[PATCH 0/4] builtin: teach sparse to linearize __builtin_fma()

[PATCH 0/4] builtin: teach sparse to linearize __builtin_fma()

[Luc Van Oostenryck]

This series is the second part experimenting with adding the
infrastructure needed for the linearization of builtins.

The first part added support for __builtin_unreachable()
but this builtin has no arguments and no return value.
Now, there is an example showing how to do it when arguments
and/or a return value are present.

Luc Van Oostenryck (4):
  add support for a new instruction: OP_FMA
  builtin: allow linearization to fail
  builtin: add declaration for __builtin_fma{,f,l}()
  builtin: teach sparse to linearize __builtin_fma()

 builtin.c   |  3 +++
 linearize.c | 29 +++++++++++++++++++++++++++--
 opcode.def  |  1 +
 3 files changed, 31 insertions(+), 2 deletions(-)

-- 
2.28.0

[PATCH 1/4] add support for a new instruction: OP_FMA

[Luc Van Oostenryck]

This will be the instruction for fused multiply-add
but the motivation for it is some experimentation with
the linearization of builtins.

Signed-off-by: Luc Van Oostenryck <luc.vanoostenryck@gmail.com>
---
 linearize.c | 2 ++
 opcode.def  | 1 +
 2 files changed, 3 insertions(+)

diff --git a/linearize.c b/linearize.c
index 5a8e74970d98..1a2677713123 100644
--- a/linearize.c
+++ b/linearize.c
@@ -244,6 +244,7 @@ static const char *opcodes[] = {
 
 	/* Special three-input */
 	[OP_SEL] = "select",
+	[OP_FMA] = "fma",
 	
 	/* Memory */
 	[OP_LOAD] = "load",
@@ -461,6 +462,7 @@ const char *show_instruction(struct instruction *insn)
 		break;
 
 	case OP_SEL:
+	case OP_FMA:
 		buf += sprintf(buf, "%s <- %s, %s, %s", show_pseudo(insn->target),
 			show_pseudo(insn->src1), show_pseudo(insn->src2), show_pseudo(insn->src3));
 		break;
diff --git a/opcode.def b/opcode.def
index 2583e2f4a602..58fe9983cbc2 100644
--- a/opcode.def
+++ b/opcode.def
@@ -91,6 +91,7 @@ OPCODE(SLICE,           BADOP,    BADOP,    BADOP, 1, OPF_TARGET)
 
 /* Select - three input values */
 OPCODE(SEL,             BADOP,    BADOP,    BADOP, 3, OPF_TARGET)
+OPCODE(FMA,             BADOP,    BADOP,    BADOP, 3, OPF_TARGET)
 
 /* Memory */
 OPCODE(LOAD,            BADOP,    BADOP,    BADOP, 1, OPF_TARGET)
-- 
2.28.0

[PATCH 2/4] builtin: allow linearization to fail

[Luc Van Oostenryck]

Allow the linearization of builtins to fail and continue with
the normal linearization of OP_CALLs. The motivation for this
is for the linearization of target specific builtins.

Signed-off-by: Luc Van Oostenryck <luc.vanoostenryck@gmail.com>
---
 linearize.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)

diff --git a/linearize.c b/linearize.c
index 1a2677713123..bf55045bcbce 100644
--- a/linearize.c
+++ b/linearize.c
@@ -1513,8 +1513,11 @@ static pseudo_t linearize_call_expression(struct entrypoint *ep, struct expressi
 	fntype = fn->ctype;
 
 	// handle builtins
-	if (fntype->op && fntype->op->linearize)
-		return fntype->op->linearize(ep, expr);
+	if (fntype->op && fntype->op->linearize) {
+		retval = fntype->op->linearize(ep, expr);
+		if (retval)
+			return retval;
+	}
 
 	ctype = &fntype->ctype;
 	if (fntype->type == SYM_NODE)
-- 
2.28.0

[PATCH 3/4] builtin: add declaration for __builtin_fma{,f,l}()

[Luc Van Oostenryck]

The motivation for this is to experiment with adding infrastructure
for the linearization of builtins.

Signed-off-by: Luc Van Oostenryck <luc.vanoostenryck@gmail.com>
---
 builtin.c | 3 +++
 1 file changed, 3 insertions(+)

diff --git a/builtin.c b/builtin.c
index 2e9be8be8adb..26b612dc401b 100644
--- a/builtin.c
+++ b/builtin.c
@@ -490,6 +490,9 @@ static const struct builtin_fn builtins_common[] = {
 	{ "__builtin_ffs", &int_ctype, 0, { &int_ctype }, .op = &ffs_op },
 	{ "__builtin_ffsl", &int_ctype, 0, { &long_ctype }, .op = &ffs_op },
 	{ "__builtin_ffsll", &int_ctype, 0, { &llong_ctype }, .op = &ffs_op },
+	{ "__builtin_fma", &double_ctype, 0, { &double_ctype, &double_ctype, &double_ctype }},
+	{ "__builtin_fmaf", &float_ctype, 0, { &float_ctype, &float_ctype, &float_ctype }},
+	{ "__builtin_fmal", &ldouble_ctype, 0, { &ldouble_ctype, &ldouble_ctype, &ldouble_ctype }},
 	{ "__builtin_frame_address", &ptr_ctype, 0, { &uint_ctype }},
 	{ "__builtin_free", &void_ctype, 0, { &ptr_ctype }},
 	{ "__builtin_huge_val", &double_ctype, 0 },
-- 
2.28.0

[PATCH 4/4] builtin: teach sparse to linearize __builtin_fma()

[Luc Van Oostenryck]

The support for the linearization of builtins was already
added for __builtin_unreachable() but this builtin has
no arguments and no return value.

So, to complete the experience of builtin linearization,
add the linearization of __builtin_fma().

Signed-off-by: Luc Van Oostenryck <luc.vanoostenryck@gmail.com>
---
 linearize.c | 20 ++++++++++++++++++++
 1 file changed, 20 insertions(+)

diff --git a/linearize.c b/linearize.c
index bf55045bcbce..7157ba511671 100644
--- a/linearize.c
+++ b/linearize.c
@@ -2585,6 +2585,23 @@ struct entrypoint *linearize_symbol(struct symbol *sym)
  * Builtin functions
  */
 
+static pseudo_t linearize_fma(struct entrypoint *ep, struct expression *expr)
+{
+	struct instruction *insn = alloc_typed_instruction(OP_FMA, expr->ctype);
+	struct expression *arg;
+
+	PREPARE_PTR_LIST(expr->args, arg);
+		insn->src1 = linearize_expression(ep, arg);
+		NEXT_PTR_LIST(arg)
+		insn->src2 = linearize_expression(ep, arg);
+		NEXT_PTR_LIST(arg)
+		insn->src3 = linearize_expression(ep, arg);
+	FINISH_PTR_LIST(arg);
+
+	add_one_insn(ep, insn);
+	return insn->target = alloc_pseudo(insn);
+}
+
 static pseudo_t linearize_unreachable(struct entrypoint *ep, struct expression *exp)
 {
 	add_unreachable(ep);
@@ -2597,6 +2614,9 @@ static struct sym_init {
 	struct symbol_op op;
 } builtins_table[] = {
 	// must be declared in builtin.c:declare_builtins[]
+	{ "__builtin_fma", linearize_fma },
+	{ "__builtin_fmaf", linearize_fma },
+	{ "__builtin_fmal", linearize_fma },
 	{ "__builtin_unreachable", linearize_unreachable },
 	{ }
 };
-- 
2.28.0