: 3.     1     13.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     12.0   0.0    0.0       ld1	{ v1.2s, v2.2s, v3.2s }, [x27], x28
-# CHECK-NEXT: 5.     1     18.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     18.0   0.0    0.0       ld1	{ v1.4h, v2.4h, v3.4h }, [x27], x28
-# CHECK-NEXT: 7.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     23.0   0.0    0.0       ld1	{ v1.4s, v2.4s, v3.4s }, [x27], x28
-# CHECK-NEXT: 9.     1     29.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     15.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1	{ v1.2d, v2.2d, v3.2d }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld1	{ v1.2s, v2.2s, v3.2s }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1	{ v1.4h, v2.4h, v3.4h }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld1	{ v1.4s, v2.4s, v3.4s }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    2.0       <total>
 
 # CHECK:      [9] Code Region - G10
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3204
+# CHECK-NEXT: Total Cycles:      608
 # CHECK-NEXT: Total uOps:        2700
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.84
-# CHECK-NEXT: IPC:               0.31
+# CHECK-NEXT: uOps Per Cycle:    4.44
+# CHECK-NEXT: IPC:               1.64
 # CHECK-NEXT: Block RThroughput: 5.7
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          012345
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     0123
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .    .   ld1	{ v1.8b, v2.8b, v3.8b }, [x27], x28
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D======eeeeeeER.    .    .    .    .   ld1	{ v1.8h, v2.8h, v3.8h }, [x27], x28
-# CHECK-NEXT: [0,3]     D============eER    .    .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===========eeeeeeER    .    .    .   ld1	{ v1.16b, v2.16b, v3.16b }, [x27], x28
-# CHECK-NEXT: [0,5]     .D=================eER   .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=================eeeeeeeER  .    .   ld1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
-# CHECK-NEXT: [0,7]     . D=======================eER .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D=======================eeeeeeeER.   ld1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
-# CHECK-NEXT: [0,9]     . D==============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER .  .   ld1	{ v1.8b, v2.8b, v3.8b }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE----R .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeER.  .   ld1	{ v1.8h, v2.8h, v3.8h }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eE----R.  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeER  .   ld1	{ v1.16b, v2.16b, v3.16b }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eE----R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeeER.   ld1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
+# CHECK-NEXT: [0,7]     . D==eE-----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeeER   ld1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
+# CHECK-NEXT: [0,9]     . D===eE-----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1627,43 +1627,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.8b, v2.8b, v3.8b }, [x27], x28
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.8h, v2.8h, v3.8h }, [x27], x28
-# CHECK-NEXT: 3.     1     13.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     12.0   0.0    0.0       ld1	{ v1.16b, v2.16b, v3.16b }, [x27], x28
-# CHECK-NEXT: 5.     1     18.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     18.0   0.0    0.0       ld1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
-# CHECK-NEXT: 7.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     24.0   0.0    0.0       ld1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
-# CHECK-NEXT: 9.     1     31.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     15.5   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1	{ v1.8h, v2.8h, v3.8h }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld1	{ v1.16b, v2.16b, v3.16b }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
+# CHECK-NEXT: 7.     1     3.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
+# CHECK-NEXT: 9.     1     4.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.6    0.1    2.2       <total>
 
 # CHECK:      [10] Code Region - G11
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3504
+# CHECK-NEXT: Total Cycles:      675
 # CHECK-NEXT: Total uOps:        3000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.86
-# CHECK-NEXT: IPC:               0.29
+# CHECK-NEXT: uOps Per Cycle:    4.44
+# CHECK-NEXT: IPC:               1.48
 # CHECK-NEXT: Block RThroughput: 6.7
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          012345678
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeeER.    .    .    .    .    .  .   ld1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
-# CHECK-NEXT: [0,1]     D=======eER    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D======eeeeeeeER   .    .    .    .  .   ld1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
-# CHECK-NEXT: [0,3]     .D=============eER  .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D============eeeeeeeER .    .    .  .   ld1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
-# CHECK-NEXT: [0,5]     . D===================eER.    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D==================eeeeeeeER    .  .   ld1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
-# CHECK-NEXT: [0,7]     .  D=========================eER   .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D========================eeeeeeeER.   ld1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
-# CHECK-NEXT: [0,9]     .   D===============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeeER.   .   ld1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
+# CHECK-NEXT: [0,1]     D=eE-----R.   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeER   .   ld1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
+# CHECK-NEXT: [0,3]     .D=eE-----R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . DeeeeeeeER  .   ld1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
+# CHECK-NEXT: [0,5]     . D=eE-----R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeeER.   ld1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
+# CHECK-NEXT: [0,7]     .  D==eE-----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D=eeeeeeeER   ld1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
+# CHECK-NEXT: [0,9]     .   D==eE-----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1673,43 +1673,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
-# CHECK-NEXT: 1.     1     8.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
-# CHECK-NEXT: 3.     1     14.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     13.0   0.0    0.0       ld1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
-# CHECK-NEXT: 5.     1     20.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     19.0   0.0    0.0       ld1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
-# CHECK-NEXT: 7.     1     26.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     25.0   0.0    0.0       ld1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
-# CHECK-NEXT: 9.     1     32.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     16.5   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
+# CHECK-NEXT: 3.     1     2.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     1.0    0.0    0.0       ld1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
+# CHECK-NEXT: 5.     1     2.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    1.0    0.0       ld1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
+# CHECK-NEXT: 7.     1     3.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     2.0    0.0    0.0       ld1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
+# CHECK-NEXT: 9.     1     3.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT:        1     1.9    0.2    2.5       <total>
 
 # CHECK:      [11] Code Region - G12
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3504
+# CHECK-NEXT: Total Cycles:      675
 # CHECK-NEXT: Total uOps:        3000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.86
-# CHECK-NEXT: IPC:               0.29
+# CHECK-NEXT: uOps Per Cycle:    4.44
+# CHECK-NEXT: IPC:               1.48
 # CHECK-NEXT: Block RThroughput: 6.7
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          012345678
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeeER.    .    .    .    .    .  .   ld1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
-# CHECK-NEXT: [0,1]     D=======eER    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D======eeeeeeeER   .    .    .    .  .   ld1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], x28
-# CHECK-NEXT: [0,3]     .D=============eER  .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D============eeeeeeeER .    .    .  .   ld1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
-# CHECK-NEXT: [0,5]     . D===================eER.    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D==================eeeeeeeER    .  .   ld1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D=========================eER   .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D========================eeeeeeeER.   ld1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D===============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeeER.   .   ld1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
+# CHECK-NEXT: [0,1]     D=eE-----R.   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeER   .   ld1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE-----R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . DeeeeeeeER  .   ld1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
+# CHECK-NEXT: [0,5]     . D=eE-----R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeeER.   ld1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D==eE-----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D=eeeeeeeER   ld1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D==eE-----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1719,43 +1719,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
-# CHECK-NEXT: 1.     1     8.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], x28
-# CHECK-NEXT: 3.     1     14.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     13.0   0.0    0.0       ld1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
-# CHECK-NEXT: 5.     1     20.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     19.0   0.0    0.0       ld1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
-# CHECK-NEXT: 7.     1     26.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     25.0   0.0    0.0       ld1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
-# CHECK-NEXT: 9.     1     32.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     16.5   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     1.0    0.0    0.0       ld1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
+# CHECK-NEXT: 5.     1     2.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    1.0    0.0       ld1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     2.0    0.0    0.0       ld1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
+# CHECK-NEXT: 9.     1     3.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT:        1     1.9    0.2    2.5       <total>
 
 # CHECK:      [12] Code Region - G13
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3604
+# CHECK-NEXT: Total Cycles:      1210
 # CHECK-NEXT: Total uOps:        2800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.78
-# CHECK-NEXT: IPC:               0.28
+# CHECK-NEXT: uOps Per Cycle:    2.31
+# CHECK-NEXT: IPC:               0.83
 # CHECK-NEXT: Block RThroughput: 5.7
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     0123456789
+# CHECK-NEXT: Index     0123456789          01
 
-# CHECK:      [0,0]     DeeeeeeeER.    .    .    .    .    .   .   ld1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
-# CHECK-NEXT: [0,1]     D=======eER    .    .    .    .    .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D======eeeeeeeER   .    .    .    .   .   ld1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
-# CHECK-NEXT: [0,3]     .D=============eER  .    .    .    .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D============eeeeeeeER .    .    .   .   ld1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
-# CHECK-NEXT: [0,5]     . D===================eER.    .    .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D==================eeeeeeeER    .   .   ld1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D=========================eER   .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .  D=========================eeeeeeeeER.   ld1	{ v1.b }[0], [x27], #1
-# CHECK-NEXT: [0,9]     .  D=================================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeeER.    .    ..   ld1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE-----R.    .    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeER    .    ..   ld1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE-----R    .    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . DeeeeeeeER   .    ..   ld1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
+# CHECK-NEXT: [0,5]     . D=eE-----R   .    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeeER .    ..   ld1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D==eE-----R .    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .  D========eeeeeeeeER   ld1	{ v1.b }[0], [x27], #1
+# CHECK-NEXT: [0,9]     .  D=========eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1765,22 +1765,22 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
-# CHECK-NEXT: 1.     1     8.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ld1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
-# CHECK-NEXT: 3.     1     14.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     13.0   0.0    0.0       ld1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
-# CHECK-NEXT: 5.     1     20.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     19.0   0.0    0.0       ld1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
-# CHECK-NEXT: 7.     1     26.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     26.0   0.0    0.0       ld1	{ v1.b }[0], [x27], #1
-# CHECK-NEXT: 9.     1     34.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     16.8   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     1.0    0.0    0.0       ld1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
+# CHECK-NEXT: 5.     1     2.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    1.0    0.0       ld1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     9.0    0.0    0.0       ld1	{ v1.b }[0], [x27], #1
+# CHECK-NEXT: 9.     1     10.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.3    0.2    2.6       <total>
 
 # CHECK:      [13] Code Region - G14
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      4003
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
@@ -1790,18 +1790,18 @@ add x0, x27, 1
 
 # CHECK:      Timeline view:
 # CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld1	{ v1.b }[8], [x27], #1
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld1	{ v1.b }[0], [x27], x28
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld1	{ v1.b }[8], [x27], x28
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld1	{ v1.h }[0], [x27], #2
-# CHECK-NEXT: [0,7]     .D===============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld1	{ v1.h }[4], [x27], #2
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT: Index     0123456789          0123456789          012
+
+# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    . .   ld1	{ v1.b }[8], [x27], #1
+# CHECK-NEXT: [0,1]     D=eE------R    .    .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    . .   ld1	{ v1.b }[0], [x27], x28
+# CHECK-NEXT: [0,3]     D=========eE------R .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    . .   ld1	{ v1.b }[8], [x27], x28
+# CHECK-NEXT: [0,5]     .D================eE------R   .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    . .   ld1	{ v1.h }[0], [x27], #2
+# CHECK-NEXT: [0,7]     .D========================eE------R.    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER   ld1	{ v1.h }[4], [x27], #2
+# CHECK-NEXT: [0,9]     . D===============================eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1811,22 +1811,22 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.b }[8], [x27], #1
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld1	{ v1.b }[0], [x27], x28
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 3.     1     10.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld1	{ v1.b }[8], [x27], x28
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 5.     1     17.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld1	{ v1.h }[0], [x27], #2
-# CHECK-NEXT: 7.     1     32.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 7.     1     25.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld1	{ v1.h }[4], [x27], #2
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.2   0.1    0.0       <total>
+# CHECK-NEXT: 9.     1     32.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     16.7   0.1    3.0       <total>
 
 # CHECK:      [14] Code Region - G15
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      4003
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
@@ -1836,18 +1836,18 @@ add x0, x27, 1
 
 # CHECK:      Timeline view:
 # CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld1	{ v1.h }[0], [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld1	{ v1.h }[4], [x27], x28
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld1	{ v1.s }[0], [x27], #4
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld1	{ v1.s }[0], [x27], x28
-# CHECK-NEXT: [0,7]     .D===============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld1	{ v1.d }[0], [x27], #8
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT: Index     0123456789          0123456789          012
+
+# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    . .   ld1	{ v1.h }[0], [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R    .    .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    . .   ld1	{ v1.h }[4], [x27], x28
+# CHECK-NEXT: [0,3]     D=========eE------R .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    . .   ld1	{ v1.s }[0], [x27], #4
+# CHECK-NEXT: [0,5]     .D================eE------R   .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    . .   ld1	{ v1.s }[0], [x27], x28
+# CHECK-NEXT: [0,7]     .D========================eE------R.    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER   ld1	{ v1.d }[0], [x27], #8
+# CHECK-NEXT: [0,9]     . D===============================eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1857,43 +1857,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.h }[0], [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld1	{ v1.h }[4], [x27], x28
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 3.     1     10.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld1	{ v1.s }[0], [x27], #4
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 5.     1     17.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld1	{ v1.s }[0], [x27], x28
-# CHECK-NEXT: 7.     1     32.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 7.     1     25.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld1	{ v1.d }[0], [x27], #8
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.2   0.1    0.0       <total>
+# CHECK-NEXT: 9.     1     32.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     16.7   0.1    3.0       <total>
 
 # CHECK:      [15] Code Region - G16
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      1203
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.50
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    1.66
+# CHECK-NEXT: IPC:               0.83
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld1	{ v1.d }[0], [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld1r	{ v1.1d }, [x27], #8
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld1r	{ v1.2d }, [x27], #8
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld1r	{ v1.2s }, [x27], #4
-# CHECK-NEXT: [0,7]     .D===============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld1r	{ v1.4h }, [x27], #2
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER   .   ld1	{ v1.d }[0], [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeeeER  .   ld1r	{ v1.1d }, [x27], #8
+# CHECK-NEXT: [0,3]     D==eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeeeER .   ld1r	{ v1.2d }, [x27], #8
+# CHECK-NEXT: [0,5]     .D==eE------R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeeeER.   ld1r	{ v1.2s }, [x27], #4
+# CHECK-NEXT: [0,7]     .D===eE------R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeeeER   ld1r	{ v1.4h }, [x27], #2
+# CHECK-NEXT: [0,9]     . D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1903,43 +1903,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1	{ v1.d }[0], [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld1r	{ v1.1d }, [x27], #8
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld1r	{ v1.2d }, [x27], #8
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld1r	{ v1.2s }, [x27], #4
-# CHECK-NEXT: 7.     1     32.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld1r	{ v1.4h }, [x27], #2
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1r	{ v1.1d }, [x27], #8
+# CHECK-NEXT: 3.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld1r	{ v1.2d }, [x27], #8
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1r	{ v1.2s }, [x27], #4
+# CHECK-NEXT: 7.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld1r	{ v1.4h }, [x27], #2
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    3.0       <total>
 
 # CHECK:      [16] Code Region - G17
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      510
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.50
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    3.92
+# CHECK-NEXT: IPC:               1.96
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld1r	{ v1.4s }, [x27], #4
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld1r	{ v1.8b }, [x27], #1
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld1r	{ v1.8h }, [x27], #2
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld1r	{ v1.16b }, [x27], #1
-# CHECK-NEXT: [0,7]     .D===============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld1r	{ v1.1d }, [x27], x28
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER   .   ld1r	{ v1.4s }, [x27], #4
+# CHECK-NEXT: [0,1]     D=eE------R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeeeER  .   ld1r	{ v1.8b }, [x27], #1
+# CHECK-NEXT: [0,3]     D==eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeeeER .   ld1r	{ v1.8h }, [x27], #2
+# CHECK-NEXT: [0,5]     .D==eE------R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeeeER.   ld1r	{ v1.16b }, [x27], #1
+# CHECK-NEXT: [0,7]     .D===eE------R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeeeER   ld1r	{ v1.1d }, [x27], x28
+# CHECK-NEXT: [0,9]     . D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1949,43 +1949,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1r	{ v1.4s }, [x27], #4
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld1r	{ v1.8b }, [x27], #1
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld1r	{ v1.8h }, [x27], #2
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld1r	{ v1.16b }, [x27], #1
-# CHECK-NEXT: 7.     1     32.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld1r	{ v1.1d }, [x27], x28
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1r	{ v1.8b }, [x27], #1
+# CHECK-NEXT: 3.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld1r	{ v1.8h }, [x27], #2
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1r	{ v1.16b }, [x27], #1
+# CHECK-NEXT: 7.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld1r	{ v1.1d }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    3.0       <total>
 
 # CHECK:      [17] Code Region - G18
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      510
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.50
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    3.92
+# CHECK-NEXT: IPC:               1.96
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld1r	{ v1.2d }, [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld1r	{ v1.2s }, [x27], x28
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld1r	{ v1.4h }, [x27], x28
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld1r	{ v1.4s }, [x27], x28
-# CHECK-NEXT: [0,7]     .D===============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld1r	{ v1.8b }, [x27], x28
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER   .   ld1r	{ v1.2d }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeeeER  .   ld1r	{ v1.2s }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeeeER .   ld1r	{ v1.4h }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eE------R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeeeER.   ld1r	{ v1.4s }, [x27], x28
+# CHECK-NEXT: [0,7]     .D===eE------R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeeeER   ld1r	{ v1.8b }, [x27], x28
+# CHECK-NEXT: [0,9]     . D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -1995,43 +1995,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1r	{ v1.2d }, [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld1r	{ v1.2s }, [x27], x28
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld1r	{ v1.4h }, [x27], x28
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld1r	{ v1.4s }, [x27], x28
-# CHECK-NEXT: 7.     1     32.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld1r	{ v1.8b }, [x27], x28
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1r	{ v1.2s }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld1r	{ v1.4h }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld1r	{ v1.4s }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld1r	{ v1.8b }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    3.0       <total>
 
 # CHECK:      [18] Code Region - G19
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      510
 # CHECK-NEXT: Total uOps:        2400
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.60
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    4.71
+# CHECK-NEXT: IPC:               1.96
 # CHECK-NEXT: Block RThroughput: 4.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld1r	{ v1.8h }, [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld1r	{ v1.16b }, [x27], x28
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld2	{ v1.2d, v2.2d }, [x27], #32
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld2	{ v1.2s, v2.2s }, [x27], #16
-# CHECK-NEXT: [0,7]     . D==============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld2	{ v1.4h, v2.4h }, [x27], #16
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER   .   ld1r	{ v1.8h }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeeeER  .   ld1r	{ v1.16b }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeeeER .   ld2	{ v1.2d, v2.2d }, [x27], #32
+# CHECK-NEXT: [0,5]     .D==eE------R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeeeER.   ld2	{ v1.2s, v2.2s }, [x27], #16
+# CHECK-NEXT: [0,7]     . D==eE------R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeeeER   ld2	{ v1.4h, v2.4h }, [x27], #16
+# CHECK-NEXT: [0,9]     . D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2041,43 +2041,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld1r	{ v1.8h }, [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld1r	{ v1.16b }, [x27], x28
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld2	{ v1.2d, v2.2d }, [x27], #32
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld2	{ v1.2s, v2.2s }, [x27], #16
-# CHECK-NEXT: 7.     1     31.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld2	{ v1.4h, v2.4h }, [x27], #16
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.1   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld1r	{ v1.16b }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld2	{ v1.2d, v2.2d }, [x27], #32
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld2	{ v1.2s, v2.2s }, [x27], #16
+# CHECK-NEXT: 7.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld2	{ v1.4h, v2.4h }, [x27], #16
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.6    0.1    3.0       <total>
 
 # CHECK:      [19] Code Region - G20
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      510
 # CHECK-NEXT: Total uOps:        2900
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.72
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    5.69
+# CHECK-NEXT: IPC:               1.96
 # CHECK-NEXT: Block RThroughput: 5.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld2	{ v1.4s, v2.4s }, [x27], #32
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld2	{ v1.8b, v2.8b }, [x27], #16
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld2	{ v1.8h, v2.8h }, [x27], #32
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     . D======================eeeeeeeeER.    .  .   ld2	{ v1.16b, v2.16b }, [x27], #32
-# CHECK-NEXT: [0,7]     . D==============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .  D=============================eeeeeeeeER.   ld2	{ v1.2d, v2.2d }, [x27], x28
-# CHECK-NEXT: [0,9]     .  D=====================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER   .   ld2	{ v1.4s, v2.4s }, [x27], #32
+# CHECK-NEXT: [0,1]     D=eE------R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeeeER  .   ld2	{ v1.8b, v2.8b }, [x27], #16
+# CHECK-NEXT: [0,3]     .D=eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeeeER .   ld2	{ v1.8h, v2.8h }, [x27], #32
+# CHECK-NEXT: [0,5]     .D==eE------R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     . D=eeeeeeeeER.   ld2	{ v1.16b, v2.16b }, [x27], #32
+# CHECK-NEXT: [0,7]     . D==eE------R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .  D=eeeeeeeeER   ld2	{ v1.2d, v2.2d }, [x27], x28
+# CHECK-NEXT: [0,9]     .  D==eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2087,43 +2087,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld2	{ v1.4s, v2.4s }, [x27], #32
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld2	{ v1.8b, v2.8b }, [x27], #16
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld2	{ v1.8h, v2.8h }, [x27], #32
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     23.0   0.0    0.0       ld2	{ v1.16b, v2.16b }, [x27], #32
-# CHECK-NEXT: 7.     1     31.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     30.0   0.0    0.0       ld2	{ v1.2d, v2.2d }, [x27], x28
-# CHECK-NEXT: 9.     1     38.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.7   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld2	{ v1.8b, v2.8b }, [x27], #16
+# CHECK-NEXT: 3.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld2	{ v1.8h, v2.8h }, [x27], #32
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       ld2	{ v1.16b, v2.16b }, [x27], #32
+# CHECK-NEXT: 7.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     2.0    0.0    0.0       ld2	{ v1.2d, v2.2d }, [x27], x28
+# CHECK-NEXT: 9.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.2    0.1    3.0       <total>
 
 # CHECK:      [20] Code Region - G21
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      510
 # CHECK-NEXT: Total uOps:        2700
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.67
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    5.29
+# CHECK-NEXT: IPC:               1.96
 # CHECK-NEXT: Block RThroughput: 5.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld2	{ v1.2s, v2.2s }, [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld2	{ v1.4h, v2.4h }, [x27], x28
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld2	{ v1.4s, v2.4s }, [x27], x28
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld2	{ v1.8b, v2.8b }, [x27], x28
-# CHECK-NEXT: [0,7]     . D==============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld2	{ v1.8h, v2.8h }, [x27], x28
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER   .   ld2	{ v1.2s, v2.2s }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeeeER  .   ld2	{ v1.4h, v2.4h }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeeeER .   ld2	{ v1.4s, v2.4s }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eE------R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeeeER.   ld2	{ v1.8b, v2.8b }, [x27], x28
+# CHECK-NEXT: [0,7]     . D==eE------R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeeeER   ld2	{ v1.8h, v2.8h }, [x27], x28
+# CHECK-NEXT: [0,9]     . D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2133,43 +2133,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld2	{ v1.2s, v2.2s }, [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld2	{ v1.4h, v2.4h }, [x27], x28
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld2	{ v1.4s, v2.4s }, [x27], x28
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld2	{ v1.8b, v2.8b }, [x27], x28
-# CHECK-NEXT: 7.     1     31.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld2	{ v1.8h, v2.8h }, [x27], x28
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.1   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld2	{ v1.4h, v2.4h }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld2	{ v1.4s, v2.4s }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld2	{ v1.8b, v2.8b }, [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld2	{ v1.8h, v2.8h }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.6    0.1    3.0       <total>
 
 # CHECK:      [21] Code Region - G22
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      3310
 # CHECK-NEXT: Total uOps:        2600
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.65
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    0.79
+# CHECK-NEXT: IPC:               0.30
 # CHECK-NEXT: Block RThroughput: 5.0
 
 # CHECK:      Timeline view:
 # CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld2	{ v1.16b, v2.16b }, [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld2	{ v1.b, v2.b }[0], [x27], #2
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld2	{ v1.b, v2.b }[8], [x27], #2
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld2	{ v1.b, v2.b }[0], [x27], x28
-# CHECK-NEXT: [0,7]     . D==============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld2	{ v1.b, v2.b }[8], [x27], x28
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT: Index     0123456789          0123456789          012
+
+# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    . .   ld2	{ v1.16b, v2.16b }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R    .    .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    . .   ld2	{ v1.b, v2.b }[0], [x27], #2
+# CHECK-NEXT: [0,3]     .D========eE------R .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    . .   ld2	{ v1.b, v2.b }[8], [x27], #2
+# CHECK-NEXT: [0,5]     .D================eE------R   .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    . .   ld2	{ v1.b, v2.b }[0], [x27], x28
+# CHECK-NEXT: [0,7]     . D=======================eE------R.    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER   ld2	{ v1.b, v2.b }[8], [x27], x28
+# CHECK-NEXT: [0,9]     . D===============================eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2179,22 +2179,22 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld2	{ v1.16b, v2.16b }, [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld2	{ v1.b, v2.b }[0], [x27], #2
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 3.     1     9.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld2	{ v1.b, v2.b }[8], [x27], #2
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 5.     1     17.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld2	{ v1.b, v2.b }[0], [x27], x28
-# CHECK-NEXT: 7.     1     31.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 7.     1     24.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld2	{ v1.b, v2.b }[8], [x27], x28
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.0   0.1    0.0       <total>
+# CHECK-NEXT: 9.     1     32.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     16.5   0.1    3.0       <total>
 
 # CHECK:      [22] Code Region - G23
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      4003
 # CHECK-NEXT: Total uOps:        2500
 
 # CHECK:      Dispatch Width:    10
@@ -2204,18 +2204,18 @@ add x0, x27, 1
 
 # CHECK:      Timeline view:
 # CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld2	{ v1.h, v2.h }[0], [x27], #4
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld2	{ v1.h, v2.h }[4], [x27], #4
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld2	{ v1.h, v2.h }[0], [x27], x28
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld2	{ v1.h, v2.h }[4], [x27], x28
-# CHECK-NEXT: [0,7]     .D===============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld2	{ v1.s, v2.s }[0], [x27], #8
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT: Index     0123456789          0123456789          012
+
+# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    . .   ld2	{ v1.h, v2.h }[0], [x27], #4
+# CHECK-NEXT: [0,1]     D=eE------R    .    .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    . .   ld2	{ v1.h, v2.h }[4], [x27], #4
+# CHECK-NEXT: [0,3]     D=========eE------R .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    . .   ld2	{ v1.h, v2.h }[0], [x27], x28
+# CHECK-NEXT: [0,5]     .D================eE------R   .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    . .   ld2	{ v1.h, v2.h }[4], [x27], x28
+# CHECK-NEXT: [0,7]     .D========================eE------R.    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER   ld2	{ v1.s, v2.s }[0], [x27], #8
+# CHECK-NEXT: [0,9]     . D===============================eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2225,43 +2225,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld2	{ v1.h, v2.h }[0], [x27], #4
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld2	{ v1.h, v2.h }[4], [x27], #4
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 3.     1     10.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld2	{ v1.h, v2.h }[0], [x27], x28
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 5.     1     17.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld2	{ v1.h, v2.h }[4], [x27], x28
-# CHECK-NEXT: 7.     1     32.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 7.     1     25.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld2	{ v1.s, v2.s }[0], [x27], #8
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.2   0.1    0.0       <total>
+# CHECK-NEXT: 9.     1     32.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     16.7   0.1    3.0       <total>
 
 # CHECK:      [23] Code Region - G24
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      2603
 # CHECK-NEXT: Total uOps:        2500
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.62
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    0.96
+# CHECK-NEXT: IPC:               0.38
 # CHECK-NEXT: Block RThroughput: 5.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld2	{ v1.s, v2.s }[0], [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld2	{ v1.d, v2.d }[0], [x27], #16
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld2	{ v1.d, v2.d }[0], [x27], x28
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld2r	{ v1.1d, v2.1d }, [x27], #16
-# CHECK-NEXT: [0,7]     .D===============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld2r	{ v1.2d, v2.2d }, [x27], #16
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     0123456789
+# CHECK-NEXT: Index     0123456789          012345678
+
+# CHECK:      [0,0]     DeeeeeeeeER    .    .    .  .   ld2	{ v1.s, v2.s }[0], [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R    .    .    .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .  .   ld2	{ v1.d, v2.d }[0], [x27], #16
+# CHECK-NEXT: [0,3]     D=========eE------R .    .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER .   ld2	{ v1.d, v2.d }[0], [x27], x28
+# CHECK-NEXT: [0,5]     .D================eE------R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D================eeeeeeeeER.   ld2r	{ v1.1d, v2.1d }, [x27], #16
+# CHECK-NEXT: [0,7]     .D=================eE------R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D================eeeeeeeeER   ld2r	{ v1.2d, v2.2d }, [x27], #16
+# CHECK-NEXT: [0,9]     . D=================eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2271,43 +2271,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld2	{ v1.s, v2.s }[0], [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld2	{ v1.d, v2.d }[0], [x27], #16
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 3.     1     10.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld2	{ v1.d, v2.d }[0], [x27], x28
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld2r	{ v1.1d, v2.1d }, [x27], #16
-# CHECK-NEXT: 7.     1     32.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld2r	{ v1.2d, v2.2d }, [x27], #16
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.2   0.1    0.0       <total>
+# CHECK-NEXT: 5.     1     17.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     17.0   0.0    0.0       ld2r	{ v1.1d, v2.1d }, [x27], #16
+# CHECK-NEXT: 7.     1     18.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     17.0   0.0    0.0       ld2r	{ v1.2d, v2.2d }, [x27], #16
+# CHECK-NEXT: 9.     1     18.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     12.5   0.1    3.0       <total>
 
 # CHECK:      [24] Code Region - G25
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      510
 # CHECK-NEXT: Total uOps:        2500
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.62
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    4.90
+# CHECK-NEXT: IPC:               1.96
 # CHECK-NEXT: Block RThroughput: 5.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld2r	{ v1.2s, v2.2s }, [x27], #8
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld2r	{ v1.4h, v2.4h }, [x27], #4
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld2r	{ v1.4s, v2.4s }, [x27], #8
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld2r	{ v1.8b, v2.8b }, [x27], #2
-# CHECK-NEXT: [0,7]     .D===============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld2r	{ v1.8h, v2.8h }, [x27], #4
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER   .   ld2r	{ v1.2s, v2.2s }, [x27], #8
+# CHECK-NEXT: [0,1]     D=eE------R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeeeER  .   ld2r	{ v1.4h, v2.4h }, [x27], #4
+# CHECK-NEXT: [0,3]     D==eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeeeER .   ld2r	{ v1.4s, v2.4s }, [x27], #8
+# CHECK-NEXT: [0,5]     .D==eE------R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeeeER.   ld2r	{ v1.8b, v2.8b }, [x27], #2
+# CHECK-NEXT: [0,7]     .D===eE------R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeeeER   ld2r	{ v1.8h, v2.8h }, [x27], #4
+# CHECK-NEXT: [0,9]     . D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2317,43 +2317,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld2r	{ v1.2s, v2.2s }, [x27], #8
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld2r	{ v1.4h, v2.4h }, [x27], #4
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld2r	{ v1.4s, v2.4s }, [x27], #8
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld2r	{ v1.8b, v2.8b }, [x27], #2
-# CHECK-NEXT: 7.     1     32.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld2r	{ v1.8h, v2.8h }, [x27], #4
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld2r	{ v1.4h, v2.4h }, [x27], #4
+# CHECK-NEXT: 3.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld2r	{ v1.4s, v2.4s }, [x27], #8
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld2r	{ v1.8b, v2.8b }, [x27], #2
+# CHECK-NEXT: 7.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld2r	{ v1.8h, v2.8h }, [x27], #4
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    3.0       <total>
 
 # CHECK:      [25] Code Region - G26
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      510
 # CHECK-NEXT: Total uOps:        2500
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.62
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    4.90
+# CHECK-NEXT: IPC:               1.96
 # CHECK-NEXT: Block RThroughput: 5.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld2r	{ v1.16b, v2.16b }, [x27], #2
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld2r	{ v1.1d, v2.1d }, [x27], x28
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld2r	{ v1.2d, v2.2d }, [x27], x28
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld2r	{ v1.2s, v2.2s }, [x27], x28
-# CHECK-NEXT: [0,7]     .D===============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld2r	{ v1.4h, v2.4h }, [x27], x28
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER   .   ld2r	{ v1.16b, v2.16b }, [x27], #2
+# CHECK-NEXT: [0,1]     D=eE------R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeeeER  .   ld2r	{ v1.1d, v2.1d }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeeeER .   ld2r	{ v1.2d, v2.2d }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eE------R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeeeER.   ld2r	{ v1.2s, v2.2s }, [x27], x28
+# CHECK-NEXT: [0,7]     .D===eE------R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeeeER   ld2r	{ v1.4h, v2.4h }, [x27], x28
+# CHECK-NEXT: [0,9]     . D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2363,43 +2363,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld2r	{ v1.16b, v2.16b }, [x27], #2
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld2r	{ v1.1d, v2.1d }, [x27], x28
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld2r	{ v1.2d, v2.2d }, [x27], x28
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld2r	{ v1.2s, v2.2s }, [x27], x28
-# CHECK-NEXT: 7.     1     32.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld2r	{ v1.4h, v2.4h }, [x27], x28
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld2r	{ v1.1d, v2.1d }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld2r	{ v1.2d, v2.2d }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld2r	{ v1.2s, v2.2s }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld2r	{ v1.4h, v2.4h }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    3.0       <total>
 
 # CHECK:      [26] Code Region - G27
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      609
 # CHECK-NEXT: Total uOps:        2800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.70
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    4.60
+# CHECK-NEXT: IPC:               1.64
 # CHECK-NEXT: Block RThroughput: 5.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld2r	{ v1.4s, v2.4s }, [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D========eeeeeeeeER .    .    .    .    .  .   ld2r	{ v1.8b, v2.8b }, [x27], x28
-# CHECK-NEXT: [0,3]     D================eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===============eeeeeeeeER   .    .    .  .   ld2r	{ v1.8h, v2.8h }, [x27], x28
-# CHECK-NEXT: [0,5]     .D=======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=======================eeeeeeeeER.    .  .   ld2r	{ v1.16b, v2.16b }, [x27], x28
-# CHECK-NEXT: [0,7]     .D===============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============================eeeeeeeeER.   ld3	{ v1.2d, v2.2d, v3.2d }, [x27], #48
-# CHECK-NEXT: [0,9]     . D======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER   .   ld2r	{ v1.4s, v2.4s }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeeeER  .   ld2r	{ v1.8b, v2.8b }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeeeER .   ld2r	{ v1.8h, v2.8h }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eE------R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeeeER.   ld2r	{ v1.16b, v2.16b }, [x27], x28
+# CHECK-NEXT: [0,7]     .D===eE------R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeeeER   ld3	{ v1.2d, v2.2d, v3.2d }, [x27], #48
+# CHECK-NEXT: [0,9]     . D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2409,43 +2409,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld2r	{ v1.4s, v2.4s }, [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     9.0    0.0    0.0       ld2r	{ v1.8b, v2.8b }, [x27], x28
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld2r	{ v1.8h, v2.8h }, [x27], x28
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld2r	{ v1.16b, v2.16b }, [x27], x28
-# CHECK-NEXT: 7.     1     32.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld3	{ v1.2d, v2.2d, v3.2d }, [x27], #48
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ld2r	{ v1.8b, v2.8b }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ld2r	{ v1.8h, v2.8h }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld2r	{ v1.16b, v2.16b }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ld3	{ v1.2d, v2.2d, v3.2d }, [x27], #48
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    3.0       <total>
 
 # CHECK:      [27] Code Region - G28
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      759
 # CHECK-NEXT: Total uOps:        3700
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.92
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    4.87
+# CHECK-NEXT: IPC:               1.32
 # CHECK-NEXT: Block RThroughput: 7.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld3	{ v1.2s, v2.2s, v3.2s }, [x27], #24
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld3	{ v1.4h, v2.4h, v3.4h }, [x27], #24
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld3	{ v1.4s, v2.4s, v3.4s }, [x27], #48
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld3	{ v1.8b, v2.8b, v3.8b }, [x27], #24
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld3	{ v1.8h, v2.8h, v3.8h }, [x27], #48
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     0123456
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER    ..   ld3	{ v1.2s, v2.2s, v3.2s }, [x27], #24
+# CHECK-NEXT: [0,1]     D=eE------R    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeER   ..   ld3	{ v1.4h, v2.4h, v3.4h }, [x27], #24
+# CHECK-NEXT: [0,3]     .D=eE------R   ..   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeeeER ..   ld3	{ v1.4s, v2.4s, v3.4s }, [x27], #48
+# CHECK-NEXT: [0,5]     . D==eE------R ..   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeeeER..   ld3	{ v1.8b, v2.8b, v3.8b }, [x27], #24
+# CHECK-NEXT: [0,7]     .  D==eE------R..   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eeeeeeeeER   ld3	{ v1.8h, v2.8h, v3.8h }, [x27], #48
+# CHECK-NEXT: [0,9]     .   D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2455,43 +2455,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld3	{ v1.2s, v2.2s, v3.2s }, [x27], #24
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld3	{ v1.4h, v2.4h, v3.4h }, [x27], #24
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld3	{ v1.4s, v2.4s, v3.4s }, [x27], #48
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld3	{ v1.8b, v2.8b, v3.8b }, [x27], #24
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld3	{ v1.8h, v2.8h, v3.8h }, [x27], #48
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld3	{ v1.4h, v2.4h, v3.4h }, [x27], #24
+# CHECK-NEXT: 3.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       ld3	{ v1.4s, v2.4s, v3.4s }, [x27], #48
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       ld3	{ v1.8b, v2.8b, v3.8b }, [x27], #24
+# CHECK-NEXT: 7.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       ld3	{ v1.8h, v2.8h, v3.8h }, [x27], #48
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    3.0       <total>
 
 # CHECK:      [28] Code Region - G29
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      759
 # CHECK-NEXT: Total uOps:        3800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.95
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    5.01
+# CHECK-NEXT: IPC:               1.32
 # CHECK-NEXT: Block RThroughput: 7.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld3	{ v1.16b, v2.16b, v3.16b }, [x27], #48
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld3	{ v1.2d, v2.2d, v3.2d }, [x27], x28
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld3	{ v1.2s, v2.2s, v3.2s }, [x27], x28
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld3	{ v1.4h, v2.4h, v3.4h }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld3	{ v1.4s, v2.4s, v3.4s }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     0123456
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER    ..   ld3	{ v1.16b, v2.16b, v3.16b }, [x27], #48
+# CHECK-NEXT: [0,1]     D=eE------R    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeER   ..   ld3	{ v1.2d, v2.2d, v3.2d }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE------R   ..   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeeeER ..   ld3	{ v1.2s, v2.2s, v3.2s }, [x27], x28
+# CHECK-NEXT: [0,5]     . D==eE------R ..   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeeeER..   ld3	{ v1.4h, v2.4h, v3.4h }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D==eE------R..   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eeeeeeeeER   ld3	{ v1.4s, v2.4s, v3.4s }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2501,43 +2501,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld3	{ v1.16b, v2.16b, v3.16b }, [x27], #48
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld3	{ v1.2d, v2.2d, v3.2d }, [x27], x28
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld3	{ v1.2s, v2.2s, v3.2s }, [x27], x28
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld3	{ v1.4h, v2.4h, v3.4h }, [x27], x28
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld3	{ v1.4s, v2.4s, v3.4s }, [x27], x28
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld3	{ v1.2d, v2.2d, v3.2d }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       ld3	{ v1.2s, v2.2s, v3.2s }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       ld3	{ v1.4h, v2.4h, v3.4h }, [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       ld3	{ v1.4s, v2.4s, v3.4s }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    3.0       <total>
 
 # CHECK:      [29] Code Region - G30
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      2010
 # CHECK-NEXT: Total uOps:        3700
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.92
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    1.84
+# CHECK-NEXT: IPC:               0.50
 # CHECK-NEXT: Block RThroughput: 7.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld3	{ v1.8b, v2.8b, v3.8b }, [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld3	{ v1.8h, v2.8h, v3.8h }, [x27], x28
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld3	{ v1.16b, v2.16b, v3.16b }, [x27], x28
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld3	{ v1.b, v2.b, v3.b }[0], [x27], #3
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld3	{ v1.b, v2.b, v3.b }[8], [x27], #3
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     0123456789
+# CHECK-NEXT: Index     0123456789          0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER    .    .    .   .   ld3	{ v1.8b, v2.8b, v3.8b }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R    .    .    .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeER   .    .    .   .   ld3	{ v1.8h, v2.8h, v3.8h }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE------R   .    .    .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeeeER .    .    .   .   ld3	{ v1.16b, v2.16b, v3.16b }, [x27], x28
+# CHECK-NEXT: [0,5]     . D==eE------R .    .    .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D========eeeeeeeeER   .   .   ld3	{ v1.b, v2.b, v3.b }[0], [x27], #3
+# CHECK-NEXT: [0,7]     .  D=========eE------R   .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D===============eeeeeeeeER   ld3	{ v1.b, v2.b, v3.b }[8], [x27], #3
+# CHECK-NEXT: [0,9]     .   D================eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2547,22 +2547,22 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld3	{ v1.8b, v2.8b, v3.8b }, [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld3	{ v1.8h, v2.8h, v3.8h }, [x27], x28
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld3	{ v1.16b, v2.16b, v3.16b }, [x27], x28
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld3	{ v1.b, v2.b, v3.b }[0], [x27], #3
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld3	{ v1.b, v2.b, v3.b }[8], [x27], #3
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld3	{ v1.8h, v2.8h, v3.8h }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       ld3	{ v1.16b, v2.16b, v3.16b }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     9.0    0.0    0.0       ld3	{ v1.b, v2.b, v3.b }[0], [x27], #3
+# CHECK-NEXT: 7.     1     10.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     16.0   0.0    0.0       ld3	{ v1.b, v2.b, v3.b }[8], [x27], #3
+# CHECK-NEXT: 9.     1     17.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     6.3    0.2    3.0       <total>
 
 # CHECK:      [30] Code Region - G31
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      4003
 # CHECK-NEXT: Total uOps:        3500
 
 # CHECK:      Dispatch Width:    10
@@ -2572,18 +2572,18 @@ add x0, x27, 1
 
 # CHECK:      Timeline view:
 # CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld3	{ v1.b, v2.b, v3.b }[0], [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld3	{ v1.b, v2.b, v3.b }[8], [x27], x28
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld3	{ v1.h, v2.h, v3.h }[0], [x27], #6
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld3	{ v1.h, v2.h, v3.h }[4], [x27], #6
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld3	{ v1.h, v2.h, v3.h }[0], [x27], x28
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT: Index     0123456789          0123456789          012
+
+# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    . .   ld3	{ v1.b, v2.b, v3.b }[0], [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R    .    .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    . .   ld3	{ v1.b, v2.b, v3.b }[8], [x27], x28
+# CHECK-NEXT: [0,3]     .D========eE------R .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    . .   ld3	{ v1.h, v2.h, v3.h }[0], [x27], #6
+# CHECK-NEXT: [0,5]     . D===============eE------R   .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    . .   ld3	{ v1.h, v2.h, v3.h }[4], [x27], #6
+# CHECK-NEXT: [0,7]     .  D======================eE------R.    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER   ld3	{ v1.h, v2.h, v3.h }[0], [x27], x28
+# CHECK-NEXT: [0,9]     .   D=============================eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2593,22 +2593,22 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld3	{ v1.b, v2.b, v3.b }[0], [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld3	{ v1.b, v2.b, v3.b }[8], [x27], x28
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 3.     1     9.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld3	{ v1.h, v2.h, v3.h }[0], [x27], #6
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 5.     1     16.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld3	{ v1.h, v2.h, v3.h }[4], [x27], #6
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 7.     1     23.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld3	{ v1.h, v2.h, v3.h }[0], [x27], x28
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 9.     1     30.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     15.5   0.1    3.0       <total>
 
 # CHECK:      [31] Code Region - G32
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      4003
 # CHECK-NEXT: Total uOps:        3500
 
 # CHECK:      Dispatch Width:    10
@@ -2618,18 +2618,18 @@ add x0, x27, 1
 
 # CHECK:      Timeline view:
 # CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld3	{ v1.h, v2.h, v3.h }[4], [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld3	{ v1.s, v2.s, v3.s }[0], [x27], #12
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld3	{ v1.s, v2.s, v3.s }[0], [x27], x28
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld3	{ v1.d, v2.d, v3.d }[0], [x27], #24
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld3	{ v1.d, v2.d, v3.d }[0], [x27], x28
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT: Index     0123456789          0123456789          012
+
+# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    . .   ld3	{ v1.h, v2.h, v3.h }[4], [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R    .    .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    . .   ld3	{ v1.s, v2.s, v3.s }[0], [x27], #12
+# CHECK-NEXT: [0,3]     .D========eE------R .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    . .   ld3	{ v1.s, v2.s, v3.s }[0], [x27], x28
+# CHECK-NEXT: [0,5]     . D===============eE------R   .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    . .   ld3	{ v1.d, v2.d, v3.d }[0], [x27], #24
+# CHECK-NEXT: [0,7]     .  D======================eE------R.    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER   ld3	{ v1.d, v2.d, v3.d }[0], [x27], x28
+# CHECK-NEXT: [0,9]     .   D=============================eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2639,43 +2639,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld3	{ v1.h, v2.h, v3.h }[4], [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld3	{ v1.s, v2.s, v3.s }[0], [x27], #12
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 3.     1     9.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld3	{ v1.s, v2.s, v3.s }[0], [x27], x28
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 5.     1     16.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld3	{ v1.d, v2.d, v3.d }[0], [x27], #24
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 7.     1     23.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld3	{ v1.d, v2.d, v3.d }[0], [x27], x28
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 9.     1     30.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     15.5   0.1    3.0       <total>
 
 # CHECK:      [32] Code Region - G33
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      759
 # CHECK-NEXT: Total uOps:        3700
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.92
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    4.87
+# CHECK-NEXT: IPC:               1.32
 # CHECK-NEXT: Block RThroughput: 7.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld3r	{ v1.1d, v2.1d, v3.1d }, [x27], #24
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld3r	{ v1.2d, v2.2d, v3.2d }, [x27], #24
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld3r	{ v1.2s, v2.2s, v3.2s }, [x27], #12
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld3r	{ v1.4h, v2.4h, v3.4h }, [x27], #6
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld3r	{ v1.4s, v2.4s, v3.4s }, [x27], #12
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     0123456
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER    ..   ld3r	{ v1.1d, v2.1d, v3.1d }, [x27], #24
+# CHECK-NEXT: [0,1]     D=eE------R    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeER   ..   ld3r	{ v1.2d, v2.2d, v3.2d }, [x27], #24
+# CHECK-NEXT: [0,3]     .D=eE------R   ..   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeeeER ..   ld3r	{ v1.2s, v2.2s, v3.2s }, [x27], #12
+# CHECK-NEXT: [0,5]     . D==eE------R ..   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeeeER..   ld3r	{ v1.4h, v2.4h, v3.4h }, [x27], #6
+# CHECK-NEXT: [0,7]     .  D==eE------R..   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eeeeeeeeER   ld3r	{ v1.4s, v2.4s, v3.4s }, [x27], #12
+# CHECK-NEXT: [0,9]     .   D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2685,43 +2685,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld3r	{ v1.1d, v2.1d, v3.1d }, [x27], #24
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld3r	{ v1.2d, v2.2d, v3.2d }, [x27], #24
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld3r	{ v1.2s, v2.2s, v3.2s }, [x27], #12
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld3r	{ v1.4h, v2.4h, v3.4h }, [x27], #6
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld3r	{ v1.4s, v2.4s, v3.4s }, [x27], #12
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld3r	{ v1.2d, v2.2d, v3.2d }, [x27], #24
+# CHECK-NEXT: 3.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       ld3r	{ v1.2s, v2.2s, v3.2s }, [x27], #12
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       ld3r	{ v1.4h, v2.4h, v3.4h }, [x27], #6
+# CHECK-NEXT: 7.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       ld3r	{ v1.4s, v2.4s, v3.4s }, [x27], #12
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    3.0       <total>
 
 # CHECK:      [33] Code Region - G34
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      759
 # CHECK-NEXT: Total uOps:        3800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.95
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    5.01
+# CHECK-NEXT: IPC:               1.32
 # CHECK-NEXT: Block RThroughput: 7.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld3r	{ v1.8b, v2.8b, v3.8b }, [x27], #3
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld3r	{ v1.8h, v2.8h, v3.8h }, [x27], #6
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld3r	{ v1.16b, v2.16b, v3.16b }, [x27], #3
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld3r	{ v1.1d, v2.1d, v3.1d }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld3r	{ v1.2d, v2.2d, v3.2d }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     0123456
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER    ..   ld3r	{ v1.8b, v2.8b, v3.8b }, [x27], #3
+# CHECK-NEXT: [0,1]     D=eE------R    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeER   ..   ld3r	{ v1.8h, v2.8h, v3.8h }, [x27], #6
+# CHECK-NEXT: [0,3]     .D=eE------R   ..   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeeeER ..   ld3r	{ v1.16b, v2.16b, v3.16b }, [x27], #3
+# CHECK-NEXT: [0,5]     . D==eE------R ..   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeeeER..   ld3r	{ v1.1d, v2.1d, v3.1d }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D==eE------R..   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eeeeeeeeER   ld3r	{ v1.2d, v2.2d, v3.2d }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2731,43 +2731,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld3r	{ v1.8b, v2.8b, v3.8b }, [x27], #3
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld3r	{ v1.8h, v2.8h, v3.8h }, [x27], #6
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld3r	{ v1.16b, v2.16b, v3.16b }, [x27], #3
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld3r	{ v1.1d, v2.1d, v3.1d }, [x27], x28
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld3r	{ v1.2d, v2.2d, v3.2d }, [x27], x28
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld3r	{ v1.8h, v2.8h, v3.8h }, [x27], #6
+# CHECK-NEXT: 3.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       ld3r	{ v1.16b, v2.16b, v3.16b }, [x27], #3
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       ld3r	{ v1.1d, v2.1d, v3.1d }, [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       ld3r	{ v1.2d, v2.2d, v3.2d }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    3.0       <total>
 
 # CHECK:      [34] Code Region - G35
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      759
 # CHECK-NEXT: Total uOps:        3700
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.92
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    4.87
+# CHECK-NEXT: IPC:               1.32
 # CHECK-NEXT: Block RThroughput: 7.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld3r	{ v1.2s, v2.2s, v3.2s }, [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld3r	{ v1.4h, v2.4h, v3.4h }, [x27], x28
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld3r	{ v1.4s, v2.4s, v3.4s }, [x27], x28
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld3r	{ v1.8b, v2.8b, v3.8b }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld3r	{ v1.8h, v2.8h, v3.8h }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     0123456
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER    ..   ld3r	{ v1.2s, v2.2s, v3.2s }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeER   ..   ld3r	{ v1.4h, v2.4h, v3.4h }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE------R   ..   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeeeER ..   ld3r	{ v1.4s, v2.4s, v3.4s }, [x27], x28
+# CHECK-NEXT: [0,5]     . D==eE------R ..   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeeeER..   ld3r	{ v1.8b, v2.8b, v3.8b }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D==eE------R..   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eeeeeeeeER   ld3r	{ v1.8h, v2.8h, v3.8h }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D===eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2777,43 +2777,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld3r	{ v1.2s, v2.2s, v3.2s }, [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld3r	{ v1.4h, v2.4h, v3.4h }, [x27], x28
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld3r	{ v1.4s, v2.4s, v3.4s }, [x27], x28
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld3r	{ v1.8b, v2.8b, v3.8b }, [x27], x28
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld3r	{ v1.8h, v2.8h, v3.8h }, [x27], x28
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld3r	{ v1.4h, v2.4h, v3.4h }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       ld3r	{ v1.4s, v2.4s, v3.4s }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       ld3r	{ v1.8b, v2.8b, v3.8b }, [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       ld3r	{ v1.8h, v2.8h, v3.8h }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    3.0       <total>
 
 # CHECK:      [35] Code Region - G36
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4204
+# CHECK-NEXT: Total Cycles:      959
 # CHECK-NEXT: Total uOps:        4600
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.09
-# CHECK-NEXT: IPC:               0.24
+# CHECK-NEXT: uOps Per Cycle:    4.80
+# CHECK-NEXT: IPC:               1.04
 # CHECK-NEXT: Block RThroughput: 9.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          012345
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .    .   ld3r	{ v1.16b, v2.16b, v3.16b }, [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeeER.    .    .    .    .    .   ld4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
-# CHECK-NEXT: [0,3]     .D================eER    .    .    .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D===============eeeeeeeeER  .    .    .    .   ld4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
-# CHECK-NEXT: [0,5]     . D=======================eER .    .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D======================eeeeeeeeER    .    .   ld4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
-# CHECK-NEXT: [0,7]     .  D==============================eER   .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D=============================eeeeeeeeeER.   ld4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
-# CHECK-NEXT: [0,9]     .   D======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     012345678
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER    .  .   ld3r	{ v1.16b, v2.16b, v3.16b }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R    .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeeER  .  .   ld4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
+# CHECK-NEXT: [0,3]     .D=eE-------R  .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeeeER .  .   ld4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
+# CHECK-NEXT: [0,5]     . D==eE------R .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D==eeeeeeeeER  .   ld4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
+# CHECK-NEXT: [0,7]     .  D===eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D===eeeeeeeeeER   ld4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
+# CHECK-NEXT: [0,9]     .   D====eE-------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2823,43 +2823,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld3r	{ v1.16b, v2.16b, v3.16b }, [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     23.0   0.0    0.0       ld4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
-# CHECK-NEXT: 7.     1     31.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     30.0   0.0    0.0       ld4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
-# CHECK-NEXT: 9.     1     39.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.8   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
+# CHECK-NEXT: 3.     1     2.0    0.0    7.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       ld4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
+# CHECK-NEXT: 5.     1     3.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    1.0    0.0       ld4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
+# CHECK-NEXT: 7.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     4.0    1.0    0.0       ld4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
+# CHECK-NEXT: 9.     1     5.0    0.0    7.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.4    3.2       <total>
 
 # CHECK:      [36] Code Region - G37
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4304
+# CHECK-NEXT: Total Cycles:      1008
 # CHECK-NEXT: Total uOps:        4800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.12
-# CHECK-NEXT: IPC:               0.23
+# CHECK-NEXT: uOps Per Cycle:    4.76
+# CHECK-NEXT: IPC:               0.99
 # CHECK-NEXT: Block RThroughput: 10.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123456
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .    ..   ld4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeeER.    .    .    .    .    ..   ld4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
-# CHECK-NEXT: [0,3]     .D================eER    .    .    .    .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D===============eeeeeeeeeER .    .    .    ..   ld4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
-# CHECK-NEXT: [0,5]     . D========================eER.    .    .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=======================eeeeeeeeeER  .    ..   ld4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D================================eER .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D===============================eeeeeeeeER.   ld4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D=======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     012345678
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER    .  .   ld4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
+# CHECK-NEXT: [0,1]     D=eE------R    .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeeER  .  .   ld4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
+# CHECK-NEXT: [0,3]     .D=eE-------R  .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D==eeeeeeeeeER  .   ld4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
+# CHECK-NEXT: [0,5]     . D===eE-------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D==eeeeeeeeeER .   ld4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D===eE-------R .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D====eeeeeeeeER   ld4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D=====eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2869,43 +2869,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
-# CHECK-NEXT: 5.     1     25.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     24.0   0.0    0.0       ld4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
-# CHECK-NEXT: 7.     1     33.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     32.0   0.0    0.0       ld4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
-# CHECK-NEXT: 9.     1     40.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.5   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
+# CHECK-NEXT: 3.     1     2.0    0.0    7.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    2.0    0.0       ld4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
+# CHECK-NEXT: 5.     1     4.0    0.0    7.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    7.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     5.0    2.0    0.0       ld4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
+# CHECK-NEXT: 9.     1     6.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.1    0.5    3.3       <total>
 
 # CHECK:      [37] Code Region - G38
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4304
+# CHECK-NEXT: Total Cycles:      1009
 # CHECK-NEXT: Total uOps:        4800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.12
-# CHECK-NEXT: IPC:               0.23
+# CHECK-NEXT: uOps Per Cycle:    4.76
+# CHECK-NEXT: IPC:               0.99
 # CHECK-NEXT: Block RThroughput: 10.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123456
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .    ..   ld4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeeER.    .    .    .    .    ..   ld4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
-# CHECK-NEXT: [0,3]     .D================eER    .    .    .    .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D===============eeeeeeeeER  .    .    .    ..   ld4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
-# CHECK-NEXT: [0,5]     . D=======================eER .    .    .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D======================eeeeeeeeeER   .    ..   ld4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D===============================eER  .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D==============================eeeeeeeeeER.   ld4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D=======================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     0123456789
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER    .   .   ld4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R    .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeeER  .   .   ld4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE-------R  .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D==eeeeeeeeER.   .   ld4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
+# CHECK-NEXT: [0,5]     . D===eE------R.   .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D==eeeeeeeeeER  .   ld4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D===eE-------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D====eeeeeeeeeER   ld4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D=====eE-------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2915,22 +2915,22 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
-# CHECK-NEXT: 3.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     16.0   0.0    0.0       ld4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
-# CHECK-NEXT: 5.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     23.0   0.0    0.0       ld4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
-# CHECK-NEXT: 7.     1     32.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     31.0   0.0    0.0       ld4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
-# CHECK-NEXT: 9.     1     40.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     20.1   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    7.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    2.0    0.0       ld4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
+# CHECK-NEXT: 5.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    7.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     5.0    2.0    0.0       ld4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
+# CHECK-NEXT: 9.     1     6.0    0.0    7.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.1    0.5    3.3       <total>
 
 # CHECK:      [38] Code Region - G39
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      4003
 # CHECK-NEXT: Total uOps:        4500
 
 # CHECK:      Dispatch Width:    10
@@ -2940,18 +2940,18 @@ add x0, x27, 1
 
 # CHECK:      Timeline view:
 # CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], #4
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], #4
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], x28
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], x28
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], #8
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT: Index     0123456789          0123456789          012
+
+# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    . .   ld4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], #4
+# CHECK-NEXT: [0,1]     D=eE------R    .    .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    . .   ld4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], #4
+# CHECK-NEXT: [0,3]     .D========eE------R .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    . .   ld4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], x28
+# CHECK-NEXT: [0,5]     . D===============eE------R   .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    . .   ld4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], x28
+# CHECK-NEXT: [0,7]     .  D======================eE------R.    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER   ld4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], #8
+# CHECK-NEXT: [0,9]     .   D=============================eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -2961,22 +2961,22 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], #4
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], #4
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 3.     1     9.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], x28
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 5.     1     16.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], x28
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 7.     1     23.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], #8
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 9.     1     30.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     15.5   0.1    3.0       <total>
 
 # CHECK:      [39] Code Region - G40
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      4003
 # CHECK-NEXT: Total uOps:        4500
 
 # CHECK:      Dispatch Width:    10
@@ -2986,18 +2986,18 @@ add x0, x27, 1
 
 # CHECK:      Timeline view:
 # CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], #8
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], x28
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], x28
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], #16
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], x28
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT: Index     0123456789          0123456789          012
+
+# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    . .   ld4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], #8
+# CHECK-NEXT: [0,1]     D=eE------R    .    .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    . .   ld4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], x28
+# CHECK-NEXT: [0,3]     .D========eE------R .    .    .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    . .   ld4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], x28
+# CHECK-NEXT: [0,5]     . D===============eE------R   .    .    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    . .   ld4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], #16
+# CHECK-NEXT: [0,7]     .  D======================eE------R.    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER   ld4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], x28
+# CHECK-NEXT: [0,9]     .   D=============================eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3007,43 +3007,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], #8
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], x28
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 3.     1     9.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], x28
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 5.     1     16.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], #16
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 7.     1     23.0   0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], x28
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 9.     1     30.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     15.5   0.1    3.0       <total>
 
 # CHECK:      [40] Code Region - G41
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      2103
 # CHECK-NEXT: Total uOps:        4600
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.15
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    2.19
+# CHECK-NEXT: IPC:               0.48
 # CHECK-NEXT: Block RThroughput: 10.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], #32
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], x28
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld4r	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld4r	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #32
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld4r	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #16
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     0123456789
+# CHECK-NEXT: Index     0123456789          0123
+
+# CHECK:      [0,0]     DeeeeeeeeER    .    .  .   ld4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], #32
+# CHECK-NEXT: [0,1]     D=eE------R    .    .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .  .   ld4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], x28
+# CHECK-NEXT: [0,3]     .D========eE------R .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=======eeeeeeeeER.  .   ld4r	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
+# CHECK-NEXT: [0,5]     . D========eE------R.  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=========eeeeeeeeER.   ld4r	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #32
+# CHECK-NEXT: [0,7]     .  D==========eE------R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D=========eeeeeeeeER   ld4r	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #16
+# CHECK-NEXT: [0,9]     .   D==========eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3053,43 +3053,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], #32
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
 # CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], x28
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld4r	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld4r	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #32
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld4r	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #16
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 3.     1     9.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     8.0    0.0    0.0       ld4r	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
+# CHECK-NEXT: 5.     1     9.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     10.0   2.0    0.0       ld4r	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #32
+# CHECK-NEXT: 7.     1     11.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     10.0   0.0    0.0       ld4r	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #16
+# CHECK-NEXT: 9.     1     11.0   0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     7.9    0.3    3.0       <total>
 
 # CHECK:      [41] Code Region - G42
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      1008
 # CHECK-NEXT: Total uOps:        4800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.20
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    4.76
+# CHECK-NEXT: IPC:               0.99
 # CHECK-NEXT: Block RThroughput: 10.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld4r	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #8
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld4r	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #16
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld4r	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #4
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld4r	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #8
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld4r	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #4
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     012345678
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER    .  .   ld4r	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #8
+# CHECK-NEXT: [0,1]     D=eE------R    .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeER   .  .   ld4r	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #16
+# CHECK-NEXT: [0,3]     .D=eE------R   .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D==eeeeeeeeER.  .   ld4r	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #4
+# CHECK-NEXT: [0,5]     . D===eE------R.  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D==eeeeeeeeER  .   ld4r	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #8
+# CHECK-NEXT: [0,7]     .  D===eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D====eeeeeeeeER   ld4r	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #4
+# CHECK-NEXT: [0,9]     .   D=====eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3099,43 +3099,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld4r	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #8
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld4r	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #16
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld4r	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #4
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld4r	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #8
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld4r	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #4
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld4r	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #16
+# CHECK-NEXT: 3.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    2.0    0.0       ld4r	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #4
+# CHECK-NEXT: 5.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld4r	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #8
+# CHECK-NEXT: 7.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     5.0    2.0    0.0       ld4r	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #4
+# CHECK-NEXT: 9.     1     6.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.1    0.5    3.0       <total>
 
 # CHECK:      [42] Code Region - G43
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      4004
+# CHECK-NEXT: Total Cycles:      1008
 # CHECK-NEXT: Total uOps:        4700
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.17
-# CHECK-NEXT: IPC:               0.25
+# CHECK-NEXT: uOps Per Cycle:    4.66
+# CHECK-NEXT: IPC:               0.99
 # CHECK-NEXT: Block RThroughput: 10.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789          0123
-
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .    .  .   ld4r	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .    .  .   ld4r	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .    .  .   ld4r	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
-# CHECK-NEXT: [0,5]     . D======================eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeeeER.    .  .   ld4r	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D=============================eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D============================eeeeeeeeER.   ld4r	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D====================================eER   add	x0, x27, #1
+# CHECK-NEXT:                     012345678
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeeeER    .  .   ld4r	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R    .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeER   .  .   ld4r	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE------R   .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D==eeeeeeeeER.  .   ld4r	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
+# CHECK-NEXT: [0,5]     . D===eE------R.  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D==eeeeeeeeER  .   ld4r	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D===eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D====eeeeeeeeER   ld4r	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D=====eE------R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3145,43 +3145,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld4r	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld4r	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld4r	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     22.0   0.0    0.0       ld4r	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
-# CHECK-NEXT: 7.     1     30.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     29.0   0.0    0.0       ld4r	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
-# CHECK-NEXT: 9.     1     37.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     19.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld4r	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    2.0    0.0       ld4r	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
+# CHECK-NEXT: 5.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ld4r	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     5.0    2.0    0.0       ld4r	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
+# CHECK-NEXT: 9.     1     6.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.1    0.5    3.0       <total>
 
 # CHECK:      [43] Code Region - G44
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3604
+# CHECK-NEXT: Total Cycles:      708
 # CHECK-NEXT: Total uOps:        3900
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.08
-# CHECK-NEXT: IPC:               0.28
+# CHECK-NEXT: uOps Per Cycle:    5.51
+# CHECK-NEXT: IPC:               1.41
 # CHECK-NEXT: Block RThroughput: 6.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456789
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeeeER    .    .    .    .    .   .   ld4r	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
-# CHECK-NEXT: [0,1]     D========eER   .    .    .    .    .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=======eeeeeeeeER .    .    .    .   .   ld4r	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
-# CHECK-NEXT: [0,3]     .D===============eER.    .    .    .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==============eeeeeeeeER   .    .   .   ld4r	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
-# CHECK-NEXT: [0,5]     . D======================eER  .    .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=====================eeeeeeER  .   .   ldp	s1, s2, [x27], #248
-# CHECK-NEXT: [0,7]     .  D===========================eER .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .  D===========================eeeeeeER.   ldp	d1, d2, [x27], #496
-# CHECK-NEXT: [0,9]     .  D=================================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeeeER   .   ld4r	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE------R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeeER  .   ld4r	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE------R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D==eeeeeeeeER   ld4r	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
+# CHECK-NEXT: [0,5]     . D===eE------R   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D==eeeeeeE-R   ldp	s1, s2, [x27], #248
+# CHECK-NEXT: [0,7]     .  D===eE-----R   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .  D===eeeeeeER   ldp	d1, d2, [x27], #496
+# CHECK-NEXT: [0,9]     .  D====eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3191,43 +3191,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ld4r	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
-# CHECK-NEXT: 1.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     8.0    0.0    0.0       ld4r	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
-# CHECK-NEXT: 3.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     15.0   0.0    0.0       ld4r	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
-# CHECK-NEXT: 5.     1     23.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     22.0   0.0    0.0       ldp	s1, s2, [x27], #248
-# CHECK-NEXT: 7.     1     28.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     28.0   0.0    0.0       ldp	d1, d2, [x27], #496
-# CHECK-NEXT: 9.     1     34.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     18.4   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ld4r	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    2.0    0.0       ld4r	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
+# CHECK-NEXT: 5.     1     4.0    0.0    6.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    1.0       ldp	s1, s2, [x27], #248
+# CHECK-NEXT: 7.     1     4.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     4.0    0.0    0.0       ldp	d1, d2, [x27], #496
+# CHECK-NEXT: 9.     1     5.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.9    0.3    2.8       <total>
 
 # CHECK:      [44] Code Region - G45
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2506
+# CHECK-NEXT: Total Cycles:      507
 # CHECK-NEXT: Total uOps:        2800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.12
-# CHECK-NEXT: IPC:               0.40
+# CHECK-NEXT: uOps Per Cycle:    5.52
+# CHECK-NEXT: IPC:               1.97
 # CHECK-NEXT: Block RThroughput: 4.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     01
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .   ldp	q1, q2, [x27], #992
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=====eeeeeeER.    .    .    .   ldp	s1, s2, [x27, #248]!
-# CHECK-NEXT: [0,3]     .D===========eER    .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===========eeeeeeER    .    .   ldp	d1, d2, [x27, #496]!
-# CHECK-NEXT: [0,5]     .D=================eER   .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     . D================eeeeeeER   .   ldp	q1, q2, [x27, #992]!
-# CHECK-NEXT: [0,7]     . D======================eER  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======================eeeeER   ldp	w1, w2, [x27], #248
-# CHECK-NEXT: [0,9]     .  D======================eE--R   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER ..   ldp	q1, q2, [x27], #992
+# CHECK-NEXT: [0,1]     D=eE----R ..   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeER..   ldp	s1, s2, [x27, #248]!
+# CHECK-NEXT: [0,3]     .D=eE----R..   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeER.   ldp	d1, d2, [x27, #496]!
+# CHECK-NEXT: [0,5]     .D==eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     . D=eeeeeeER   ldp	q1, q2, [x27, #992]!
+# CHECK-NEXT: [0,7]     . D==eE----R   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeE-R   ldp	w1, w2, [x27], #248
+# CHECK-NEXT: [0,9]     .  D==eE---R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3237,43 +3237,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ldp	q1, q2, [x27], #992
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     6.0    0.0    0.0       ldp	s1, s2, [x27, #248]!
-# CHECK-NEXT: 3.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     12.0   0.0    0.0       ldp	d1, d2, [x27, #496]!
-# CHECK-NEXT: 5.     1     18.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     17.0   0.0    0.0       ldp	q1, q2, [x27, #992]!
-# CHECK-NEXT: 7.     1     23.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     23.0   0.0    0.0       ldp	w1, w2, [x27], #248
-# CHECK-NEXT: 9.     1     23.0   0.0    2.0       add	x0, x27, #1
-# CHECK-NEXT:        1     14.2   0.1    0.2       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       ldp	s1, s2, [x27, #248]!
+# CHECK-NEXT: 3.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ldp	d1, d2, [x27, #496]!
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       ldp	q1, q2, [x27, #992]!
+# CHECK-NEXT: 7.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    1.0       ldp	w1, w2, [x27], #248
+# CHECK-NEXT: 9.     1     3.0    0.0    3.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.2    0.1    2.0       <total>
 
 # CHECK:      [45] Code Region - G46
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1304
+# CHECK-NEXT: Total Cycles:      1006
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.53
-# CHECK-NEXT: IPC:               0.77
+# CHECK-NEXT: uOps Per Cycle:    1.99
+# CHECK-NEXT: IPC:               0.99
 # CHECK-NEXT: Block RThroughput: 10.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456
+# CHECK-NEXT:                     012345
 # CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeER   .    ..   ldp	x1, x2, [x27], #496
-# CHECK-NEXT: [0,1]     D=eE--R   .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D=eeeeER  .    ..   ldp	w1, w2, [x27, #248]!
-# CHECK-NEXT: [0,3]     D==eE--R  .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D=eeeeER .    ..   ldp	x1, x2, [x27, #496]!
-# CHECK-NEXT: [0,5]     .D==eE--R .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D==eeeeeER    ..   ldpsw	x1, x2, [x27], #248
-# CHECK-NEXT: [0,7]     .D=======eER   ..   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======eeeeeER.   ldpsw	x1, x2, [x27, #248]!
-# CHECK-NEXT: [0,9]     . D===========eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeER   .    .   ldp	x1, x2, [x27], #496
+# CHECK-NEXT: [0,1]     D=eE--R   .    .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeER  .    .   ldp	w1, w2, [x27, #248]!
+# CHECK-NEXT: [0,3]     D==eE--R  .    .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeER .    .   ldp	x1, x2, [x27, #496]!
+# CHECK-NEXT: [0,5]     .D==eE--R .    .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeER    .   ldpsw	x1, x2, [x27], #248
+# CHECK-NEXT: [0,7]     .D===eE---R    .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D======eeeeeER   ldpsw	x1, x2, [x27, #248]!
+# CHECK-NEXT: [0,9]     . D=======eE---R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3289,10 +3289,10 @@ add x0, x27, 1
 # CHECK-NEXT: 4.     1     2.0    0.0    0.0       ldp	x1, x2, [x27, #496]!
 # CHECK-NEXT: 5.     1     3.0    0.0    2.0       add	x0, x27, #1
 # CHECK-NEXT: 6.     1     3.0    0.0    0.0       ldpsw	x1, x2, [x27], #248
-# CHECK-NEXT: 7.     1     8.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     7.0    0.0    0.0       ldpsw	x1, x2, [x27, #248]!
-# CHECK-NEXT: 9.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     4.3    0.1    0.6       <total>
+# CHECK-NEXT: 7.     1     4.0    0.0    3.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     7.0    4.0    0.0       ldpsw	x1, x2, [x27, #248]!
+# CHECK-NEXT: 9.     1     8.0    0.0    3.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.5    0.5    1.2       <total>
 
 # CHECK:      [46] Code Region - G47
 
@@ -3344,28 +3344,28 @@ add x0, x27, 1
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      508
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    0.67
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    3.94
+# CHECK-NEXT: IPC:               1.97
 # CHECK-NEXT: Block RThroughput: 3.8
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     012
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   ldr	b1, [x27, #254]!
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D======eeeeeeER.    .    .    .  .   ldr	h1, [x27, #254]!
-# CHECK-NEXT: [0,3]     D============eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===========eeeeeeER    .    .  .   ldr	s1, [x27, #254]!
-# CHECK-NEXT: [0,5]     .D=================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=================eeeeeeER   .  .   ldr	d1, [x27, #254]!
-# CHECK-NEXT: [0,7]     .D=======================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======================eeeeeeER.   ldr	q1, [x27, #254]!
-# CHECK-NEXT: [0,9]     . D============================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER . .   ldr	b1, [x27, #254]!
+# CHECK-NEXT: [0,1]     D=eE----R . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeeeER. .   ldr	h1, [x27, #254]!
+# CHECK-NEXT: [0,3]     D==eE----R. .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeER .   ldr	s1, [x27, #254]!
+# CHECK-NEXT: [0,5]     .D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeeeER.   ldr	d1, [x27, #254]!
+# CHECK-NEXT: [0,7]     .D===eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeeeER   ldr	q1, [x27, #254]!
+# CHECK-NEXT: [0,9]     . D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3375,16 +3375,16 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       ldr	b1, [x27, #254]!
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     7.0    0.0    0.0       ldr	h1, [x27, #254]!
-# CHECK-NEXT: 3.     1     13.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     12.0   0.0    0.0       ldr	s1, [x27, #254]!
-# CHECK-NEXT: 5.     1     18.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     18.0   0.0    0.0       ldr	d1, [x27, #254]!
-# CHECK-NEXT: 7.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     23.0   0.0    0.0       ldr	q1, [x27, #254]!
-# CHECK-NEXT: 9.     1     29.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     15.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       ldr	h1, [x27, #254]!
+# CHECK-NEXT: 3.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       ldr	s1, [x27, #254]!
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       ldr	d1, [x27, #254]!
+# CHECK-NEXT: 7.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       ldr	q1, [x27, #254]!
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    2.0       <total>
 
 # CHECK:      [48] Code Region - G49
 
@@ -3528,28 +3528,27 @@ add x0, x27, 1
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      704
+# CHECK-NEXT: Total Cycles:      504
 # CHECK-NEXT: Total uOps:        1700
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    2.41
-# CHECK-NEXT: IPC:               1.42
+# CHECK-NEXT: uOps Per Cycle:    3.37
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0
-# CHECK-NEXT: Index     0123456789
+# CHECK-NEXT: Index     012345678
 
-# CHECK:      [0,0]     DeeeeER   .   ldrsh	x1, [x27, #254]!
-# CHECK-NEXT: [0,1]     D=eE--R   .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D=eeeeER  .   ldrsw	x1, [x27], #254
-# CHECK-NEXT: [0,3]     D==eE--R  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     D==eeeeER .   ldrsw	x1, [x27, #254]!
-# CHECK-NEXT: [0,5]     D===eE--R .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D==eeE-R .   st1	{ v1.1d }, [x27], #8
-# CHECK-NEXT: [0,7]     .D====eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .D====eeER.   st1	{ v1.2d }, [x27], #16
-# CHECK-NEXT: [0,9]     .D======eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeER .   ldrsh	x1, [x27, #254]!
+# CHECK-NEXT: [0,1]     D=eE--R .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeER.   ldrsw	x1, [x27], #254
+# CHECK-NEXT: [0,3]     D==eE--R.   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     D==eeeeER   ldrsw	x1, [x27, #254]!
+# CHECK-NEXT: [0,5]     D===eE--R   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeE-R   st1	{ v1.1d }, [x27], #8
+# CHECK-NEXT: [0,7]     .D===eE-R   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .D===eeER   st1	{ v1.2d }, [x27], #16
+# CHECK-NEXT: [0,9]     .D====eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3565,37 +3564,36 @@ add x0, x27, 1
 # CHECK-NEXT: 4.     1     3.0    0.0    0.0       ldrsw	x1, [x27, #254]!
 # CHECK-NEXT: 5.     1     4.0    0.0    2.0       add	x0, x27, #1
 # CHECK-NEXT: 6.     1     3.0    0.0    1.0       st1	{ v1.1d }, [x27], #8
-# CHECK-NEXT: 7.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     5.0    0.0    0.0       st1	{ v1.2d }, [x27], #16
-# CHECK-NEXT: 9.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     3.5    0.1    0.7       <total>
+# CHECK-NEXT: 7.     1     4.0    0.0    1.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     4.0    0.0    0.0       st1	{ v1.2d }, [x27], #16
+# CHECK-NEXT: 9.     1     5.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.1    0.1    0.8       <total>
 
 # CHECK:      [52] Code Region - G53
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      504
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.99
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    3.97
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
-# CHECK-NEXT: Index     0123456789
+# CHECK-NEXT: Index     012345678
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.2s }, [x27], #8
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D==eeER   .  .   st1	{ v1.4h }, [x27], #8
-# CHECK-NEXT: [0,3]     D====eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===eeER .  .   st1	{ v1.4s }, [x27], #16
-# CHECK-NEXT: [0,5]     .D=====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=====eeER  .   st1	{ v1.8b }, [x27], #8
-# CHECK-NEXT: [0,7]     .D=======eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======eeER.   st1	{ v1.8h }, [x27], #16
-# CHECK-NEXT: [0,9]     . D========eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.  .   st1	{ v1.2s }, [x27], #8
+# CHECK-NEXT: [0,1]     D=eER.  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeER  .   st1	{ v1.4h }, [x27], #8
+# CHECK-NEXT: [0,3]     D==eER  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeER .   st1	{ v1.4s }, [x27], #16
+# CHECK-NEXT: [0,5]     .D==eER .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeER.   st1	{ v1.8b }, [x27], #8
+# CHECK-NEXT: [0,7]     .D===eER.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeER   st1	{ v1.8h }, [x27], #16
+# CHECK-NEXT: [0,9]     . D===eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3605,43 +3603,42 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.2s }, [x27], #8
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     3.0    0.0    0.0       st1	{ v1.4h }, [x27], #8
-# CHECK-NEXT: 3.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     4.0    0.0    0.0       st1	{ v1.4s }, [x27], #16
-# CHECK-NEXT: 5.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     6.0    0.0    0.0       st1	{ v1.8b }, [x27], #8
-# CHECK-NEXT: 7.     1     8.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     7.0    0.0    0.0       st1	{ v1.8h }, [x27], #16
-# CHECK-NEXT: 9.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     5.2    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.4h }, [x27], #8
+# CHECK-NEXT: 3.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st1	{ v1.4s }, [x27], #16
+# CHECK-NEXT: 5.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st1	{ v1.8b }, [x27], #8
+# CHECK-NEXT: 7.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st1	{ v1.8h }, [x27], #16
+# CHECK-NEXT: 9.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    0.0       <total>
 
 # CHECK:      [53] Code Region - G54
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      504
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.99
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    3.97
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
-# CHECK-NEXT: Index     0123456789
+# CHECK-NEXT: Index     012345678
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.16b }, [x27], #16
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D==eeER   .  .   st1	{ v1.1d }, [x27], x28
-# CHECK-NEXT: [0,3]     D====eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===eeER .  .   st1	{ v1.2d }, [x27], x28
-# CHECK-NEXT: [0,5]     .D=====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=====eeER  .   st1	{ v1.2s }, [x27], x28
-# CHECK-NEXT: [0,7]     .D=======eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======eeER.   st1	{ v1.4h }, [x27], x28
-# CHECK-NEXT: [0,9]     . D========eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.  .   st1	{ v1.16b }, [x27], #16
+# CHECK-NEXT: [0,1]     D=eER.  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeER  .   st1	{ v1.1d }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eER  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeER .   st1	{ v1.2d }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eER .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeER.   st1	{ v1.2s }, [x27], x28
+# CHECK-NEXT: [0,7]     .D===eER.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeER   st1	{ v1.4h }, [x27], x28
+# CHECK-NEXT: [0,9]     . D===eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3651,43 +3648,42 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.16b }, [x27], #16
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     3.0    0.0    0.0       st1	{ v1.1d }, [x27], x28
-# CHECK-NEXT: 3.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     4.0    0.0    0.0       st1	{ v1.2d }, [x27], x28
-# CHECK-NEXT: 5.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     6.0    0.0    0.0       st1	{ v1.2s }, [x27], x28
-# CHECK-NEXT: 7.     1     8.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     7.0    0.0    0.0       st1	{ v1.4h }, [x27], x28
-# CHECK-NEXT: 9.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     5.2    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.1d }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st1	{ v1.2d }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st1	{ v1.2s }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st1	{ v1.4h }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    0.0       <total>
 
 # CHECK:      [54] Code Region - G55
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      504
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.99
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    3.97
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
-# CHECK-NEXT: Index     0123456789
+# CHECK-NEXT: Index     012345678
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.4s }, [x27], x28
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D==eeER   .  .   st1	{ v1.8b }, [x27], x28
-# CHECK-NEXT: [0,3]     D====eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===eeER .  .   st1	{ v1.8h }, [x27], x28
-# CHECK-NEXT: [0,5]     .D=====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=====eeER  .   st1	{ v1.16b }, [x27], x28
-# CHECK-NEXT: [0,7]     .D=======eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======eeER.   st1	{ v1.1d, v2.1d }, [x27], #16
-# CHECK-NEXT: [0,9]     . D========eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.  .   st1	{ v1.4s }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eER.  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeER  .   st1	{ v1.8b }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eER  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeER .   st1	{ v1.8h }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eER .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeER.   st1	{ v1.16b }, [x27], x28
+# CHECK-NEXT: [0,7]     .D===eER.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeER   st1	{ v1.1d, v2.1d }, [x27], #16
+# CHECK-NEXT: [0,9]     . D===eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3697,43 +3693,42 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.4s }, [x27], x28
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     3.0    0.0    0.0       st1	{ v1.8b }, [x27], x28
-# CHECK-NEXT: 3.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     4.0    0.0    0.0       st1	{ v1.8h }, [x27], x28
-# CHECK-NEXT: 5.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     6.0    0.0    0.0       st1	{ v1.16b }, [x27], x28
-# CHECK-NEXT: 7.     1     8.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     7.0    0.0    0.0       st1	{ v1.1d, v2.1d }, [x27], #16
-# CHECK-NEXT: 9.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     5.2    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.8b }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st1	{ v1.8h }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st1	{ v1.16b }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st1	{ v1.1d, v2.1d }, [x27], #16
+# CHECK-NEXT: 9.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    0.0       <total>
 
 # CHECK:      [55] Code Region - G56
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      504
 # CHECK-NEXT: Total uOps:        2400
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    2.39
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    4.76
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 3.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
-# CHECK-NEXT: Index     0123456789
+# CHECK-NEXT: Index     012345678
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.2d, v2.2d }, [x27], #32
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D==eeER   .  .   st1	{ v1.2s, v2.2s }, [x27], #16
-# CHECK-NEXT: [0,3]     D====eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===eeER .  .   st1	{ v1.4h, v2.4h }, [x27], #16
-# CHECK-NEXT: [0,5]     .D=====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=====eeER  .   st1	{ v1.4s, v2.4s }, [x27], #32
-# CHECK-NEXT: [0,7]     .D=======eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======eeER.   st1	{ v1.8b, v2.8b }, [x27], #16
-# CHECK-NEXT: [0,9]     . D========eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.  .   st1	{ v1.2d, v2.2d }, [x27], #32
+# CHECK-NEXT: [0,1]     D=eER.  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeER  .   st1	{ v1.2s, v2.2s }, [x27], #16
+# CHECK-NEXT: [0,3]     D==eER  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeER .   st1	{ v1.4h, v2.4h }, [x27], #16
+# CHECK-NEXT: [0,5]     .D==eER .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeER.   st1	{ v1.4s, v2.4s }, [x27], #32
+# CHECK-NEXT: [0,7]     .D===eER.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeER   st1	{ v1.8b, v2.8b }, [x27], #16
+# CHECK-NEXT: [0,9]     . D===eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3743,43 +3738,42 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.2d, v2.2d }, [x27], #32
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     3.0    0.0    0.0       st1	{ v1.2s, v2.2s }, [x27], #16
-# CHECK-NEXT: 3.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     4.0    0.0    0.0       st1	{ v1.4h, v2.4h }, [x27], #16
-# CHECK-NEXT: 5.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     6.0    0.0    0.0       st1	{ v1.4s, v2.4s }, [x27], #32
-# CHECK-NEXT: 7.     1     8.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     7.0    0.0    0.0       st1	{ v1.8b, v2.8b }, [x27], #16
-# CHECK-NEXT: 9.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     5.2    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.2s, v2.2s }, [x27], #16
+# CHECK-NEXT: 3.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st1	{ v1.4h, v2.4h }, [x27], #16
+# CHECK-NEXT: 5.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st1	{ v1.4s, v2.4s }, [x27], #32
+# CHECK-NEXT: 7.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st1	{ v1.8b, v2.8b }, [x27], #16
+# CHECK-NEXT: 9.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    0.0       <total>
 
 # CHECK:      [56] Code Region - G57
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      504
 # CHECK-NEXT: Total uOps:        2600
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    2.59
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    5.16
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 4.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
-# CHECK-NEXT: Index     0123456789
+# CHECK-NEXT: Index     012345678
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.8h, v2.8h }, [x27], #32
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=eeER   .  .   st1	{ v1.16b, v2.16b }, [x27], #32
-# CHECK-NEXT: [0,3]     .D===eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===eeER .  .   st1	{ v1.1d, v2.1d }, [x27], x28
-# CHECK-NEXT: [0,5]     .D=====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     . D====eeER  .   st1	{ v1.2d, v2.2d }, [x27], x28
-# CHECK-NEXT: [0,7]     . D======eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======eeER.   st1	{ v1.2s, v2.2s }, [x27], x28
-# CHECK-NEXT: [0,9]     . D========eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.  .   st1	{ v1.8h, v2.8h }, [x27], #32
+# CHECK-NEXT: [0,1]     D=eER.  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeER  .   st1	{ v1.16b, v2.16b }, [x27], #32
+# CHECK-NEXT: [0,3]     .D=eER  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeER .   st1	{ v1.1d, v2.1d }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eER .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     . D=eeER.   st1	{ v1.2d, v2.2d }, [x27], x28
+# CHECK-NEXT: [0,7]     . D==eER.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeER   st1	{ v1.2s, v2.2s }, [x27], x28
+# CHECK-NEXT: [0,9]     . D===eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3789,43 +3783,42 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.8h, v2.8h }, [x27], #32
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.16b, v2.16b }, [x27], #32
-# CHECK-NEXT: 3.     1     4.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     4.0    0.0    0.0       st1	{ v1.1d, v2.1d }, [x27], x28
-# CHECK-NEXT: 5.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     5.0    0.0    0.0       st1	{ v1.2d, v2.2d }, [x27], x28
-# CHECK-NEXT: 7.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     7.0    0.0    0.0       st1	{ v1.2s, v2.2s }, [x27], x28
-# CHECK-NEXT: 9.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     4.8    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st1	{ v1.16b, v2.16b }, [x27], #32
+# CHECK-NEXT: 3.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st1	{ v1.1d, v2.1d }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st1	{ v1.2d, v2.2d }, [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st1	{ v1.2s, v2.2s }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.1    0.0       <total>
 
 # CHECK:      [57] Code Region - G58
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      504
 # CHECK-NEXT: Total uOps:        2600
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    2.59
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    5.16
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 4.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
-# CHECK-NEXT: Index     0123456789
+# CHECK-NEXT: Index     012345678
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.4h, v2.4h }, [x27], x28
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D==eeER   .  .   st1	{ v1.4s, v2.4s }, [x27], x28
-# CHECK-NEXT: [0,3]     D====eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D===eeER .  .   st1	{ v1.8b, v2.8b }, [x27], x28
-# CHECK-NEXT: [0,5]     .D=====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D=====eeER  .   st1	{ v1.8h, v2.8h }, [x27], x28
-# CHECK-NEXT: [0,7]     .D=======eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D======eeER.   st1	{ v1.16b, v2.16b }, [x27], x28
-# CHECK-NEXT: [0,9]     . D========eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.  .   st1	{ v1.4h, v2.4h }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eER.  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeER  .   st1	{ v1.4s, v2.4s }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eER  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeER .   st1	{ v1.8b, v2.8b }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eER .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeER.   st1	{ v1.8h, v2.8h }, [x27], x28
+# CHECK-NEXT: [0,7]     .D===eER.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeER   st1	{ v1.16b, v2.16b }, [x27], x28
+# CHECK-NEXT: [0,9]     . D===eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3835,43 +3828,42 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.4h, v2.4h }, [x27], x28
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     3.0    0.0    0.0       st1	{ v1.4s, v2.4s }, [x27], x28
-# CHECK-NEXT: 3.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     4.0    0.0    0.0       st1	{ v1.8b, v2.8b }, [x27], x28
-# CHECK-NEXT: 5.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     6.0    0.0    0.0       st1	{ v1.8h, v2.8h }, [x27], x28
-# CHECK-NEXT: 7.     1     8.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     7.0    0.0    0.0       st1	{ v1.16b, v2.16b }, [x27], x28
-# CHECK-NEXT: 9.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     5.2    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.4s, v2.4s }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st1	{ v1.8b, v2.8b }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st1	{ v1.8h, v2.8h }, [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st1	{ v1.16b, v2.16b }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    0.0       <total>
 
 # CHECK:      [58] Code Region - G59
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      703
 # CHECK-NEXT: Total uOps:        3400
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    3.39
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    4.84
+# CHECK-NEXT: IPC:               1.42
 # CHECK-NEXT: Block RThroughput: 6.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
 # CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.1d, v2.1d, v3.1d }, [x27], #24
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=eeER   .  .   st1	{ v1.2d, v2.2d, v3.2d }, [x27], #48
-# CHECK-NEXT: [0,3]     .D===eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==eeER .  .   st1	{ v1.2s, v2.2s, v3.2s }, [x27], #24
-# CHECK-NEXT: [0,5]     . D====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D===eeER  .   st1	{ v1.4h, v2.4h, v3.4h }, [x27], #24
-# CHECK-NEXT: [0,7]     .  D=====eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D====eeER.   st1	{ v1.4s, v2.4s, v3.4s }, [x27], #48
-# CHECK-NEXT: [0,9]     .   D======eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.   .   st1	{ v1.1d, v2.1d, v3.1d }, [x27], #24
+# CHECK-NEXT: [0,1]     D=eER.   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeER   .   st1	{ v1.2d, v2.2d, v3.2d }, [x27], #48
+# CHECK-NEXT: [0,3]     .D=eER   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . DeeER  .   st1	{ v1.2s, v2.2s, v3.2s }, [x27], #24
+# CHECK-NEXT: [0,5]     . D=eER  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeER.   st1	{ v1.4h, v2.4h, v3.4h }, [x27], #24
+# CHECK-NEXT: [0,7]     .  D==eER.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D=eeER   st1	{ v1.4s, v2.4s, v3.4s }, [x27], #48
+# CHECK-NEXT: [0,9]     .   D==eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3881,43 +3873,42 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.1d, v2.1d, v3.1d }, [x27], #24
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.2d, v2.2d, v3.2d }, [x27], #48
-# CHECK-NEXT: 3.     1     4.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     3.0    0.0    0.0       st1	{ v1.2s, v2.2s, v3.2s }, [x27], #24
-# CHECK-NEXT: 5.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     4.0    0.0    0.0       st1	{ v1.4h, v2.4h, v3.4h }, [x27], #24
-# CHECK-NEXT: 7.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     5.0    0.0    0.0       st1	{ v1.4s, v2.4s, v3.4s }, [x27], #48
-# CHECK-NEXT: 9.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     4.0    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st1	{ v1.2d, v2.2d, v3.2d }, [x27], #48
+# CHECK-NEXT: 3.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     1.0    0.0    0.0       st1	{ v1.2s, v2.2s, v3.2s }, [x27], #24
+# CHECK-NEXT: 5.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    1.0    0.0       st1	{ v1.4h, v2.4h, v3.4h }, [x27], #24
+# CHECK-NEXT: 7.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     2.0    0.0    0.0       st1	{ v1.4s, v2.4s, v3.4s }, [x27], #48
+# CHECK-NEXT: 9.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     1.9    0.2    0.0       <total>
 
 # CHECK:      [59] Code Region - G60
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      703
 # CHECK-NEXT: Total uOps:        3600
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    3.59
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    5.12
+# CHECK-NEXT: IPC:               1.42
 # CHECK-NEXT: Block RThroughput: 6.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
 # CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.8b, v2.8b, v3.8b }, [x27], #24
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=eeER   .  .   st1	{ v1.8h, v2.8h, v3.8h }, [x27], #48
-# CHECK-NEXT: [0,3]     .D===eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==eeER .  .   st1	{ v1.16b, v2.16b, v3.16b }, [x27], #48
-# CHECK-NEXT: [0,5]     . D====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D===eeER  .   st1	{ v1.1d, v2.1d, v3.1d }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D=====eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D====eeER.   st1	{ v1.2d, v2.2d, v3.2d }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D======eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.   .   st1	{ v1.8b, v2.8b, v3.8b }, [x27], #24
+# CHECK-NEXT: [0,1]     D=eER.   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeER   .   st1	{ v1.8h, v2.8h, v3.8h }, [x27], #48
+# CHECK-NEXT: [0,3]     .D=eER   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . DeeER  .   st1	{ v1.16b, v2.16b, v3.16b }, [x27], #48
+# CHECK-NEXT: [0,5]     . D=eER  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeER.   st1	{ v1.1d, v2.1d, v3.1d }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D==eER.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D=eeER   st1	{ v1.2d, v2.2d, v3.2d }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D==eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3927,43 +3918,42 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.8b, v2.8b, v3.8b }, [x27], #24
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.8h, v2.8h, v3.8h }, [x27], #48
-# CHECK-NEXT: 3.     1     4.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     3.0    0.0    0.0       st1	{ v1.16b, v2.16b, v3.16b }, [x27], #48
-# CHECK-NEXT: 5.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     4.0    0.0    0.0       st1	{ v1.1d, v2.1d, v3.1d }, [x27], x28
-# CHECK-NEXT: 7.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     5.0    0.0    0.0       st1	{ v1.2d, v2.2d, v3.2d }, [x27], x28
-# CHECK-NEXT: 9.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     4.0    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st1	{ v1.8h, v2.8h, v3.8h }, [x27], #48
+# CHECK-NEXT: 3.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     1.0    0.0    0.0       st1	{ v1.16b, v2.16b, v3.16b }, [x27], #48
+# CHECK-NEXT: 5.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    1.0    0.0       st1	{ v1.1d, v2.1d, v3.1d }, [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     2.0    0.0    0.0       st1	{ v1.2d, v2.2d, v3.2d }, [x27], x28
+# CHECK-NEXT: 9.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     1.9    0.2    0.0       <total>
 
 # CHECK:      [60] Code Region - G61
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      703
 # CHECK-NEXT: Total uOps:        3400
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    3.39
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    4.84
+# CHECK-NEXT: IPC:               1.42
 # CHECK-NEXT: Block RThroughput: 6.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
 # CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.2s, v2.2s, v3.2s }, [x27], x28
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=eeER   .  .   st1	{ v1.4h, v2.4h, v3.4h }, [x27], x28
-# CHECK-NEXT: [0,3]     .D===eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==eeER .  .   st1	{ v1.4s, v2.4s, v3.4s }, [x27], x28
-# CHECK-NEXT: [0,5]     . D====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D===eeER  .   st1	{ v1.8b, v2.8b, v3.8b }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D=====eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D====eeER.   st1	{ v1.8h, v2.8h, v3.8h }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D======eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.   .   st1	{ v1.2s, v2.2s, v3.2s }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eER.   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeER   .   st1	{ v1.4h, v2.4h, v3.4h }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eER   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . DeeER  .   st1	{ v1.4s, v2.4s, v3.4s }, [x27], x28
+# CHECK-NEXT: [0,5]     . D=eER  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  DeeER .   st1	{ v1.8b, v2.8b, v3.8b }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D=eER .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D=eeER   st1	{ v1.8h, v2.8h, v3.8h }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D==eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -3973,43 +3963,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.2s, v2.2s, v3.2s }, [x27], x28
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.4h, v2.4h, v3.4h }, [x27], x28
-# CHECK-NEXT: 3.     1     4.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     3.0    0.0    0.0       st1	{ v1.4s, v2.4s, v3.4s }, [x27], x28
-# CHECK-NEXT: 5.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     4.0    0.0    0.0       st1	{ v1.8b, v2.8b, v3.8b }, [x27], x28
-# CHECK-NEXT: 7.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     5.0    0.0    0.0       st1	{ v1.8h, v2.8h, v3.8h }, [x27], x28
-# CHECK-NEXT: 9.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     4.0    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st1	{ v1.4h, v2.4h, v3.4h }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     1.0    0.0    0.0       st1	{ v1.4s, v2.4s, v3.4s }, [x27], x28
+# CHECK-NEXT: 5.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     1.0    0.0    0.0       st1	{ v1.8b, v2.8b, v3.8b }, [x27], x28
+# CHECK-NEXT: 7.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     2.0    1.0    0.0       st1	{ v1.8h, v2.8h, v3.8h }, [x27], x28
+# CHECK-NEXT: 9.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     1.7    0.2    0.0       <total>
 
 # CHECK:      [61] Code Region - G62
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      704
 # CHECK-NEXT: Total uOps:        3600
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    3.59
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    5.11
+# CHECK-NEXT: IPC:               1.42
 # CHECK-NEXT: Block RThroughput: 6.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
+# CHECK-NEXT:                     0
 # CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.16b, v2.16b, v3.16b }, [x27], x28
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=eeER   .  .   st1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
-# CHECK-NEXT: [0,3]     .D===eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==eeER .  .   st1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
-# CHECK-NEXT: [0,5]     . D====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D===eeER  .   st1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
-# CHECK-NEXT: [0,7]     .  D=====eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D====eeER.   st1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
-# CHECK-NEXT: [0,9]     .   D======eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.    .   st1	{ v1.16b, v2.16b, v3.16b }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eER.    .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeER    .   st1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
+# CHECK-NEXT: [0,3]     .D=eER    .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeER  .   st1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
+# CHECK-NEXT: [0,5]     . D==eER  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeER .   st1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
+# CHECK-NEXT: [0,7]     .  D==eER .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eeER   st1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
+# CHECK-NEXT: [0,9]     .   D===eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4019,43 +4009,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.16b, v2.16b, v3.16b }, [x27], x28
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
-# CHECK-NEXT: 3.     1     4.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     3.0    0.0    0.0       st1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
-# CHECK-NEXT: 5.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     4.0    0.0    0.0       st1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
-# CHECK-NEXT: 7.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     5.0    0.0    0.0       st1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
-# CHECK-NEXT: 9.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     4.0    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], #32
+# CHECK-NEXT: 3.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       st1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
+# CHECK-NEXT: 5.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
+# CHECK-NEXT: 7.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       st1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
+# CHECK-NEXT: 9.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    0.0       <total>
 
 # CHECK:      [62] Code Region - G63
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      804
 # CHECK-NEXT: Total uOps:        4200
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    4.18
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    5.22
+# CHECK-NEXT: IPC:               1.24
 # CHECK-NEXT: Block RThroughput: 8.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
+# CHECK-NEXT:                     01
 # CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=eeER   .  .   st1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
-# CHECK-NEXT: [0,3]     .D===eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==eeER .  .   st1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
-# CHECK-NEXT: [0,5]     . D====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D===eeER  .   st1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
-# CHECK-NEXT: [0,7]     .  D=====eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D====eeER.   st1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D======eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.    ..   st1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
+# CHECK-NEXT: [0,1]     D=eER.    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeER    ..   st1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
+# CHECK-NEXT: [0,3]     .D=eER    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeER  ..   st1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
+# CHECK-NEXT: [0,5]     . D==eER  ..   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeER ..   st1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
+# CHECK-NEXT: [0,7]     .  D==eER ..   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D===eeER   st1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D====eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4065,43 +4055,42 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
-# CHECK-NEXT: 3.     1     4.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     3.0    0.0    0.0       st1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
-# CHECK-NEXT: 5.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     4.0    0.0    0.0       st1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
-# CHECK-NEXT: 7.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     5.0    0.0    0.0       st1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], x28
-# CHECK-NEXT: 9.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     4.0    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
+# CHECK-NEXT: 3.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       st1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
+# CHECK-NEXT: 5.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
+# CHECK-NEXT: 7.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     4.0    2.0    0.0       st1	{ v1.1d, v2.1d, v3.1d, v4.1d }, [x27], x28
+# CHECK-NEXT: 9.     1     5.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.5    0.4    0.0       <total>
 
 # CHECK:      [63] Code Region - G64
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1004
+# CHECK-NEXT: Total Cycles:      703
 # CHECK-NEXT: Total uOps:        3800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    3.78
-# CHECK-NEXT: IPC:               1.00
+# CHECK-NEXT: uOps Per Cycle:    5.41
+# CHECK-NEXT: IPC:               1.42
 # CHECK-NEXT: Block RThroughput: 7.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123
 # CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeER.    .  .   st1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
-# CHECK-NEXT: [0,1]     D==eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=eeER   .  .   st1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
-# CHECK-NEXT: [0,3]     .D===eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==eeER .  .   st1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
-# CHECK-NEXT: [0,5]     . D====eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D===eeER  .   st1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D=====eER .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D====eeER.   st1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D======eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.   .   st1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eER.   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeER   .   st1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eER   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeER .   st1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
+# CHECK-NEXT: [0,5]     . D==eER .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeER.   st1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D==eER.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D=eeER   st1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D==eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4111,43 +4100,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
-# CHECK-NEXT: 3.     1     4.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     3.0    0.0    0.0       st1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
-# CHECK-NEXT: 5.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     4.0    0.0    0.0       st1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
-# CHECK-NEXT: 7.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     5.0    0.0    0.0       st1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
-# CHECK-NEXT: 9.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     4.0    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st1	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       st1	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st1	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     2.0    0.0    0.0       st1	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
+# CHECK-NEXT: 9.     1     3.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.1    0.2    0.0       <total>
 
 # CHECK:      [64] Code Region - G65
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1604
+# CHECK-NEXT: Total Cycles:      706
 # CHECK-NEXT: Total uOps:        3200
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    2.00
-# CHECK-NEXT: IPC:               0.62
+# CHECK-NEXT: uOps Per Cycle:    4.53
+# CHECK-NEXT: IPC:               1.42
 # CHECK-NEXT: Block RThroughput: 5.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789
+# CHECK-NEXT:                     012
 # CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeER.    .    .   .   st1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
-# CHECK-NEXT: [0,1]     D==eER    .    .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=eeER   .    .   .   st1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
-# CHECK-NEXT: [0,3]     .D===eER  .    .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==eeeeER    .   .   st1	{ v1.b }[0], [x27], #1
-# CHECK-NEXT: [0,5]     . D======eER   .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     . D======eeeeER.   .   st1	{ v1.b }[8], [x27], #1
-# CHECK-NEXT: [0,7]     . D==========eER   .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .  D=========eeeeER.   st1	{ v1.b }[0], [x27], x28
-# CHECK-NEXT: [0,9]     .  D=============eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeER.    . .   st1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eER.    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeER    . .   st1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eER    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D==eeeeER .   st1	{ v1.b }[0], [x27], #1
+# CHECK-NEXT: [0,5]     . D===eE--R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     . D===eeeeER.   st1	{ v1.b }[8], [x27], #1
+# CHECK-NEXT: [0,7]     . D====eE--R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .  D===eeeeER   st1	{ v1.b }[0], [x27], x28
+# CHECK-NEXT: [0,9]     .  D====eE--R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4157,43 +4146,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
-# CHECK-NEXT: 1.     1     3.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
-# CHECK-NEXT: 3.     1     4.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     3.0    0.0    0.0       st1	{ v1.b }[0], [x27], #1
-# CHECK-NEXT: 5.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     7.0    0.0    0.0       st1	{ v1.b }[8], [x27], #1
-# CHECK-NEXT: 7.     1     11.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     10.0   0.0    0.0       st1	{ v1.b }[0], [x27], x28
-# CHECK-NEXT: 9.     1     14.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     6.2    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st1	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    2.0    0.0       st1	{ v1.b }[0], [x27], #1
+# CHECK-NEXT: 5.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     4.0    0.0    0.0       st1	{ v1.b }[8], [x27], #1
+# CHECK-NEXT: 7.     1     5.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     4.0    0.0    0.0       st1	{ v1.b }[0], [x27], x28
+# CHECK-NEXT: 9.     1     5.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT:        1     3.1    0.3    0.6       <total>
 
 # CHECK:      [65] Code Region - G66
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2004
+# CHECK-NEXT: Total Cycles:      506
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.00
-# CHECK-NEXT: IPC:               0.50
+# CHECK-NEXT: uOps Per Cycle:    3.95
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789
-# CHECK-NEXT: Index     0123456789          0123
+# CHECK-NEXT:                     0
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeER   .    .    .  .   st1	{ v1.b }[8], [x27], x28
-# CHECK-NEXT: [0,1]     D====eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D====eeeeER    .    .  .   st1	{ v1.h }[0], [x27], #2
-# CHECK-NEXT: [0,3]     D========eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D=======eeeeER.    .  .   st1	{ v1.h }[4], [x27], #2
-# CHECK-NEXT: [0,5]     .D===========eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D===========eeeeER .  .   st1	{ v1.h }[0], [x27], x28
-# CHECK-NEXT: [0,7]     .D===============eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============eeeeER.   st1	{ v1.h }[4], [x27], x28
-# CHECK-NEXT: [0,9]     . D==================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeER   .   st1	{ v1.b }[8], [x27], x28
+# CHECK-NEXT: [0,1]     D=eE--R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeER  .   st1	{ v1.h }[0], [x27], #2
+# CHECK-NEXT: [0,3]     D==eE--R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeER .   st1	{ v1.h }[4], [x27], #2
+# CHECK-NEXT: [0,5]     .D==eE--R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeER.   st1	{ v1.h }[0], [x27], x28
+# CHECK-NEXT: [0,7]     .D===eE--R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeER   st1	{ v1.h }[4], [x27], x28
+# CHECK-NEXT: [0,9]     . D===eE--R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4203,43 +4192,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.b }[8], [x27], x28
-# CHECK-NEXT: 1.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     5.0    0.0    0.0       st1	{ v1.h }[0], [x27], #2
-# CHECK-NEXT: 3.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     8.0    0.0    0.0       st1	{ v1.h }[4], [x27], #2
-# CHECK-NEXT: 5.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     12.0   0.0    0.0       st1	{ v1.h }[0], [x27], x28
-# CHECK-NEXT: 7.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     15.0   0.0    0.0       st1	{ v1.h }[4], [x27], x28
-# CHECK-NEXT: 9.     1     19.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     10.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.h }[0], [x27], #2
+# CHECK-NEXT: 3.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st1	{ v1.h }[4], [x27], #2
+# CHECK-NEXT: 5.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st1	{ v1.h }[0], [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st1	{ v1.h }[4], [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    1.0       <total>
 
 # CHECK:      [66] Code Region - G67
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2004
+# CHECK-NEXT: Total Cycles:      506
 # CHECK-NEXT: Total uOps:        2200
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.10
-# CHECK-NEXT: IPC:               0.50
+# CHECK-NEXT: uOps Per Cycle:    4.35
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 3.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789
-# CHECK-NEXT: Index     0123456789          0123
+# CHECK-NEXT:                     0
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeER   .    .    .  .   st1	{ v1.s }[0], [x27], #4
-# CHECK-NEXT: [0,1]     D====eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D====eeeeER    .    .  .   st1	{ v1.s }[0], [x27], x28
-# CHECK-NEXT: [0,3]     D========eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D=======eeeeER.    .  .   st1	{ v1.d }[0], [x27], #8
-# CHECK-NEXT: [0,5]     .D===========eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D===========eeeeER .  .   st1	{ v1.d }[0], [x27], x28
-# CHECK-NEXT: [0,7]     .D===============eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============eeeeER.   st2	{ v1.2d, v2.2d }, [x27], #32
-# CHECK-NEXT: [0,9]     . D==================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeER   .   st1	{ v1.s }[0], [x27], #4
+# CHECK-NEXT: [0,1]     D=eE--R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeER  .   st1	{ v1.s }[0], [x27], x28
+# CHECK-NEXT: [0,3]     D==eE--R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeER .   st1	{ v1.d }[0], [x27], #8
+# CHECK-NEXT: [0,5]     .D==eE--R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeER.   st1	{ v1.d }[0], [x27], x28
+# CHECK-NEXT: [0,7]     .D===eE--R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeER   st2	{ v1.2d, v2.2d }, [x27], #32
+# CHECK-NEXT: [0,9]     . D===eE--R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4249,43 +4238,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st1	{ v1.s }[0], [x27], #4
-# CHECK-NEXT: 1.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     5.0    0.0    0.0       st1	{ v1.s }[0], [x27], x28
-# CHECK-NEXT: 3.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     8.0    0.0    0.0       st1	{ v1.d }[0], [x27], #8
-# CHECK-NEXT: 5.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     12.0   0.0    0.0       st1	{ v1.d }[0], [x27], x28
-# CHECK-NEXT: 7.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     15.0   0.0    0.0       st2	{ v1.2d, v2.2d }, [x27], #32
-# CHECK-NEXT: 9.     1     19.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     10.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st1	{ v1.s }[0], [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st1	{ v1.d }[0], [x27], #8
+# CHECK-NEXT: 5.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st1	{ v1.d }[0], [x27], x28
+# CHECK-NEXT: 7.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st2	{ v1.2d, v2.2d }, [x27], #32
+# CHECK-NEXT: 9.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    1.0       <total>
 
 # CHECK:      [67] Code Region - G68
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2004
+# CHECK-NEXT: Total Cycles:      506
 # CHECK-NEXT: Total uOps:        2400
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.20
-# CHECK-NEXT: IPC:               0.50
+# CHECK-NEXT: uOps Per Cycle:    4.74
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 3.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789
-# CHECK-NEXT: Index     0123456789          0123
+# CHECK-NEXT:                     0
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeER   .    .    .  .   st2	{ v1.2s, v2.2s }, [x27], #16
-# CHECK-NEXT: [0,1]     D====eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D====eeeeER    .    .  .   st2	{ v1.4h, v2.4h }, [x27], #16
-# CHECK-NEXT: [0,3]     D========eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D=======eeeeER.    .  .   st2	{ v1.4s, v2.4s }, [x27], #32
-# CHECK-NEXT: [0,5]     .D===========eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D===========eeeeER .  .   st2	{ v1.8b, v2.8b }, [x27], #16
-# CHECK-NEXT: [0,7]     .D===============eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============eeeeER.   st2	{ v1.8h, v2.8h }, [x27], #32
-# CHECK-NEXT: [0,9]     . D==================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeER   .   st2	{ v1.2s, v2.2s }, [x27], #16
+# CHECK-NEXT: [0,1]     D=eE--R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeER  .   st2	{ v1.4h, v2.4h }, [x27], #16
+# CHECK-NEXT: [0,3]     D==eE--R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeER .   st2	{ v1.4s, v2.4s }, [x27], #32
+# CHECK-NEXT: [0,5]     .D==eE--R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeER.   st2	{ v1.8b, v2.8b }, [x27], #16
+# CHECK-NEXT: [0,7]     .D===eE--R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeER   st2	{ v1.8h, v2.8h }, [x27], #32
+# CHECK-NEXT: [0,9]     . D===eE--R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4295,43 +4284,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st2	{ v1.2s, v2.2s }, [x27], #16
-# CHECK-NEXT: 1.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     5.0    0.0    0.0       st2	{ v1.4h, v2.4h }, [x27], #16
-# CHECK-NEXT: 3.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     8.0    0.0    0.0       st2	{ v1.4s, v2.4s }, [x27], #32
-# CHECK-NEXT: 5.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     12.0   0.0    0.0       st2	{ v1.8b, v2.8b }, [x27], #16
-# CHECK-NEXT: 7.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     15.0   0.0    0.0       st2	{ v1.8h, v2.8h }, [x27], #32
-# CHECK-NEXT: 9.     1     19.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     10.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st2	{ v1.4h, v2.4h }, [x27], #16
+# CHECK-NEXT: 3.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st2	{ v1.4s, v2.4s }, [x27], #32
+# CHECK-NEXT: 5.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st2	{ v1.8b, v2.8b }, [x27], #16
+# CHECK-NEXT: 7.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st2	{ v1.8h, v2.8h }, [x27], #32
+# CHECK-NEXT: 9.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    1.0       <total>
 
 # CHECK:      [68] Code Region - G69
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2004
+# CHECK-NEXT: Total Cycles:      506
 # CHECK-NEXT: Total uOps:        2600
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.30
-# CHECK-NEXT: IPC:               0.50
+# CHECK-NEXT: uOps Per Cycle:    5.14
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 4.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789
-# CHECK-NEXT: Index     0123456789          0123
+# CHECK-NEXT:                     0
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeER   .    .    .  .   st2	{ v1.16b, v2.16b }, [x27], #32
-# CHECK-NEXT: [0,1]     D====eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D===eeeeER    .    .  .   st2	{ v1.2d, v2.2d }, [x27], x28
-# CHECK-NEXT: [0,3]     .D=======eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D=======eeeeER.    .  .   st2	{ v1.2s, v2.2s }, [x27], x28
-# CHECK-NEXT: [0,5]     .D===========eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     . D==========eeeeER .  .   st2	{ v1.4h, v2.4h }, [x27], x28
-# CHECK-NEXT: [0,7]     . D==============eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============eeeeER.   st2	{ v1.4s, v2.4s }, [x27], x28
-# CHECK-NEXT: [0,9]     . D==================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeER   .   st2	{ v1.16b, v2.16b }, [x27], #32
+# CHECK-NEXT: [0,1]     D=eE--R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeER  .   st2	{ v1.2d, v2.2d }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE--R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeER .   st2	{ v1.2s, v2.2s }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eE--R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     . D=eeeeER.   st2	{ v1.4h, v2.4h }, [x27], x28
+# CHECK-NEXT: [0,7]     . D==eE--R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeER   st2	{ v1.4s, v2.4s }, [x27], x28
+# CHECK-NEXT: [0,9]     . D===eE--R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4341,43 +4330,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st2	{ v1.16b, v2.16b }, [x27], #32
-# CHECK-NEXT: 1.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     4.0    0.0    0.0       st2	{ v1.2d, v2.2d }, [x27], x28
-# CHECK-NEXT: 3.     1     8.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     8.0    0.0    0.0       st2	{ v1.2s, v2.2s }, [x27], x28
-# CHECK-NEXT: 5.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     11.0   0.0    0.0       st2	{ v1.4h, v2.4h }, [x27], x28
-# CHECK-NEXT: 7.     1     15.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     15.0   0.0    0.0       st2	{ v1.4s, v2.4s }, [x27], x28
-# CHECK-NEXT: 9.     1     19.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     9.8    0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st2	{ v1.2d, v2.2d }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st2	{ v1.2s, v2.2s }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st2	{ v1.4h, v2.4h }, [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st2	{ v1.4s, v2.4s }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.1    1.0       <total>
 
 # CHECK:      [69] Code Region - G70
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2004
+# CHECK-NEXT: Total Cycles:      506
 # CHECK-NEXT: Total uOps:        2400
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.20
-# CHECK-NEXT: IPC:               0.50
+# CHECK-NEXT: uOps Per Cycle:    4.74
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 3.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789
-# CHECK-NEXT: Index     0123456789          0123
+# CHECK-NEXT:                     0
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeER   .    .    .  .   st2	{ v1.8b, v2.8b }, [x27], x28
-# CHECK-NEXT: [0,1]     D====eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D====eeeeER    .    .  .   st2	{ v1.8h, v2.8h }, [x27], x28
-# CHECK-NEXT: [0,3]     D========eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D=======eeeeER.    .  .   st2	{ v1.16b, v2.16b }, [x27], x28
-# CHECK-NEXT: [0,5]     .D===========eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D===========eeeeER .  .   st2	{ v1.b, v2.b }[0], [x27], #2
-# CHECK-NEXT: [0,7]     .D===============eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============eeeeER.   st2	{ v1.b, v2.b }[8], [x27], #2
-# CHECK-NEXT: [0,9]     . D==================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeER   .   st2	{ v1.8b, v2.8b }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE--R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeER  .   st2	{ v1.8h, v2.8h }, [x27], x28
+# CHECK-NEXT: [0,3]     D==eE--R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeER .   st2	{ v1.16b, v2.16b }, [x27], x28
+# CHECK-NEXT: [0,5]     .D==eE--R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeER.   st2	{ v1.b, v2.b }[0], [x27], #2
+# CHECK-NEXT: [0,7]     .D===eE--R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeER   st2	{ v1.b, v2.b }[8], [x27], #2
+# CHECK-NEXT: [0,9]     . D===eE--R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4387,43 +4376,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st2	{ v1.8b, v2.8b }, [x27], x28
-# CHECK-NEXT: 1.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     5.0    0.0    0.0       st2	{ v1.8h, v2.8h }, [x27], x28
-# CHECK-NEXT: 3.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     8.0    0.0    0.0       st2	{ v1.16b, v2.16b }, [x27], x28
-# CHECK-NEXT: 5.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     12.0   0.0    0.0       st2	{ v1.b, v2.b }[0], [x27], #2
-# CHECK-NEXT: 7.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     15.0   0.0    0.0       st2	{ v1.b, v2.b }[8], [x27], #2
-# CHECK-NEXT: 9.     1     19.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     10.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st2	{ v1.8h, v2.8h }, [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st2	{ v1.16b, v2.16b }, [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st2	{ v1.b, v2.b }[0], [x27], #2
+# CHECK-NEXT: 7.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st2	{ v1.b, v2.b }[8], [x27], #2
+# CHECK-NEXT: 9.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    1.0       <total>
 
 # CHECK:      [70] Code Region - G71
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2004
+# CHECK-NEXT: Total Cycles:      506
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.00
-# CHECK-NEXT: IPC:               0.50
+# CHECK-NEXT: uOps Per Cycle:    3.95
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789
-# CHECK-NEXT: Index     0123456789          0123
+# CHECK-NEXT:                     0
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeER   .    .    .  .   st2	{ v1.b, v2.b }[0], [x27], x28
-# CHECK-NEXT: [0,1]     D====eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D====eeeeER    .    .  .   st2	{ v1.b, v2.b }[8], [x27], x28
-# CHECK-NEXT: [0,3]     D========eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D=======eeeeER.    .  .   st2	{ v1.h, v2.h }[0], [x27], #4
-# CHECK-NEXT: [0,5]     .D===========eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D===========eeeeER .  .   st2	{ v1.h, v2.h }[4], [x27], #4
-# CHECK-NEXT: [0,7]     .D===============eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============eeeeER.   st2	{ v1.h, v2.h }[0], [x27], x28
-# CHECK-NEXT: [0,9]     . D==================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeER   .   st2	{ v1.b, v2.b }[0], [x27], x28
+# CHECK-NEXT: [0,1]     D=eE--R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeER  .   st2	{ v1.b, v2.b }[8], [x27], x28
+# CHECK-NEXT: [0,3]     D==eE--R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeER .   st2	{ v1.h, v2.h }[0], [x27], #4
+# CHECK-NEXT: [0,5]     .D==eE--R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeER.   st2	{ v1.h, v2.h }[4], [x27], #4
+# CHECK-NEXT: [0,7]     .D===eE--R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeER   st2	{ v1.h, v2.h }[0], [x27], x28
+# CHECK-NEXT: [0,9]     . D===eE--R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4433,43 +4422,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st2	{ v1.b, v2.b }[0], [x27], x28
-# CHECK-NEXT: 1.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     5.0    0.0    0.0       st2	{ v1.b, v2.b }[8], [x27], x28
-# CHECK-NEXT: 3.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     8.0    0.0    0.0       st2	{ v1.h, v2.h }[0], [x27], #4
-# CHECK-NEXT: 5.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     12.0   0.0    0.0       st2	{ v1.h, v2.h }[4], [x27], #4
-# CHECK-NEXT: 7.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     15.0   0.0    0.0       st2	{ v1.h, v2.h }[0], [x27], x28
-# CHECK-NEXT: 9.     1     19.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     10.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st2	{ v1.b, v2.b }[8], [x27], x28
+# CHECK-NEXT: 3.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st2	{ v1.h, v2.h }[0], [x27], #4
+# CHECK-NEXT: 5.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st2	{ v1.h, v2.h }[4], [x27], #4
+# CHECK-NEXT: 7.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st2	{ v1.h, v2.h }[0], [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    1.0       <total>
 
 # CHECK:      [71] Code Region - G72
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2004
+# CHECK-NEXT: Total Cycles:      506
 # CHECK-NEXT: Total uOps:        2000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.00
-# CHECK-NEXT: IPC:               0.50
+# CHECK-NEXT: uOps Per Cycle:    3.95
+# CHECK-NEXT: IPC:               1.98
 # CHECK-NEXT: Block RThroughput: 2.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789
-# CHECK-NEXT: Index     0123456789          0123
+# CHECK-NEXT:                     0
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeER   .    .    .  .   st2	{ v1.h, v2.h }[4], [x27], x28
-# CHECK-NEXT: [0,1]     D====eER  .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D====eeeeER    .    .  .   st2	{ v1.s, v2.s }[0], [x27], #8
-# CHECK-NEXT: [0,3]     D========eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D=======eeeeER.    .  .   st2	{ v1.s, v2.s }[0], [x27], x28
-# CHECK-NEXT: [0,5]     .D===========eER    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .D===========eeeeER .  .   st2	{ v1.d, v2.d }[0], [x27], #16
-# CHECK-NEXT: [0,7]     .D===============eER.  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     . D==============eeeeER.   st2	{ v1.d, v2.d }[0], [x27], x28
-# CHECK-NEXT: [0,9]     . D==================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeER   .   st2	{ v1.h, v2.h }[4], [x27], x28
+# CHECK-NEXT: [0,1]     D=eE--R   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eeeeER  .   st2	{ v1.s, v2.s }[0], [x27], #8
+# CHECK-NEXT: [0,3]     D==eE--R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeER .   st2	{ v1.s, v2.s }[0], [x27], x28
+# CHECK-NEXT: [0,5]     .D==eE--R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .D==eeeeER.   st2	{ v1.d, v2.d }[0], [x27], #16
+# CHECK-NEXT: [0,7]     .D===eE--R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     . D==eeeeER   st2	{ v1.d, v2.d }[0], [x27], x28
+# CHECK-NEXT: [0,9]     . D===eE--R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4479,43 +4468,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st2	{ v1.h, v2.h }[4], [x27], x28
-# CHECK-NEXT: 1.     1     5.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     5.0    0.0    0.0       st2	{ v1.s, v2.s }[0], [x27], #8
-# CHECK-NEXT: 3.     1     9.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     8.0    0.0    0.0       st2	{ v1.s, v2.s }[0], [x27], x28
-# CHECK-NEXT: 5.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     12.0   0.0    0.0       st2	{ v1.d, v2.d }[0], [x27], #16
-# CHECK-NEXT: 7.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     15.0   0.0    0.0       st2	{ v1.d, v2.d }[0], [x27], x28
-# CHECK-NEXT: 9.     1     19.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     10.2   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     2.0    0.0    0.0       st2	{ v1.s, v2.s }[0], [x27], #8
+# CHECK-NEXT: 3.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    0.0    0.0       st2	{ v1.s, v2.s }[0], [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st2	{ v1.d, v2.d }[0], [x27], #16
+# CHECK-NEXT: 7.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    0.0    0.0       st2	{ v1.d, v2.d }[0], [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.7    0.1    1.0       <total>
 
 # CHECK:      [72] Code Region - G73
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      1804
+# CHECK-NEXT: Total Cycles:      607
 # CHECK-NEXT: Total uOps:        2800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.55
-# CHECK-NEXT: IPC:               0.55
+# CHECK-NEXT: uOps Per Cycle:    4.61
+# CHECK-NEXT: IPC:               1.65
 # CHECK-NEXT: Block RThroughput: 4.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789
-# CHECK-NEXT: Index     0123456789          01
+# CHECK-NEXT:                     012
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeER .    .    .    ..   st2g	x26, [x27], #4064
-# CHECK-NEXT: [0,1]     D=eER.    .    .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     D=eER.    .    .    ..   st2g	x26, [x27, #4064]!
-# CHECK-NEXT: [0,3]     D==eER    .    .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .D=eeeeeeER    .    ..   st3	{ v1.2d, v2.2d, v3.2d }, [x27], #48
-# CHECK-NEXT: [0,5]     .D=======eER   .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     . D======eeeeeER    ..   st3	{ v1.2s, v2.2s, v3.2s }, [x27], #24
-# CHECK-NEXT: [0,7]     . D===========eER   ..   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .  D==========eeeeeER.   st3	{ v1.4h, v2.4h, v3.4h }, [x27], #24
-# CHECK-NEXT: [0,9]     .  D===============eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeER .    . .   st2g	x26, [x27], #4064
+# CHECK-NEXT: [0,1]     D=eER.    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     D=eER.    . .   st2g	x26, [x27, #4064]!
+# CHECK-NEXT: [0,3]     D==eER    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .D=eeeeeeER .   st3	{ v1.2d, v2.2d, v3.2d }, [x27], #48
+# CHECK-NEXT: [0,5]     .D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     . D=eeeeeER .   st3	{ v1.2s, v2.2s, v3.2s }, [x27], #24
+# CHECK-NEXT: [0,7]     . D==eE---R .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .  D==eeeeeER   st3	{ v1.4h, v2.4h, v3.4h }, [x27], #24
+# CHECK-NEXT: [0,9]     .  D===eE---R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4529,39 +4518,39 @@ add x0, x27, 1
 # CHECK-NEXT: 2.     1     2.0    0.0    0.0       st2g	x26, [x27, #4064]!
 # CHECK-NEXT: 3.     1     3.0    0.0    0.0       add	x0, x27, #1
 # CHECK-NEXT: 4.     1     2.0    0.0    0.0       st3	{ v1.2d, v2.2d, v3.2d }, [x27], #48
-# CHECK-NEXT: 5.     1     8.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     7.0    0.0    0.0       st3	{ v1.2s, v2.2s, v3.2s }, [x27], #24
-# CHECK-NEXT: 7.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     11.0   0.0    0.0       st3	{ v1.4h, v2.4h, v3.4h }, [x27], #24
-# CHECK-NEXT: 9.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     6.4    0.1    0.0       <total>
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st3	{ v1.2s, v2.2s, v3.2s }, [x27], #24
+# CHECK-NEXT: 7.     1     3.0    0.0    3.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       st3	{ v1.4h, v2.4h, v3.4h }, [x27], #24
+# CHECK-NEXT: 9.     1     4.0    0.0    3.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.5    0.2    1.0       <total>
 
 # CHECK:      [73] Code Region - G74
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2904
+# CHECK-NEXT: Total Cycles:      708
 # CHECK-NEXT: Total uOps:        3800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.31
-# CHECK-NEXT: IPC:               0.34
+# CHECK-NEXT: uOps Per Cycle:    5.37
+# CHECK-NEXT: IPC:               1.41
 # CHECK-NEXT: Block RThroughput: 7.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          012
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    . .   st3	{ v1.4s, v2.4s, v3.4s }, [x27], #48
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    . .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=====eeeeeER .    .    .    . .   st3	{ v1.8b, v2.8b, v3.8b }, [x27], #24
-# CHECK-NEXT: [0,3]     .D==========eER.    .    .    . .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D=========eeeeeeER.    .    . .   st3	{ v1.8h, v2.8h, v3.8h }, [x27], #48
-# CHECK-NEXT: [0,5]     . D===============eER    .    . .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D==============eeeeeeER    . .   st3	{ v1.16b, v2.16b, v3.16b }, [x27], #48
-# CHECK-NEXT: [0,7]     .  D====================eER   . .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D===================eeeeeeER.   st3	{ v1.2d, v2.2d, v3.2d }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D=========================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER .   .   st3	{ v1.4s, v2.4s, v3.4s }, [x27], #48
+# CHECK-NEXT: [0,1]     D=eE----R .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeER .   .   st3	{ v1.8b, v2.8b, v3.8b }, [x27], #24
+# CHECK-NEXT: [0,3]     .D=eE---R .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeER  .   st3	{ v1.8h, v2.8h, v3.8h }, [x27], #48
+# CHECK-NEXT: [0,5]     . D==eE----R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeER .   st3	{ v1.16b, v2.16b, v3.16b }, [x27], #48
+# CHECK-NEXT: [0,7]     .  D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eeeeeeER   st3	{ v1.2d, v2.2d, v3.2d }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4571,43 +4560,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st3	{ v1.4s, v2.4s, v3.4s }, [x27], #48
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     6.0    0.0    0.0       st3	{ v1.8b, v2.8b, v3.8b }, [x27], #24
-# CHECK-NEXT: 3.     1     11.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     10.0   0.0    0.0       st3	{ v1.8h, v2.8h, v3.8h }, [x27], #48
-# CHECK-NEXT: 5.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     15.0   0.0    0.0       st3	{ v1.16b, v2.16b, v3.16b }, [x27], #48
-# CHECK-NEXT: 7.     1     21.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     20.0   0.0    0.0       st3	{ v1.2d, v2.2d, v3.2d }, [x27], x28
-# CHECK-NEXT: 9.     1     26.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     13.3   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st3	{ v1.8b, v2.8b, v3.8b }, [x27], #24
+# CHECK-NEXT: 3.     1     2.0    0.0    3.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       st3	{ v1.8h, v2.8h, v3.8h }, [x27], #48
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st3	{ v1.16b, v2.16b, v3.16b }, [x27], #48
+# CHECK-NEXT: 7.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       st3	{ v1.2d, v2.2d, v3.2d }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    1.9       <total>
 
 # CHECK:      [74] Code Region - G75
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2704
+# CHECK-NEXT: Total Cycles:      707
 # CHECK-NEXT: Total uOps:        3400
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.26
-# CHECK-NEXT: IPC:               0.37
+# CHECK-NEXT: uOps Per Cycle:    4.81
+# CHECK-NEXT: IPC:               1.41
 # CHECK-NEXT: Block RThroughput: 6.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     0123
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeER  .    .    .    .    .   st3	{ v1.2s, v2.2s, v3.2s }, [x27], x28
-# CHECK-NEXT: [0,1]     D=====eER .    .    .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D====eeeeeER  .    .    .    .   st3	{ v1.4h, v2.4h, v3.4h }, [x27], x28
-# CHECK-NEXT: [0,3]     .D=========eER .    .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D========eeeeeeER .    .    .   st3	{ v1.4s, v2.4s, v3.4s }, [x27], x28
-# CHECK-NEXT: [0,5]     . D==============eER.    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=============eeeeeER .    .   st3	{ v1.8b, v2.8b, v3.8b }, [x27], x28
-# CHECK-NEXT: [0,7]     .  D==================eER.    .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D=================eeeeeeER.   st3	{ v1.8h, v2.8h, v3.8h }, [x27], x28
-# CHECK-NEXT: [0,9]     .   D=======================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeER  .  .   st3	{ v1.2s, v2.2s, v3.2s }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE---R  .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeER .  .   st3	{ v1.4h, v2.4h, v3.4h }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE---R .  .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . DeeeeeeER  .   st3	{ v1.4s, v2.4s, v3.4s }, [x27], x28
+# CHECK-NEXT: [0,5]     . D=eE----R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  DeeeeeER  .   st3	{ v1.8b, v2.8b, v3.8b }, [x27], x28
+# CHECK-NEXT: [0,7]     .  D=eE---R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D=eeeeeeER   st3	{ v1.8h, v2.8h, v3.8h }, [x27], x28
+# CHECK-NEXT: [0,9]     .   D==eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4617,43 +4606,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st3	{ v1.2s, v2.2s, v3.2s }, [x27], x28
-# CHECK-NEXT: 1.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     5.0    0.0    0.0       st3	{ v1.4h, v2.4h, v3.4h }, [x27], x28
-# CHECK-NEXT: 3.     1     10.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     9.0    0.0    0.0       st3	{ v1.4s, v2.4s, v3.4s }, [x27], x28
-# CHECK-NEXT: 5.     1     15.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     14.0   0.0    0.0       st3	{ v1.8b, v2.8b, v3.8b }, [x27], x28
-# CHECK-NEXT: 7.     1     19.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     18.0   0.0    0.0       st3	{ v1.8h, v2.8h, v3.8h }, [x27], x28
-# CHECK-NEXT: 9.     1     24.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     12.1   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    3.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st3	{ v1.4h, v2.4h, v3.4h }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    3.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     1.0    0.0    0.0       st3	{ v1.4s, v2.4s, v3.4s }, [x27], x28
+# CHECK-NEXT: 5.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     1.0    0.0    0.0       st3	{ v1.8b, v2.8b, v3.8b }, [x27], x28
+# CHECK-NEXT: 7.     1     2.0    0.0    3.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     2.0    1.0    0.0       st3	{ v1.8h, v2.8h, v3.8h }, [x27], x28
+# CHECK-NEXT: 9.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     1.7    0.2    1.7       <total>
 
 # CHECK:      [75] Code Region - G76
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      757
 # CHECK-NEXT: Total uOps:        4000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.33
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    5.28
+# CHECK-NEXT: IPC:               1.32
 # CHECK-NEXT: Block RThroughput: 7.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   st3	{ v1.16b, v2.16b, v3.16b }, [x27], x28
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=====eeeeeeER.    .    .    .  .   st3	{ v1.b, v2.b, v3.b }[0], [x27], #3
-# CHECK-NEXT: [0,3]     .D===========eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==========eeeeeeER    .    .  .   st3	{ v1.b, v2.b, v3.b }[8], [x27], #3
-# CHECK-NEXT: [0,5]     . D================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D===============eeeeeeER   .  .   st3	{ v1.b, v2.b, v3.b }[0], [x27], x28
-# CHECK-NEXT: [0,7]     .  D=====================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D====================eeeeeeER.   st3	{ v1.b, v2.b, v3.b }[8], [x27], x28
-# CHECK-NEXT: [0,9]     .   D==========================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER .   .   st3	{ v1.16b, v2.16b, v3.16b }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE----R .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeER.   .   st3	{ v1.b, v2.b, v3.b }[0], [x27], #3
+# CHECK-NEXT: [0,3]     .D=eE----R.   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeER  .   st3	{ v1.b, v2.b, v3.b }[8], [x27], #3
+# CHECK-NEXT: [0,5]     . D==eE----R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeER .   st3	{ v1.b, v2.b, v3.b }[0], [x27], x28
+# CHECK-NEXT: [0,7]     .  D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eeeeeeER   st3	{ v1.b, v2.b, v3.b }[8], [x27], x28
+# CHECK-NEXT: [0,9]     .   D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4663,43 +4652,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st3	{ v1.16b, v2.16b, v3.16b }, [x27], x28
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     6.0    0.0    0.0       st3	{ v1.b, v2.b, v3.b }[0], [x27], #3
-# CHECK-NEXT: 3.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     11.0   0.0    0.0       st3	{ v1.b, v2.b, v3.b }[8], [x27], #3
-# CHECK-NEXT: 5.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     16.0   0.0    0.0       st3	{ v1.b, v2.b, v3.b }[0], [x27], x28
-# CHECK-NEXT: 7.     1     22.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     21.0   0.0    0.0       st3	{ v1.b, v2.b, v3.b }[8], [x27], x28
-# CHECK-NEXT: 9.     1     27.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     14.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st3	{ v1.b, v2.b, v3.b }[0], [x27], #3
+# CHECK-NEXT: 3.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       st3	{ v1.b, v2.b, v3.b }[8], [x27], #3
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st3	{ v1.b, v2.b, v3.b }[0], [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       st3	{ v1.b, v2.b, v3.b }[8], [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    2.0       <total>
 
 # CHECK:      [76] Code Region - G77
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      757
 # CHECK-NEXT: Total uOps:        4000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.33
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    5.28
+# CHECK-NEXT: IPC:               1.32
 # CHECK-NEXT: Block RThroughput: 7.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   st3	{ v1.h, v2.h, v3.h }[0], [x27], #6
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=====eeeeeeER.    .    .    .  .   st3	{ v1.h, v2.h, v3.h }[4], [x27], #6
-# CHECK-NEXT: [0,3]     .D===========eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==========eeeeeeER    .    .  .   st3	{ v1.h, v2.h, v3.h }[0], [x27], x28
-# CHECK-NEXT: [0,5]     . D================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D===============eeeeeeER   .  .   st3	{ v1.h, v2.h, v3.h }[4], [x27], x28
-# CHECK-NEXT: [0,7]     .  D=====================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D====================eeeeeeER.   st3	{ v1.s, v2.s, v3.s }[0], [x27], #12
-# CHECK-NEXT: [0,9]     .   D==========================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER .   .   st3	{ v1.h, v2.h, v3.h }[0], [x27], #6
+# CHECK-NEXT: [0,1]     D=eE----R .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeER.   .   st3	{ v1.h, v2.h, v3.h }[4], [x27], #6
+# CHECK-NEXT: [0,3]     .D=eE----R.   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeER  .   st3	{ v1.h, v2.h, v3.h }[0], [x27], x28
+# CHECK-NEXT: [0,5]     . D==eE----R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeER .   st3	{ v1.h, v2.h, v3.h }[4], [x27], x28
+# CHECK-NEXT: [0,7]     .  D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eeeeeeER   st3	{ v1.s, v2.s, v3.s }[0], [x27], #12
+# CHECK-NEXT: [0,9]     .   D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4709,43 +4698,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st3	{ v1.h, v2.h, v3.h }[0], [x27], #6
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     6.0    0.0    0.0       st3	{ v1.h, v2.h, v3.h }[4], [x27], #6
-# CHECK-NEXT: 3.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     11.0   0.0    0.0       st3	{ v1.h, v2.h, v3.h }[0], [x27], x28
-# CHECK-NEXT: 5.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     16.0   0.0    0.0       st3	{ v1.h, v2.h, v3.h }[4], [x27], x28
-# CHECK-NEXT: 7.     1     22.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     21.0   0.0    0.0       st3	{ v1.s, v2.s, v3.s }[0], [x27], #12
-# CHECK-NEXT: 9.     1     27.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     14.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st3	{ v1.h, v2.h, v3.h }[4], [x27], #6
+# CHECK-NEXT: 3.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       st3	{ v1.h, v2.h, v3.h }[0], [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st3	{ v1.h, v2.h, v3.h }[4], [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       st3	{ v1.s, v2.s, v3.s }[0], [x27], #12
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    2.0       <total>
 
 # CHECK:      [77] Code Region - G78
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2904
+# CHECK-NEXT: Total Cycles:      807
 # CHECK-NEXT: Total uOps:        4200
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.45
-# CHECK-NEXT: IPC:               0.34
+# CHECK-NEXT: uOps Per Cycle:    5.20
+# CHECK-NEXT: IPC:               1.24
 # CHECK-NEXT: Block RThroughput: 8.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          012
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    . .   st3	{ v1.s, v2.s, v3.s }[0], [x27], x28
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    . .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=====eeeeeeER.    .    .    . .   st3	{ v1.d, v2.d, v3.d }[0], [x27], #24
-# CHECK-NEXT: [0,3]     .D===========eER    .    .    . .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==========eeeeeeER    .    . .   st3	{ v1.d, v2.d, v3.d }[0], [x27], x28
-# CHECK-NEXT: [0,5]     . D================eER   .    . .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D===============eeeeeER    . .   st4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
-# CHECK-NEXT: [0,7]     .  D====================eER   . .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D===================eeeeeeER.   st4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
-# CHECK-NEXT: [0,9]     .   D=========================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER .   .   st3	{ v1.s, v2.s, v3.s }[0], [x27], x28
+# CHECK-NEXT: [0,1]     D=eE----R .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeER.   .   st3	{ v1.d, v2.d, v3.d }[0], [x27], #24
+# CHECK-NEXT: [0,3]     .D=eE----R.   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeER  .   st3	{ v1.d, v2.d, v3.d }[0], [x27], x28
+# CHECK-NEXT: [0,5]     . D==eE----R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeER  .   st4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
+# CHECK-NEXT: [0,7]     .  D==eE---R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eeeeeeER   st4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
+# CHECK-NEXT: [0,9]     .   D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4755,43 +4744,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st3	{ v1.s, v2.s, v3.s }[0], [x27], x28
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     6.0    0.0    0.0       st3	{ v1.d, v2.d, v3.d }[0], [x27], #24
-# CHECK-NEXT: 3.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     11.0   0.0    0.0       st3	{ v1.d, v2.d, v3.d }[0], [x27], x28
-# CHECK-NEXT: 5.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     16.0   0.0    0.0       st4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
-# CHECK-NEXT: 7.     1     21.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     20.0   0.0    0.0       st4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
-# CHECK-NEXT: 9.     1     26.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     13.7   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st3	{ v1.d, v2.d, v3.d }[0], [x27], #24
+# CHECK-NEXT: 3.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       st3	{ v1.d, v2.d, v3.d }[0], [x27], x28
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], #64
+# CHECK-NEXT: 7.     1     3.0    0.0    3.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       st4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], #32
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    1.9       <total>
 
 # CHECK:      [78] Code Region - G79
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3304
+# CHECK-NEXT: Total Cycles:      1205
 # CHECK-NEXT: Total uOps:        5800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.76
-# CHECK-NEXT: IPC:               0.30
+# CHECK-NEXT: uOps Per Cycle:    4.81
+# CHECK-NEXT: IPC:               0.83
 # CHECK-NEXT: Block RThroughput: 12.0
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123456
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     0123456
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .    ..   st4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=====eeeeeeeER    .    .    .    ..   st4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
-# CHECK-NEXT: [0,3]     . D===========eER   .    .    .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .  D==========eeeeeeER   .    .    ..   st4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
-# CHECK-NEXT: [0,5]     .  D================eER  .    .    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .   D===============eeeeeeeER .    ..   st4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
-# CHECK-NEXT: [0,7]     .    D=====================eER.    ..   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .    .D====================eeeeeeeER.   st4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
-# CHECK-NEXT: [0,9]     .    . D==========================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER .    ..   st4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
+# CHECK-NEXT: [0,1]     D=eE----R .    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeeER    ..   st4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
+# CHECK-NEXT: [0,3]     . DeE-----R    ..   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .  DeeeeeeER   ..   st4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
+# CHECK-NEXT: [0,5]     .  D=eE----R   ..   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .   D==eeeeeeeER.   st4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
+# CHECK-NEXT: [0,7]     .    D==eE-----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .    .D=eeeeeeeER   st4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
+# CHECK-NEXT: [0,9]     .    . D=eE-----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4801,43 +4790,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], #32
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     6.0    0.0    0.0       st4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
-# CHECK-NEXT: 3.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     11.0   0.0    0.0       st4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
-# CHECK-NEXT: 5.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     16.0   0.0    0.0       st4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
-# CHECK-NEXT: 7.     1     22.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     21.0   0.0    0.0       st4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
-# CHECK-NEXT: 9.     1     27.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     14.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], #64
+# CHECK-NEXT: 3.     1     1.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     1.0    1.0    0.0       st4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], #32
+# CHECK-NEXT: 5.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    2.0    0.0       st4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], #64
+# CHECK-NEXT: 7.     1     3.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     2.0    0.0    0.0       st4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], #64
+# CHECK-NEXT: 9.     1     2.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT:        1     1.8    0.4    2.3       <total>
 
 # CHECK:      [79] Code Region - G80
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      1006
 # CHECK-NEXT: Total uOps:        4800
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.60
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    4.77
+# CHECK-NEXT: IPC:               0.99
 # CHECK-NEXT: Block RThroughput: 9.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     012345
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeER  .    .    .    .    .  .   st4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
-# CHECK-NEXT: [0,1]     D=====eER .    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D====eeeeeeER .    .    .    .  .   st4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
-# CHECK-NEXT: [0,3]     .D==========eER.    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D=========eeeeeeER.    .    .  .   st4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
-# CHECK-NEXT: [0,5]     . D===============eER    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D==============eeeeeeeER   .  .   st4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
-# CHECK-NEXT: [0,7]     .   D====================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .    D===================eeeeeeER.   st4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
-# CHECK-NEXT: [0,9]     .    D=========================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeER  .    .   st4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
+# CHECK-NEXT: [0,1]     D=eE---R  .    .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeER.    .   st4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE----R.    .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D==eeeeeeER  .   st4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
+# CHECK-NEXT: [0,5]     . D===eE----R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D==eeeeeeeER.   st4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
+# CHECK-NEXT: [0,7]     .   D==eE-----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .    D==eeeeeeER   st4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
+# CHECK-NEXT: [0,9]     .    D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4847,43 +4836,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st4	{ v1.2d, v2.2d, v3.2d, v4.2d }, [x27], x28
-# CHECK-NEXT: 1.     1     6.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     5.0    0.0    0.0       st4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
-# CHECK-NEXT: 3.     1     11.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     10.0   0.0    0.0       st4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
-# CHECK-NEXT: 5.     1     16.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     15.0   0.0    0.0       st4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
-# CHECK-NEXT: 7.     1     21.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     20.0   0.0    0.0       st4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
-# CHECK-NEXT: 9.     1     26.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     13.1   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    3.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st4	{ v1.2s, v2.2s, v3.2s, v4.2s }, [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    2.0    0.0       st4	{ v1.4h, v2.4h, v3.4h, v4.4h }, [x27], x28
+# CHECK-NEXT: 5.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     3.0    0.0    0.0       st4	{ v1.4s, v2.4s, v3.4s, v4.4s }, [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       st4	{ v1.8b, v2.8b, v3.8b, v4.8b }, [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.6    0.4    2.0       <total>
 
 # CHECK:      [80] Code Region - G81
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3204
+# CHECK-NEXT: Total Cycles:      1058
 # CHECK-NEXT: Total uOps:        5200
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.62
-# CHECK-NEXT: IPC:               0.31
+# CHECK-NEXT: uOps Per Cycle:    4.91
+# CHECK-NEXT: IPC:               0.95
 # CHECK-NEXT: Block RThroughput: 10.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          012345
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     01234567
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeeER.    .    .    .    .    .   st4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
-# CHECK-NEXT: [0,1]     .D======eER    .    .    .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     . D=====eeeeeeeER   .    .    .    .   st4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
-# CHECK-NEXT: [0,3]     .  D===========eER  .    .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     .   D==========eeeeeeER  .    .    .   st4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], #4
-# CHECK-NEXT: [0,5]     .   D================eER .    .    .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .    D===============eeeeeeER .    .   st4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], #4
-# CHECK-NEXT: [0,7]     .    D=====================eER.    .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .    .D====================eeeeeeER.   st4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], x28
-# CHECK-NEXT: [0,9]     .    .D==========================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeeER.    . .   st4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
+# CHECK-NEXT: [0,1]     .DeE-----R.    . .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     . DeeeeeeeER   . .   st4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
+# CHECK-NEXT: [0,3]     .  DeE-----R   . .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     .   D==eeeeeeER. .   st4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], #4
+# CHECK-NEXT: [0,5]     .   D===eE----R. .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .    D===eeeeeeER.   st4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], #4
+# CHECK-NEXT: [0,7]     .    D====eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .    .D===eeeeeeER   st4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], x28
+# CHECK-NEXT: [0,9]     .    .D====eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4893,43 +4882,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st4	{ v1.8h, v2.8h, v3.8h, v4.8h }, [x27], x28
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     6.0    0.0    0.0       st4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
-# CHECK-NEXT: 3.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     11.0   0.0    0.0       st4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], #4
-# CHECK-NEXT: 5.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     16.0   0.0    0.0       st4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], #4
-# CHECK-NEXT: 7.     1     22.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     21.0   0.0    0.0       st4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], x28
-# CHECK-NEXT: 9.     1     27.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     14.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     1.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    1.0    0.0       st4	{ v1.16b, v2.16b, v3.16b, v4.16b }, [x27], x28
+# CHECK-NEXT: 3.     1     1.0    0.0    5.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     3.0    3.0    0.0       st4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], #4
+# CHECK-NEXT: 5.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     4.0    1.0    0.0       st4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], #4
+# CHECK-NEXT: 7.     1     5.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     4.0    0.0    0.0       st4	{ v1.b, v2.b, v3.b, v4.b }[0], [x27], x28
+# CHECK-NEXT: 9.     1     5.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.9    0.6    2.2       <total>
 
 # CHECK:      [81] Code Region - G82
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      3004
+# CHECK-NEXT: Total Cycles:      757
 # CHECK-NEXT: Total uOps:        4000
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.33
-# CHECK-NEXT: IPC:               0.33
+# CHECK-NEXT: uOps Per Cycle:    5.28
+# CHECK-NEXT: IPC:               1.32
 # CHECK-NEXT: Block RThroughput: 7.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789          0123
-# CHECK-NEXT: Index     0123456789          0123456789
+# CHECK-NEXT:                     01234
+# CHECK-NEXT: Index     0123456789
 
-# CHECK:      [0,0]     DeeeeeeER .    .    .    .    .  .   st4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], x28
-# CHECK-NEXT: [0,1]     D======eER.    .    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=====eeeeeeER.    .    .    .  .   st4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], #8
-# CHECK-NEXT: [0,3]     .D===========eER    .    .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==========eeeeeeER    .    .  .   st4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], #8
-# CHECK-NEXT: [0,5]     . D================eER   .    .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D===============eeeeeeER   .  .   st4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], x28
-# CHECK-NEXT: [0,7]     .  D=====================eER  .  .   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D====================eeeeeeER.   st4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], x28
-# CHECK-NEXT: [0,9]     .   D==========================eER   add	x0, x27, #1
+# CHECK:      [0,0]     DeeeeeeER .   .   st4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], x28
+# CHECK-NEXT: [0,1]     D=eE----R .   .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeER.   .   st4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], #8
+# CHECK-NEXT: [0,3]     .D=eE----R.   .   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeeeER  .   st4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], #8
+# CHECK-NEXT: [0,5]     . D==eE----R  .   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeeeER .   st4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], x28
+# CHECK-NEXT: [0,7]     .  D==eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eeeeeeER   st4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], x28
+# CHECK-NEXT: [0,9]     .   D===eE----R   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4939,43 +4928,43 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st4	{ v1.b, v2.b, v3.b, v4.b }[8], [x27], x28
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     6.0    0.0    0.0       st4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], #8
-# CHECK-NEXT: 3.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     11.0   0.0    0.0       st4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], #8
-# CHECK-NEXT: 5.     1     17.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     16.0   0.0    0.0       st4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], x28
-# CHECK-NEXT: 7.     1     22.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     21.0   0.0    0.0       st4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], x28
-# CHECK-NEXT: 9.     1     27.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     14.0   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], #8
+# CHECK-NEXT: 3.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       st4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], #8
+# CHECK-NEXT: 5.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st4	{ v1.h, v2.h, v3.h, v4.h }[0], [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    0.0       st4	{ v1.h, v2.h, v3.h, v4.h }[4], [x27], x28
+# CHECK-NEXT: 9.     1     4.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    2.0       <total>
 
 # CHECK:      [82] Code Region - G83
 
 # CHECK:      Iterations:        100
 # CHECK-NEXT: Instructions:      1000
-# CHECK-NEXT: Total Cycles:      2104
+# CHECK-NEXT: Total Cycles:      704
 # CHECK-NEXT: Total uOps:        3600
 
 # CHECK:      Dispatch Width:    10
-# CHECK-NEXT: uOps Per Cycle:    1.71
-# CHECK-NEXT: IPC:               0.48
+# CHECK-NEXT: uOps Per Cycle:    5.11
+# CHECK-NEXT: IPC:               1.42
 # CHECK-NEXT: Block RThroughput: 6.5
 
 # CHECK:      Timeline view:
-# CHECK-NEXT:                     0123456789
-# CHECK-NEXT: Index     0123456789          01234
-
-# CHECK:      [0,0]     DeeeeeeER .    .    .   .   st4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], #16
-# CHECK-NEXT: [0,1]     D======eER.    .    .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,2]     .D=====eeeeeeER.    .   .   st4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], x28
-# CHECK-NEXT: [0,3]     .D===========eER    .   .   add	x0, x27, #1
-# CHECK-NEXT: [0,4]     . D==========eeeeER .   .   st4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], #32
-# CHECK-NEXT: [0,5]     . D==============eER.   .   add	x0, x27, #1
-# CHECK-NEXT: [0,6]     .  D=============eeeeER .   st4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], x28
-# CHECK-NEXT: [0,7]     .  D=================eER.   add	x0, x27, #1
-# CHECK-NEXT: [0,8]     .   D================eER.   stg	x26, [x27], #4064
-# CHECK-NEXT: [0,9]     .   D=================eER   add	x0, x27, #1
+# CHECK-NEXT:                     0
+# CHECK-NEXT: Index     0123456789
+
+# CHECK:      [0,0]     DeeeeeeER .   st4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], #16
+# CHECK-NEXT: [0,1]     D=eE----R .   add	x0, x27, #1
+# CHECK-NEXT: [0,2]     .DeeeeeeER.   st4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], x28
+# CHECK-NEXT: [0,3]     .D=eE----R.   add	x0, x27, #1
+# CHECK-NEXT: [0,4]     . D=eeeeER.   st4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], #32
+# CHECK-NEXT: [0,5]     . D==eE--R.   add	x0, x27, #1
+# CHECK-NEXT: [0,6]     .  D=eeeeER   st4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], x28
+# CHECK-NEXT: [0,7]     .  D==eE--R   add	x0, x27, #1
+# CHECK-NEXT: [0,8]     .   D==eE-R   stg	x26, [x27], #4064
+# CHECK-NEXT: [0,9]     .   D===eER   add	x0, x27, #1
 
 # CHECK:      Average Wait times (based on the timeline view):
 # CHECK-NEXT: [0]: Executions
@@ -4985,16 +4974,16 @@ add x0, x27, 1
 
 # CHECK:            [0]    [1]    [2]    [3]
 # CHECK-NEXT: 0.     1     1.0    1.0    0.0       st4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], #16
-# CHECK-NEXT: 1.     1     7.0    0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 2.     1     6.0    0.0    0.0       st4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], x28
-# CHECK-NEXT: 3.     1     12.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 4.     1     11.0   0.0    0.0       st4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], #32
-# CHECK-NEXT: 5.     1     15.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 6.     1     14.0   0.0    0.0       st4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], x28
-# CHECK-NEXT: 7.     1     18.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT: 8.     1     17.0   0.0    0.0       stg	x26, [x27], #4064
-# CHECK-NEXT: 9.     1     18.0   0.0    0.0       add	x0, x27, #1
-# CHECK-NEXT:        1     11.9   0.1    0.0       <total>
+# CHECK-NEXT: 1.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 2.     1     1.0    0.0    0.0       st4	{ v1.s, v2.s, v3.s, v4.s }[0], [x27], x28
+# CHECK-NEXT: 3.     1     2.0    0.0    4.0       add	x0, x27, #1
+# CHECK-NEXT: 4.     1     2.0    1.0    0.0       st4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], #32
+# CHECK-NEXT: 5.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 6.     1     2.0    0.0    0.0       st4	{ v1.d, v2.d, v3.d, v4.d }[0], [x27], x28
+# CHECK-NEXT: 7.     1     3.0    0.0    2.0       add	x0, x27, #1
+# CHECK-NEXT: 8.     1     3.0    1.0    1.0       stg	x26, [x27], #4064
+# CHECK-NEXT: 9.     1     4.0    0.0    0.0       add	x0, x27, #1
+# CHECK-NEXT:        1     2.3    0.3    1.3       <total>
 
 # CHECK:      [83] Code Region - G84
 


        


More information about the llvm-commits mailing list