From 1cf34058c764db6e1651bb5b99fb659bb76f8737 Mon Sep 17 00:00:00 2001
From: Paulo Matos <pmatos@igalia.com>
Date: Tue, 3 Dec 2024 17:41:18 +0100
Subject: [PATCH] instcountci: Cache predicate register generation from pattern

---
 unittests/InstructionCountCI/X87ldst-SVE.json | 24 ++++---------------
 1 file changed, 4 insertions(+), 20 deletions(-)

diff --git a/unittests/InstructionCountCI/X87ldst-SVE.json b/unittests/InstructionCountCI/X87ldst-SVE.json
index d82b68d9b1..ec0abde349 100644
--- a/unittests/InstructionCountCI/X87ldst-SVE.json
+++ b/unittests/InstructionCountCI/X87ldst-SVE.json
@@ -34,7 +34,7 @@
     },
     "2-store 80bit": {
       "x86InstructionCount": 2,
-      "ExpectedInstructionCount": 25,
+      "ExpectedInstructionCount": 24,
       "x86Insts": [
         "fstp tword [rax]",
         "fstp tword [rax+10]"
@@ -56,7 +56,6 @@
         "add x21, x4, #0xa (10)",
         "add x0, x28, x20, lsl #4",
         "ldr q2, [x0, #1040]",
-        "ptrue p2.h, vl5",
         "st1h {z2.h}, p2, [x21]",
         "ldrb w21, [x28, #1298]",
         "lsl w22, w22, w20",
@@ -69,7 +68,7 @@
     },
     "8-store 80bit": {
       "x86InstructionCount": 8,
-      "ExpectedInstructionCount": 97,
+      "ExpectedInstructionCount": 90,
       "x86Insts": [
         "fstp tword [rax]",
         "fstp tword [rax+10]",
@@ -97,7 +96,6 @@
         "add x21, x4, #0xa (10)",
         "add x0, x28, x20, lsl #4",
         "ldr q2, [x0, #1040]",
-        "ptrue p2.h, vl5",
         "st1h {z2.h}, p2, [x21]",
         "ldrb w21, [x28, #1298]",
         "lsl w23, w22, w20",
@@ -109,7 +107,6 @@
         "add x21, x4, #0x14 (20)",
         "add x0, x28, x20, lsl #4",
         "ldr q2, [x0, #1040]",
-        "ptrue p2.h, vl5",
         "st1h {z2.h}, p2, [x21]",
         "ldrb w21, [x28, #1298]",
         "lsl w23, w22, w20",
@@ -121,7 +118,6 @@
         "add x21, x4, #0x1e (30)",
         "add x0, x28, x20, lsl #4",
         "ldr q2, [x0, #1040]",
-        "ptrue p2.h, vl5",
         "st1h {z2.h}, p2, [x21]",
         "ldrb w21, [x28, #1298]",
         "lsl w23, w22, w20",
@@ -133,7 +129,6 @@
         "add x21, x4, #0x28 (40)",
         "add x0, x28, x20, lsl #4",
         "ldr q2, [x0, #1040]",
-        "ptrue p2.h, vl5",
         "st1h {z2.h}, p2, [x21]",
         "ldrb w21, [x28, #1298]",
         "lsl w23, w22, w20",
@@ -145,7 +140,6 @@
         "add x21, x4, #0x32 (50)",
         "add x0, x28, x20, lsl #4",
         "ldr q2, [x0, #1040]",
-        "ptrue p2.h, vl5",
         "st1h {z2.h}, p2, [x21]",
         "ldrb w21, [x28, #1298]",
         "lsl w23, w22, w20",
@@ -157,7 +151,6 @@
         "add x21, x4, #0x3c (60)",
         "add x0, x28, x20, lsl #4",
         "ldr q2, [x0, #1040]",
-        "ptrue p2.h, vl5",
         "st1h {z2.h}, p2, [x21]",
         "ldrb w21, [x28, #1298]",
         "lsl w23, w22, w20",
@@ -169,7 +162,6 @@
         "add x21, x4, #0x46 (70)",
         "add x0, x28, x20, lsl #4",
         "ldr q2, [x0, #1040]",
-        "ptrue p2.h, vl5",
         "st1h {z2.h}, p2, [x21]",
         "ldrb w21, [x28, #1298]",
         "lsl w22, w22, w20",
@@ -201,7 +193,7 @@
     },
     "2-load 80bit": {
       "x86InstructionCount": 2,
-      "ExpectedInstructionCount": 22,
+      "ExpectedInstructionCount": 21,
       "x86Insts": [
         "fld tword [rax]",
         "fld tword [rax+10]"
@@ -210,7 +202,6 @@
         "ptrue p2.h, vl5",
         "ld1h {z2.h}, p2/z, [x4]",
         "add x20, x4, #0xa (10)",
-        "ptrue p2.h, vl5",
         "ld1h {z3.h}, p2/z, [x20]",
         "ldrb w20, [x28, #1019]",
         "sub w20, w20, #0x2 (2)",
@@ -233,7 +224,7 @@
     },
     "8-load 80bit": {
       "x86InstructionCount": 8,
-      "ExpectedInstructionCount": 59,
+      "ExpectedInstructionCount": 52,
       "x86Insts": [
         "fld tword [rax]",
         "fld tword [rax+10]",
@@ -248,25 +239,18 @@
         "ptrue p2.h, vl5",
         "ld1h {z2.h}, p2/z, [x4]",
         "add x20, x4, #0xa (10)",
-        "ptrue p2.h, vl5",
         "ld1h {z3.h}, p2/z, [x20]",
         "add x20, x4, #0x14 (20)",
-        "ptrue p2.h, vl5",
         "ld1h {z4.h}, p2/z, [x20]",
         "add x20, x4, #0x1e (30)",
-        "ptrue p2.h, vl5",
         "ld1h {z5.h}, p2/z, [x20]",
         "add x20, x4, #0x28 (40)",
-        "ptrue p2.h, vl5",
         "ld1h {z6.h}, p2/z, [x20]",
         "add x20, x4, #0x32 (50)",
-        "ptrue p2.h, vl5",
         "ld1h {z7.h}, p2/z, [x20]",
         "add x20, x4, #0x3c (60)",
-        "ptrue p2.h, vl5",
         "ld1h {z8.h}, p2/z, [x20]",
         "add x20, x4, #0x46 (70)",
-        "ptrue p2.h, vl5",
         "ld1h {z9.h}, p2/z, [x20]",
         "ldrb w20, [x28, #1019]",
         "sub w20, w20, #0x8 (8)",