diff --git a/unittests/InstructionCountCI/FEXOpt/MultiInst.json b/unittests/InstructionCountCI/FEXOpt/MultiInst.json
index 9610c2947..4696cad39 100644
--- a/unittests/InstructionCountCI/FEXOpt/MultiInst.json
+++ b/unittests/InstructionCountCI/FEXOpt/MultiInst.json
@@ -34,7 +34,7 @@
     },
     "push rax, rbx": {
       "x86InstructionCount": 2,
-      "ExpectedInstructionCount": 2,
+      "ExpectedInstructionCount": 1,
       "Comment": [
         "Mergable 64-bit pushes"
       ],
@@ -43,8 +43,7 @@
         "push rbx"
       ],
       "ExpectedArm64ASM": [
-        "str x4, [x8, #-8]!",
-        "str x6, [x8, #-8]!"
+        "stp x6, x4, [x8, #-16]!"
       ]
     },
     "adds xmm0, xmm1, xmm2": {
@@ -1056,7 +1055,7 @@
     },
     "Sekiro spill block": {
       "x86InstructionCount": 119,
-      "ExpectedInstructionCount": 126,
+      "ExpectedInstructionCount": 118,
       "Comment": [
         "This block of code came from the settings screen when it loaded",
         "It was originally at RIP: 0x14232cca0 and has been deobfuscated"
@@ -1184,14 +1183,10 @@
       ],
       "ExpectedArm64ASM": [
         "str x7, [x8, #8]",
-        "str x6, [x8, #-8]!",
-        "str x9, [x8, #-8]!",
-        "str x10, [x8, #-8]!",
-        "str x11, [x8, #-8]!",
-        "str x16, [x8, #-8]!",
-        "str x17, [x8, #-8]!",
-        "str x19, [x8, #-8]!",
-        "str x29, [x8, #-8]!",
+        "stp x9, x6, [x8, #-16]!",
+        "stp x11, x10, [x8, #-16]!",
+        "stp x17, x16, [x8, #-16]!",
+        "stp x29, x19, [x8, #-16]!",
         "sub x8, x8, #0x18 (24)",
         "ldr w7, [x5, #36]",
         "ldr w10, [x5]",
@@ -1300,14 +1295,10 @@
         "mvn w27, w8",
         "adds x26, x8, #0x18 (24)",
         "mov x8, x26",
-        "ldr x29, [x8], #8",
-        "ldr x19, [x8], #8",
-        "ldr x17, [x8], #8",
-        "ldr x16, [x8], #8",
-        "ldr x11, [x8], #8",
-        "ldr x10, [x8], #8",
-        "ldr x9, [x8], #8",
-        "ldr x6, [x8], #8",
+        "ldp x29, x19, [x8], #16",
+        "ldp x17, x16, [x8], #16",
+        "ldp x11, x10, [x8], #16",
+        "ldp x9, x6, [x8], #16",
         "cfinv"
       ]
     },
diff --git a/unittests/InstructionCountCI/FlagM/HotBlocks.json b/unittests/InstructionCountCI/FlagM/HotBlocks.json
index e57ceb3f2..e11d0ec42 100644
--- a/unittests/InstructionCountCI/FlagM/HotBlocks.json
+++ b/unittests/InstructionCountCI/FlagM/HotBlocks.json
@@ -253,7 +253,7 @@
     },
     "bytemark DivideInternalFPF": {
       "x86InstructionCount": 13,
-      "ExpectedInstructionCount": 15,
+      "ExpectedInstructionCount": 12,
       "Comment": [
         "Saw this in bytemark"
       ],
@@ -273,12 +273,9 @@
         "cmp    eax,0x18"
       ],
       "ExpectedArm64ASM": [
-        "str x9, [x8, #-8]!",
-        "str x29, [x8, #-8]!",
-        "str x19, [x8, #-8]!",
-        "str x17, [x8, #-8]!",
-        "str x16, [x8, #-8]!",
-        "str x6, [x8, #-8]!",
+        "stp x29, x9, [x8, #-16]!",
+        "stp x17, x19, [x8, #-16]!",
+        "stp x6, x16, [x8, #-16]!",
         "sub x8, x8, #0x18 (24)",
         "str x5, [x8]",
         "ldrb w14, [x11]",
diff --git a/unittests/InstructionCountCI/FlagM/HotBlocks_TSO_32Bit.json b/unittests/InstructionCountCI/FlagM/HotBlocks_TSO_32Bit.json
index 6bbfd644e..307647d5e 100644
--- a/unittests/InstructionCountCI/FlagM/HotBlocks_TSO_32Bit.json
+++ b/unittests/InstructionCountCI/FlagM/HotBlocks_TSO_32Bit.json
@@ -17,7 +17,7 @@
   "Instructions": {
     "The Sims 1 hot block": {
       "x86InstructionCount": 47,
-      "ExpectedInstructionCount": 98,
+      "ExpectedInstructionCount": 97,
       "Comment": [
         "Hottest in-game block from The Sims 1, Legacy Collection",
         "Consumed 6.13% of a CPU core on Oryon-1",
@@ -92,8 +92,7 @@
         "stlur w4, [x9, #-4]",
         "ldapur w5, [x9, #8]",
         "nop",
-        "str w6, [x8, #-4]!",
-        "str w10, [x8, #-4]!",
+        "stp w10, w6, [x8, #-8]!",
         "ldapur w10, [x9, #24]",
         "nop",
         "str w11, [x8, #-4]!",
diff --git a/unittests/InstructionCountCI/FlagM/Primary_32Bit.json b/unittests/InstructionCountCI/FlagM/Primary_32Bit.json
index e4bac5ca6..8813249ea 100644
--- a/unittests/InstructionCountCI/FlagM/Primary_32Bit.json
+++ b/unittests/InstructionCountCI/FlagM/Primary_32Bit.json
@@ -247,67 +247,51 @@
       ]
     },
     "pusha": {
-      "ExpectedInstructionCount": 10,
+      "ExpectedInstructionCount": 5,
       "Comment": "0x60",
       "ExpectedArm64ASM": [
-        "mov w20, w8",
-        "str w4, [x20, #-4]!",
-        "str w7, [x20, #-4]!",
-        "str w5, [x20, #-4]!",
-        "str w6, [x20, #-4]!",
-        "str w8, [x20, #-4]!",
-        "str w9, [x20, #-4]!",
-        "str w10, [x20, #-4]!",
-        "mov w8, w20",
-        "str w11, [x8, #-4]!"
+        "mov x20, x8",
+        "stp w7, w4, [x8, #-8]!",
+        "stp w6, w5, [x8, #-8]!",
+        "stp w9, w20, [x8, #-8]!",
+        "stp w11, w10, [x8, #-8]!"
       ]
     },
     "pushad": {
-      "ExpectedInstructionCount": 10,
+      "ExpectedInstructionCount": 5,
       "Comment": "0x60",
       "ExpectedArm64ASM": [
-        "mov w20, w8",
-        "str w4, [x20, #-4]!",
-        "str w7, [x20, #-4]!",
-        "str w5, [x20, #-4]!",
-        "str w6, [x20, #-4]!",
-        "str w8, [x20, #-4]!",
-        "str w9, [x20, #-4]!",
-        "str w10, [x20, #-4]!",
-        "mov w8, w20",
-        "str w11, [x8, #-4]!"
+        "mov x20, x8",
+        "stp w7, w4, [x8, #-8]!",
+        "stp w6, w5, [x8, #-8]!",
+        "stp w9, w20, [x8, #-8]!",
+        "stp w11, w10, [x8, #-8]!"
       ]
     },
     "popa": {
-      "ExpectedInstructionCount": 10,
+      "ExpectedInstructionCount": 7,
       "Comment": "0x61",
       "ExpectedArm64ASM": [
         "mov x20, x8",
-        "ldr w11, [x20], #4",
-        "ldr w10, [x20], #4",
+        "ldp w11, w10, [x20], #8",
         "ldr w9, [x20], #4",
         "add x20, x20, #0x4 (4)",
         "mov x8, x20",
-        "ldr w6, [x8], #4",
-        "ldr w5, [x8], #4",
-        "ldr w7, [x8], #4",
-        "ldr w4, [x8], #4"
+        "ldp w6, w5, [x8], #8",
+        "ldp w7, w4, [x8], #8"
       ]
     },
     "popad": {
-      "ExpectedInstructionCount": 10,
+      "ExpectedInstructionCount": 7,
       "Comment": "0x61",
       "ExpectedArm64ASM": [
         "mov x20, x8",
-        "ldr w11, [x20], #4",
-        "ldr w10, [x20], #4",
+        "ldp w11, w10, [x20], #8",
         "ldr w9, [x20], #4",
         "add x20, x20, #0x4 (4)",
         "mov x8, x20",
-        "ldr w6, [x8], #4",
-        "ldr w5, [x8], #4",
-        "ldr w7, [x8], #4",
-        "ldr w4, [x8], #4"
+        "ldp w6, w5, [x8], #8",
+        "ldp w7, w4, [x8], #8"
       ]
     },
     "o16 pushf": {
diff --git a/unittests/InstructionCountCI/FlagM/x87-Crysis2Max-fmodel.json b/unittests/InstructionCountCI/FlagM/x87-Crysis2Max-fmodel.json
index 8c75d53d1..92c838b55 100644
--- a/unittests/InstructionCountCI/FlagM/x87-Crysis2Max-fmodel.json
+++ b/unittests/InstructionCountCI/FlagM/x87-Crysis2Max-fmodel.json
@@ -8585,8 +8585,7 @@
         "fmov s2, s0",
         "str s2, [x7, #960]",
         "mov x8, x9",
-        "ldr w9, [x8], #4",
-        "ldr w20, [x8], #4",
+        "ldp w9, w20, [x8], #8",
         "ldrb w21, [x28, #1019]",
         "add w21, w21, #0x3 (3)",
         "and w21, w21, #0x7",
@@ -8597,7 +8596,7 @@
         "strb w21, [x28, #1298]"
       ],
       "x86InstructionCount": 809,
-      "ExpectedInstructionCount": 7770
+      "ExpectedInstructionCount": 7769
     }
   }
 }
diff --git a/unittests/InstructionCountCI/FlagM/x87-HalfLife.json b/unittests/InstructionCountCI/FlagM/x87-HalfLife.json
index 99139a050..4032800ca 100644
--- a/unittests/InstructionCountCI/FlagM/x87-HalfLife.json
+++ b/unittests/InstructionCountCI/FlagM/x87-HalfLife.json
@@ -1063,7 +1063,7 @@
     },
     "Block4": {
       "x86InstructionCount": 54,
-      "ExpectedInstructionCount": 80,
+      "ExpectedInstructionCount": 78,
       "x86Insts": [
         "push ebp",
         "push edi",
@@ -1121,10 +1121,8 @@
         "fld dword [esp + 0x2c]"
       ],
       "ExpectedArm64ASM": [
-        "str w9, [x8, #-4]!",
-        "str w11, [x8, #-4]!",
-        "str w10, [x8, #-4]!",
-        "str w6, [x8, #-4]!",
+        "stp w11, w9, [x8, #-8]!",
+        "stp w6, w10, [x8, #-8]!",
         "subs w26, w8, #0x4c (76)",
         "mov x27, x8",
         "mov x8, x26",
@@ -1564,7 +1562,7 @@
     },
     "Block6": {
       "x86InstructionCount": 39,
-      "ExpectedInstructionCount": 297,
+      "ExpectedInstructionCount": 295,
       "x86Insts": [
         "push ebp",
         "push edi",
@@ -1607,10 +1605,8 @@
         "fstp st1"
       ],
       "ExpectedArm64ASM": [
-        "str w9, [x8, #-4]!",
-        "str w11, [x8, #-4]!",
-        "str w10, [x8, #-4]!",
-        "str w6, [x8, #-4]!",
+        "stp w11, w9, [x8, #-8]!",
+        "stp w6, w10, [x8, #-8]!",
         "subs w20, w8, #0x4 (4)",
         "mov x27, x8",
         "mov x8, x20",
@@ -2611,7 +2607,7 @@
     },
     "Block10": {
       "x86InstructionCount": 125,
-      "ExpectedInstructionCount": 22,
+      "ExpectedInstructionCount": 21,
       "x86Insts": [
         "push esi",
         "push ebx",
@@ -2740,8 +2736,7 @@
         "xor eax,dword gs:[0x14]"
       ],
       "ExpectedArm64ASM": [
-        "str w10, [x8, #-4]!",
-        "str w6, [x8, #-4]!",
+        "stp w6, w10, [x8, #-8]!",
         "sub w8, w8, #0xa4 (164)",
         "ldr w6, [x8, #176]",
         "add w10, w8, #0x18 (24)",
diff --git a/unittests/InstructionCountCI/FlagM/x87-Oblivion.json b/unittests/InstructionCountCI/FlagM/x87-Oblivion.json
index 06ad7eb8d..caef346be 100644
--- a/unittests/InstructionCountCI/FlagM/x87-Oblivion.json
+++ b/unittests/InstructionCountCI/FlagM/x87-Oblivion.json
@@ -17705,7 +17705,7 @@
     },
     "Block4": {
       "x86InstructionCount": 2050,
-      "ExpectedInstructionCount": 40,
+      "ExpectedInstructionCount": 39,
       "x86Insts": [
         "fldz",
         "push 0x0",
@@ -19768,8 +19768,7 @@
         "movk w21, #0x100, lsl #16",
         "str w21, [x8, #-4]!",
         "mov w21, #0x37",
-        "str w21, [x8, #-4]!",
-        "str w7, [x8, #-4]!",
+        "stp w7, w21, [x8, #-8]!",
         "str x30, [sp, #-16]!",
         "mov v0.16b, v2.16b",
         "ldr x0, [x28, #1456]",
diff --git a/unittests/InstructionCountCI/FlagM/x87-Psychonauts.json b/unittests/InstructionCountCI/FlagM/x87-Psychonauts.json
index 15ccf61e7..3975e25d1 100644
--- a/unittests/InstructionCountCI/FlagM/x87-Psychonauts.json
+++ b/unittests/InstructionCountCI/FlagM/x87-Psychonauts.json
@@ -26756,7 +26756,7 @@
     },
     "Block8": {
       "x86InstructionCount": 231,
-      "ExpectedInstructionCount": 3724,
+      "ExpectedInstructionCount": 3723,
       "x86Insts": [
         "fadd dword [esp + 0x40]",
         "lea edx,[ecx + ecx*0x2]",
@@ -30693,8 +30693,7 @@
         "add w20, w20, #0x1 (1)",
         "and w20, w20, #0x7",
         "strb w20, [x28, #1019]",
-        "ldr w11, [x8], #4",
-        "ldr w10, [x8], #4",
+        "ldp w11, w10, [x8], #8",
         "ldrb w22, [x28, #1298]",
         "lsl w23, w21, w20",
         "bic w22, w22, w23",
@@ -30719,7 +30718,7 @@
     },
     "Block9": {
       "x86InstructionCount": 222,
-      "ExpectedInstructionCount": 3646,
+      "ExpectedInstructionCount": 3645,
       "x86Insts": [
         "fadd dword [esp + 0x40]",
         "lea edx,[ecx + ecx*0x2]",
@@ -34569,8 +34568,7 @@
         "add w20, w20, #0x1 (1)",
         "and w20, w20, #0x7",
         "strb w20, [x28, #1019]",
-        "ldr w11, [x8], #4",
-        "ldr w10, [x8], #4",
+        "ldp w11, w10, [x8], #8",
         "ldrb w22, [x28, #1298]",
         "lsl w23, w21, w20",
         "bic w22, w22, w23",
diff --git a/unittests/InstructionCountCI/Primary_32Bit.json b/unittests/InstructionCountCI/Primary_32Bit.json
index 2c27708d4..fa0b4a854 100644
--- a/unittests/InstructionCountCI/Primary_32Bit.json
+++ b/unittests/InstructionCountCI/Primary_32Bit.json
@@ -262,67 +262,51 @@
       ]
     },
     "pusha": {
-      "ExpectedInstructionCount": 10,
+      "ExpectedInstructionCount": 5,
       "Comment": "0x60",
       "ExpectedArm64ASM": [
-        "mov w20, w8",
-        "str w4, [x20, #-4]!",
-        "str w7, [x20, #-4]!",
-        "str w5, [x20, #-4]!",
-        "str w6, [x20, #-4]!",
-        "str w8, [x20, #-4]!",
-        "str w9, [x20, #-4]!",
-        "str w10, [x20, #-4]!",
-        "mov w8, w20",
-        "str w11, [x8, #-4]!"
+        "mov x20, x8",
+        "stp w7, w4, [x8, #-8]!",
+        "stp w6, w5, [x8, #-8]!",
+        "stp w9, w20, [x8, #-8]!",
+        "stp w11, w10, [x8, #-8]!"
       ]
     },
     "pushad": {
-      "ExpectedInstructionCount": 10,
+      "ExpectedInstructionCount": 5,
       "Comment": "0x60",
       "ExpectedArm64ASM": [
-        "mov w20, w8",
-        "str w4, [x20, #-4]!",
-        "str w7, [x20, #-4]!",
-        "str w5, [x20, #-4]!",
-        "str w6, [x20, #-4]!",
-        "str w8, [x20, #-4]!",
-        "str w9, [x20, #-4]!",
-        "str w10, [x20, #-4]!",
-        "mov w8, w20",
-        "str w11, [x8, #-4]!"
+        "mov x20, x8",
+        "stp w7, w4, [x8, #-8]!",
+        "stp w6, w5, [x8, #-8]!",
+        "stp w9, w20, [x8, #-8]!",
+        "stp w11, w10, [x8, #-8]!"
       ]
     },
     "popa": {
-      "ExpectedInstructionCount": 10,
+      "ExpectedInstructionCount": 7,
       "Comment": "0x61",
       "ExpectedArm64ASM": [
         "mov x20, x8",
-        "ldr w11, [x20], #4",
-        "ldr w10, [x20], #4",
+        "ldp w11, w10, [x20], #8",
         "ldr w9, [x20], #4",
         "add x20, x20, #0x4 (4)",
         "mov x8, x20",
-        "ldr w6, [x8], #4",
-        "ldr w5, [x8], #4",
-        "ldr w7, [x8], #4",
-        "ldr w4, [x8], #4"
+        "ldp w6, w5, [x8], #8",
+        "ldp w7, w4, [x8], #8"
       ]
     },
     "popad": {
-      "ExpectedInstructionCount": 10,
+      "ExpectedInstructionCount": 7,
       "Comment": "0x61",
       "ExpectedArm64ASM": [
         "mov x20, x8",
-        "ldr w11, [x20], #4",
-        "ldr w10, [x20], #4",
+        "ldp w11, w10, [x20], #8",
         "ldr w9, [x20], #4",
         "add x20, x20, #0x4 (4)",
         "mov x8, x20",
-        "ldr w6, [x8], #4",
-        "ldr w5, [x8], #4",
-        "ldr w7, [x8], #4",
-        "ldr w4, [x8], #4"
+        "ldp w6, w5, [x8], #8",
+        "ldp w7, w4, [x8], #8"
       ]
     },
     "aam": {