Added test of vector arguments.

.github/workflows/ci.yml

@@ -87,6 +87,22 @@ jobs:
           fi
         done
         codecov --required
+  avx:
+    runs-on: ubuntu-latest
+    env:
+      CFLAGS: -mavx
+    steps:
+    - uses: actions/checkout@v3
+    - name: Run tests
+      run: scripts/travis.sh
+#  avx2:
+#    runs-on: ubuntu-latest
+#    env:
+#      CFLAGS: -mavx2 -mavx512f
+#    steps:
+#    - uses: actions/checkout@v3
+#    - name: Run tests
+#      run: scripts/travis.sh
   x86:
     runs-on: ubuntu-latest
     env:

arch/x86_64/table.S.tpl

@@ -34,18 +34,27 @@ _${lib_suffix}_save_regs_and_resolve:
 #define DEC_STACK(d) subq $$d, %rsp; .cfi_adjust_cfa_offset d
 #define INC_STACK(d) addq $$d, %rsp; .cfi_adjust_cfa_offset -d
 
+#define PUSH_MMX_REG(reg) DEC_STACK(8); movq %reg, (%rsp); .cfi_rel_offset reg, 0
+#define POP_MMX_REG(reg) movq (%rsp), %reg; .cfi_restore reg; INC_STACK(8)
+
 #define PUSH_XMM_REG(reg) DEC_STACK(16); movdqa %reg, (%rsp); .cfi_rel_offset reg, 0
 #define POP_XMM_REG(reg) movdqa (%rsp), %reg; .cfi_restore reg; INC_STACK(16)
 
+// TODO: cfi_offset/cfi_restore
+#define PUSH_YMM_REG(reg) DEC_STACK(32); vmovdqu %reg, (%rsp)
+#define POP_YMM_REG(reg) vmovdqu (%rsp), %reg; INC_STACK(32)
+
+// TODO: cfi_offset/cfi_restore
+#define PUSH_ZMM_REG(reg) DEC_STACK(64); vmovdqu32 %reg, (%rsp)
+#define POP_ZMM_REG(reg) vmovdqu32 (%rsp), %reg; INC_STACK(64)
+
   // Slow path which calls dlsym, taken only on first call.
   // All registers are stored to handle arbitrary calling conventions
   // (except x87 FPU registers which do not have to be preserved).
   // For Dwarf directives, read https://www.imperialviolet.org/2017/01/18/cfi.html.
 
   .cfi_def_cfa_offset 8  // Return address
 
-  // FIXME: AVX (YMM, ZMM) registers are NOT saved to simplify code.
-
   PUSH_REG(rdi)  // 16
   mov 0x10(%rsp), %rdi
   PUSH_REG(rax)
@@ -62,6 +71,26 @@ _${lib_suffix}_save_regs_and_resolve:
   PUSH_REG(r13)  // 16
   PUSH_REG(r14)
   PUSH_REG(r15)  // 16
+
+#ifdef __AVX2__
+  PUSH_ZMM_REG(zmm0)
+  PUSH_ZMM_REG(zmm1)
+  PUSH_ZMM_REG(zmm2)
+  PUSH_ZMM_REG(zmm3)
+  PUSH_ZMM_REG(zmm4)
+  PUSH_ZMM_REG(zmm5)
+  PUSH_ZMM_REG(zmm6)
+  PUSH_ZMM_REG(zmm7)
+#elif defined __AVX__
+  PUSH_YMM_REG(ymm0)
+  PUSH_YMM_REG(ymm1)
+  PUSH_YMM_REG(ymm2)
+  PUSH_YMM_REG(ymm3)
+  PUSH_YMM_REG(ymm4)
+  PUSH_YMM_REG(ymm5)
+  PUSH_YMM_REG(ymm6)
+  PUSH_YMM_REG(ymm7)
+#elif defined __SSE__
   PUSH_XMM_REG(xmm0)
   PUSH_XMM_REG(xmm1)
   PUSH_XMM_REG(xmm2)
@@ -70,10 +99,39 @@ _${lib_suffix}_save_regs_and_resolve:
   PUSH_XMM_REG(xmm5)
   PUSH_XMM_REG(xmm6)
   PUSH_XMM_REG(xmm7)
+#elif defined __MMX__
+  PUSH_MMX_REG(mm0)
+  PUSH_MMX_REG(mm1)
+  PUSH_MMX_REG(mm2)
+  PUSH_MMX_REG(mm3)
+  PUSH_MMX_REG(mm4)
+  PUSH_MMX_REG(mm5)
+  PUSH_MMX_REG(mm6)
+  PUSH_MMX_REG(mm7)
+#endif
 
   // Stack is just 8-byte aligned but callee will re-align to 16
   call _${lib_suffix}_tramp_resolve
 
+#ifdef __AVX2__
+  POP_ZMM_REG(zmm7)
+  POP_ZMM_REG(zmm6)
+  POP_ZMM_REG(zmm5)
+  POP_ZMM_REG(zmm4)
+  POP_ZMM_REG(zmm3)
+  POP_ZMM_REG(zmm2)
+  POP_ZMM_REG(zmm1)
+  POP_ZMM_REG(zmm0)  // 16
+#elif defined __AVX__
+  POP_YMM_REG(ymm7)
+  POP_YMM_REG(ymm6)
+  POP_YMM_REG(ymm5)
+  POP_YMM_REG(ymm4)
+  POP_YMM_REG(ymm3)
+  POP_YMM_REG(ymm2)
+  POP_YMM_REG(ymm1)
+  POP_YMM_REG(ymm0)  // 16
+#elif defined __SSE__
   POP_XMM_REG(xmm7)
   POP_XMM_REG(xmm6)
   POP_XMM_REG(xmm5)
@@ -82,6 +140,17 @@ _${lib_suffix}_save_regs_and_resolve:
   POP_XMM_REG(xmm2)
   POP_XMM_REG(xmm1)
   POP_XMM_REG(xmm0)  // 16
+#elif defined __MMX__
+  POP_MMX_REG(mm7)
+  POP_MMX_REG(mm6)
+  POP_MMX_REG(mm5)
+  POP_MMX_REG(mm4)
+  POP_MMX_REG(mm3)
+  POP_MMX_REG(mm2)
+  POP_MMX_REG(mm1)
+  POP_MMX_REG(mm0)  // 16
+#endif
+
   POP_REG(r15)
   POP_REG(r14)  // 16
   POP_REG(r13)

scripts/travis.sh

@@ -41,3 +41,7 @@ if ! echo "$ARCH" | grep -q powerpc; then
   tests/many-functions/run.sh $ARCH
 fi
 tests/stack-args/run.sh $ARCH
+if ! echo "$ARCH" | grep -q 'powerpc\|mips\|riscv'; then
+  # TODO: support vector types for remaining platforms
+  tests/vector-args/run.sh $ARCH
+fi

tests/stack-args/run.sh

@@ -7,12 +7,9 @@
 # Use of this source code is governed by MIT license that can be
 # found in the LICENSE.txt file.
 
-# This is a simple test for Implib.so functionality.
+# This is a simple test that verifies that parameters are correctly passed on stack.
 # Run it like
 #   ./run.sh ARCH
-# where ARCH stands for any supported arch (arm, x86_64, etc., see `implib-gen -h' for full list).
-# Note that you may need to install qemu-user for respective platform
-# (i386 also needs gcc-multilib).
 
 set -eu
 

tests/vector-args/interposed.c

@@ -0,0 +1,34 @@
+/*
+ * Copyright 2024 Yury Gribov
+ *
+ * The MIT License (MIT)
+ * 
+ * Use of this source code is governed by MIT license that can be
+ * found in the LICENSE.txt file.
+ */
+
+#include "interposed.h"
+
+__attribute__((visibility("default")))
+vector_type foo(vector_type x) {
+  return 3 *x;
+}
+
+static
+#ifdef __clang__
+# if __clang_major__ >= 14
+__attribute__((noipa))
+# else
+__attribute__((noinline))
+# endif
+#else
+__attribute__((noinline,noclone))
+#endif
+vector_type dummy(vector_type x0, vector_type x1, vector_type x2, vector_type x3, vector_type x4, vector_type x5, vector_type x6, vector_type x7) {
+  return x0 + x1 + x2 + x3 + x4 + x5 + x6 + x7;
+}
+
+__attribute__((constructor)) void touch_vector_regs() {
+  vector_type zero = {0};
+  dummy(zero, zero, zero, zero, zero, zero, zero, zero);
+}

tests/vector-args/interposed.h

@@ -0,0 +1,33 @@
+/*
+ * Copyright 2024 Yury Gribov
+ *
+ * The MIT License (MIT)
+ * 
+ * Use of this source code is governed by MIT license that can be
+ * found in the LICENSE.txt file.
+ */
+
+#ifndef INTERPOSED_H
+#define INTERPOSED_H
+
+// Determine number of 32-bit ints in native vector type
+// for each supported platform
+#if defined __AVX2__ /* ZMM regs */
+# define VECTOR_SIZE 16
+#elif defined __AVX__ /* YMM regs */
+# define VECTOR_SIZE 8
+#elif defined __SSE__ /* XMM regs */ \
+    || defined __aarch64__ /* NEON regs */
+# define VECTOR_SIZE 4
+#elif defined __MMX__ /* MMX regs */ \
+    || defined __arm__  /* NEON regs */
+# define VECTOR_SIZE 2
+#else
+# error "Unknown platform"
+#endif
+
+typedef int vector_type __attribute__((vector_size(sizeof(int) * VECTOR_SIZE)));
+
+extern vector_type foo(vector_type x);
+
+#endif

tests/vector-args/main.c

@@ -0,0 +1,36 @@
+/*
+ * Copyright 2024 Yury Gribov
+ *
+ * The MIT License (MIT)
+ * 
+ * Use of this source code is governed by MIT license that can be
+ * found in the LICENSE.txt file.
+ */
+
+#include <stdio.h>
+
+#include "interposed.h"
+
+#if VECTOR_SIZE == 2
+# define VECTOR_INIT {1, 2}
+#elif VECTOR_SIZE == 4
+# define VECTOR_INIT {1, 2, 3, 4}
+#elif VECTOR_SIZE == 8
+# define VECTOR_INIT {1, 2, 3, 4, 5, 6, 7}
+#elif VECTOR_SIZE == 16
+# define VECTOR_INIT {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15}
+#else
+# error "Unsupported vector size"
+#endif
+
+int main() {
+  vector_type x = VECTOR_INIT, res = foo(x), ref = 3 * x;
+  int i;
+  for (i = 0; i < VECTOR_SIZE; ++i) {
+    if (res[i] != ref[i]) {
+      printf("NOT OK\n");
+      return 1;
+    }
+  }
+  return 0;
+}

tests/vector-args/run.sh

@@ -0,0 +1,38 @@
+#!/bin/sh
+
+# Copyright 2024 Yury Gribov
+#
+# The MIT License (MIT)
+# 
+# Use of this source code is governed by MIT license that can be
+# found in the LICENSE.txt file.
+
+# This test verifies that Implib trampolines save/restore vector regs correctly.
+# Run it like
+#   ./run.sh ARCH
+
+set -eu
+
+cd $(dirname $0)
+
+if test -n "${1:-}"; then
+  ARCH="$1"
+fi
+
+. ../common.sh
+
+CFLAGS="-g -O2 $CFLAGS"
+
+# Build shlib to test against
+$CC $CFLAGS -shared -fPIC interposed.c -o libinterposed.so
+
+# Prepare implib
+${PYTHON:-} ../../implib-gen.py -q --target $TARGET libinterposed.so
+
+# Build app
+$CC $CFLAGS main.c libinterposed.so.tramp.S libinterposed.so.init.c $LIBS
+
+LD_LIBRARY_PATH=.:${LD_LIBRARY_PATH:-} $INTERP ./a.out > a.out.log
+diff test.ref a.out.log
+
+echo SUCCESS