i#6777: AArch64-ci: Add runner for 128 bit sve

.github/workflows/ci-aarchxx.yml

@@ -50,8 +50,17 @@ jobs:
       fail-fast: false
       matrix:
         # This job will run in parallel.
-        os: [ubuntu-20-arm64, ubuntu-20-arm64-sve]
+        include:
+          - os: ubuntu-20-arm64
+            sve: false
+          - os: ubuntu-20-arm64-sve
+            sve: true
+            sve_length: 256
+          - os: ubuntu-20-arm64-sve
+            sve: true
+            sve_length: 128
     runs-on: ${{ matrix.os }}
+    name: AArch64 ${{matrix.sve && 'SVE' || ''}} precommit ${{matrix.sve_length}}
     steps:
       - name: Check out repository code
         uses: actions/checkout@v3
@@ -81,6 +90,14 @@ jobs:
         run: ../suite/runsuite_wrapper.pl travis
         env:
           CI_BRANCH: ${{ github.ref }}
+        if: ${{ matrix.sve == false || matrix.sve_length == '' }}
+
+      - name: Run Suite with sve vector length ${{ matrix.sve_length }}
+        working-directory: build
+        run: ../tools/run_with_vector_length.py ${{ matrix.sve_length }} ../suite/runsuite_wrapper.pl travis
+        env:
+          CI_BRANCH: ${{ github.ref }}
+        if: ${{ matrix.sve == true && matrix.sve_length != ''}}
 
       - name: Send failure mail to dynamorio-devs
         if: failure() && github.ref == 'refs/heads/master'

core/arch/aarch64/proc.c

@@ -113,7 +113,7 @@ get_processor_specific_info(void)
      * - Release or development test build.
      */
 #        if !defined(DR_HOST_NOT_TARGET)
-    if (proc_has_feature(FEATURE_SVE)) {
+    if (proc_has_feature(FEATURE_SVE) && false) {
         uint64 vl;
         /* This RDVL instruction is inserted as raw hex because we don't build
          * with SVE enabled: i.e. not -march=armv8-a+sve, so that we can run a

tools/get_vector_length.py

@@ -0,0 +1,34 @@
+#!/usr/bin/env python3
+
+"""
+Print the vector length in the current context
+"""
+
+import ctypes
+import sys
+
+# These constants must match linux/prctl.h
+PR_SVE_GET_VL = 51
+PR_SVE_VL_LEN_MASK = 0xffff
+
+def get_vector_length():
+    """
+    Use Prctl(2) to set the process tree's sve vector length. See
+    https://man7.org/linux/man-pages/man2/prctl.2.html for more details
+    """
+    libc = ctypes.cdll.LoadLibrary('libc.so.6')
+    return_code = libc.prctl(
+        PR_SVE_GET_VL)
+    if return_code < 0:
+        print(f'Failed to get VL, rc: {return_code}')
+        sys.exit(1)
+    print((return_code & PR_SVE_VL_LEN_MASK) * 8)
+
+if __name__ == '__main__':
+    if any(arg in ["--help", "-h"] for arg in sys.argv) or len(sys.argv) > 1:
+        print(f"usage: {sys.argv[0]}")
+        print("\nGets the current vector length of this process tree")
+        sys.exit(0)
+
+    get_vector_length()
+

tools/run_with_vector_length.py

@@ -0,0 +1,50 @@
+#!/usr/bin/env python3
+
+"""
+Run a program as a subprocess after configuring the vector length of the
+system to match the value in the first argument.
+"""
+
+import ctypes
+import subprocess
+import sys
+
+# These constants must match linux/prctl.h
+PR_SVE_SET_VL = 50
+
+PR_SVE_VL_INHERIT = 1 << 17
+PR_SVE_SET_VL_ONEXEC = 1 << 18
+PR_SVE_VL_LEN_MASK = 0xffff
+
+def set_vector_length(vector_length):
+    """
+    Use Prctl(2) to set the process tree's sve vector length. See
+    https://man7.org/linux/man-pages/man2/prctl.2.html for more details
+    """
+    vector_length_bytes =  vector_length // 8
+    libc = ctypes.cdll.LoadLibrary('libc.so.6')
+    return_code = libc.prctl(
+        PR_SVE_SET_VL,
+        PR_SVE_SET_VL_ONEXEC | PR_SVE_VL_INHERIT | vector_length_bytes)
+    if return_code < 0:
+        print(f'Failed to set VL, rc: {return_code}')
+        sys.exit(1)
+    set_vector_length = (return_code & PR_SVE_VL_LEN_MASK) * 8
+
+    if set_vector_length != vector_length:
+        print(
+            f"Requested vector length {vector_length} not set, maximum of {set_vector_length} available",
+            file=sys.stderr)
+        sys.exit(1)
+
+if __name__ == '__main__':
+    if any(arg in ["--help", "-h"] for arg in sys.argv) or len(sys.argv) < 3:
+        print(f"usage: {sys.argv[0]} [vector_length] [program] [program args]...")
+        print("\nSet the vector length of this process tree to [vector_length] bits")
+        print("and then execute [program]")
+        sys.exit(0)
+
+    set_vector_length(int(sys.argv[1]))
+
+    command = sys.argv[2:]
+    sys.exit(subprocess.run(command, check=False).returncode)

tools/set_vector_length.py

@@ -0,0 +1,43 @@
+#!/usr/bin/env python3
+
+"""
+Run a program as a subprocess after configuring the vector length of the
+system to match the value in the first argument.
+"""
+
+import ctypes
+import subprocess
+import sys
+
+# These constants must match linux/prctl.h
+PR_SVE_SET_VL = 50
+PR_SVE_GET_VL = 51
+
+PR_SVE_VL_INHERIT = 1 << 17
+PR_SVE_SET_VL_ONEXEC = 1 << 18
+PR_SVE_VL_LEN_MASK = 0xffff
+
+def set_vector_length(vector_length):
+    """
+    Use Prctl(2) to set the process tree's sve vector length. See
+    https://man7.org/linux/man-pages/man2/prctl.2.html for more details
+    """
+    libc = ctypes.cdll.LoadLibrary('libc.so.6')
+    return_code = libc.prctl(
+        PR_SVE_SET_VL,
+        PR_SVE_SET_VL_ONEXEC | PR_SVE_VL_INHERIT | vector_length)
+    if return_code < 0:
+        print(f'Failed to set VL, rc: {return_code}')
+        sys.exit(1)
+
+if __name__ == '__main__':
+    if any(arg in ["--help", "-h"] for arg in sys.argv) or len(sys.argv) < 3:
+        print(f"usage: {sys.argv[0]} [vector_length] [program] [program args]...")
+        print("\nSet the vector length of this process tree to [vector_length]")
+        print("and then execute [program]")
+        sys.exit(0)
+
+    set_vector_length(int(sys.argv[1]))
+
+    command = sys.argv[2:]
+    subprocess.run(command, check=False)