Fixed kernel stub argument marshalling code to ensure proper alignment

Previously, shamrock was generating load instructions in the kernel stub with
a strict type alignment, which was not being followed by the argument
marshalling code.

This resulted in a NEON vld1 instruction failing in the kernel stub,
and trashing a base register which was not 16 byte aligned.

Now the marshalling code calculates the proper alignement for each argument,
and is based on a buffer aligned to double16 to begin with.

With this patch, all the vector sub-tests of the Khronos basic_parameter_types
test now pass.

Signed-off-by: Gil Pitney <gil.pitney@linaro.org>

1 files changed, 10 insertions, 6 deletions

diff --git a/src/core/cpu/kernel.cpp b/src/core/cpu/kernel.cpp
index e81391f..49e4dcc 100644
--- a/src/core/cpu/kernel.cpp
+++ b/src/core/cpu/kernel.cpp
@@ -290,10 +290,11 @@ size_t CPUKernel::typeOffset(size_t &offset, size_t type_len)
 
     // Align offset to stype_len
     type_len = next_power_of_two(type_len);
-    size_t mask = ~(type_len - 1);
+    size_t mask = (type_len - 1);
 
-    while (rs & mask != rs)
-        rs++;
+    if (rs&mask) {
+      rs += (type_len - rs%type_len);
+    }
 
     // Where to try to place the next value
     offset = rs + type_len;
@@ -566,9 +567,8 @@ void *CPUKernelWorkGroup::callArgs(std::vector<void *> &locals_to_free)
         CPUKernel::typeOffset(args_size, arg->valueSize() * arg->vecDim());
     }
 
-    rs = std::malloc(args_size);
-
-    if (!rs)
+    int retval = posix_memalign(&rs, 128, args_size);  // align for type double16 size.
+    if (retval || !rs)
         return NULL;
 
     size_t arg_offset = 0;
@@ -649,6 +649,10 @@ bool CPUKernelWorkGroup::run()
     std::vector<void *> locals_to_free;
     llvm::Function *kernel_func = p_kernel->callFunction();
 
+#if 0  // Let's see the stub's IR:
+    kernel_func->dump();
+#endif
+
     if (!kernel_func)
         return false;