Hi,
the patch below adds support for a few unary floating point buitlins
that can be implemented with a single HSA instruction. More effort in
the area of builtins is needed, the motivation for this was a a
benchmark that previously failed with a sorry message. Committed to
the hsa branch.
Martin
2015-08-25 Martin Jambor <mjam...@suse.cz>
* hsa-gen.c (gen_hsa_unaryop_for_builtin): New function.
(gen_hsa_insns_for_call): Add support for a few unary fp
operations.
---
gcc/ChangeLog.hsa | 4 +++
gcc/hsa-gen.c | 86 +++++++++++++++++++++++++++++++++++++++++++++++--------
2 files changed, 78 insertions(+), 12 deletions(-)
diff --git a/gcc/hsa-gen.c b/gcc/hsa-gen.c
index 14cf890..1e23996 100644
--- a/gcc/hsa-gen.c
+++ b/gcc/hsa-gen.c
@@ -3213,6 +3213,26 @@ gen_hsa_insns_for_kernel_call (hsa_bb *hbb, gcall *call)
hsa_cfun->kernel_dispatch_count++;
}
+/* Helper functions to create a single unary HSA operations out of calls to
+ builtins. OPCODE is the HSA operation to be generated. STMT is a gimple
+ call to a builtin. HBB is the HSA BB to which the instruction should be
+ added and SSA_MAP is used to map gimple SSA names to HSA pseudoreisters. */
+
+static void
+gen_hsa_unaryop_for_builtin (int opcode, gimple stmt, hsa_bb *hbb,
+ vec <hsa_op_reg_p> *ssa_map)
+{
+ tree lhs = gimple_call_lhs (stmt);
+ /* FIXME: Since calls without a LHS are not removed, double check that
+ they cannot have side effects. */
+ if (!lhs)
+ return;
+ hsa_op_reg *dest = hsa_reg_for_gimple_ssa (lhs, ssa_map);
+ hsa_op_base *op = hsa_reg_or_immed_for_gimple_op (gimple_call_arg (stmt, 0),
+ hbb, ssa_map, NULL);
+ gen_hsa_unary_operation (opcode, dest, op, hbb);
+}
+
/* Generate HSA instructions for the given call statement STMT. Instructions
will be appended to HBB. SSA_MAP maps gimple SSA names to HSA pseudo
registers. */
@@ -3284,22 +3304,64 @@ specialop:
break;
}
+ case BUILT_IN_FABS:
+ case BUILT_IN_FABSF:
+ gen_hsa_unaryop_for_builtin (BRIG_OPCODE_ABS, stmt, hbb, ssa_map);
+ break;
+
+ case BUILT_IN_CEIL:
+ case BUILT_IN_CEILF:
+ gen_hsa_unaryop_for_builtin (BRIG_OPCODE_CEIL, stmt, hbb, ssa_map);
+ break;
+
+ case BUILT_IN_FLOOR:
+ case BUILT_IN_FLOORF:
+ gen_hsa_unaryop_for_builtin (BRIG_OPCODE_FLOOR, stmt, hbb, ssa_map);
+ break;
+
+ case BUILT_IN_RINT:
+ case BUILT_IN_RINTF:
+ gen_hsa_unaryop_for_builtin (BRIG_OPCODE_RINT, stmt, hbb, ssa_map);
+ break;
+
case BUILT_IN_SQRT:
case BUILT_IN_SQRTF:
- /* FIXME: Since calls without a LHS are not removed, double check that
- they cannot have side effects. */
- if (!lhs)
- return;
- dest = hsa_reg_for_gimple_ssa (lhs, ssa_map);
- insn = new hsa_insn_basic (2, BRIG_OPCODE_SQRT, dest->type);
- insn->operands[0] = dest;
- dest->set_definition (insn);
- insn->operands[1]
- = hsa_reg_or_immed_for_gimple_op (gimple_call_arg (stmt, 0),
- hbb, ssa_map, insn);
- hbb->append_insn (insn);
+ /* TODO: Perhaps produce BRIG_OPCODE_NSQRT with -ffast-math? */
+ gen_hsa_unaryop_for_builtin (BRIG_OPCODE_SQRT, stmt, hbb, ssa_map);
+ break;
+
+ case BUILT_IN_TRUNC:
+ case BUILT_IN_TRUNCF:
+ gen_hsa_unaryop_for_builtin (BRIG_OPCODE_TRUNC, stmt, hbb, ssa_map);
break;
+ case BUILT_IN_COS:
+ case BUILT_IN_COSF:
+ /* FIXME: Using the native instruction may not be precise enough.
+ Perhaps only allow if using -ffast-math? */
+ gen_hsa_unaryop_for_builtin (BRIG_OPCODE_NCOS, stmt, hbb, ssa_map);
+ break;
+
+ case BUILT_IN_EXP2:
+ case BUILT_IN_EXP2F:
+ /* FIXME: Using the native instruction may not be precise enough.
+ Perhaps only allow if using -ffast-math? */
+ gen_hsa_unaryop_for_builtin (BRIG_OPCODE_NEXP2, stmt, hbb, ssa_map);
+ break;
+
+ case BUILT_IN_LOG2:
+ case BUILT_IN_LOG2F:
+ /* FIXME: Using the native instruction may not be precise enough.
+ Perhaps only allow if using -ffast-math? */
+ gen_hsa_unaryop_for_builtin (BRIG_OPCODE_NLOG2, stmt, hbb, ssa_map);
+ break;
+
+ case BUILT_IN_SIN:
+ case BUILT_IN_SINF:
+ /* FIXME: Using the native instruction may not be precise enough.
+ Perhaps only allow if using -ffast-math? */
+ gen_hsa_unaryop_for_builtin (BRIG_OPCODE_NSIN, stmt, hbb, ssa_map);
+
case BUILT_IN_ATOMIC_LOAD_1:
case BUILT_IN_ATOMIC_LOAD_2:
case BUILT_IN_ATOMIC_LOAD_4:
--
2.4.6
