This patch resolves PR tree-optimization/101895 a missed optimization
regression, by adding a constant folding simplification to match.pd to
simplify the transform "mult; vec_perm; plus" into "vec_perm; mult; plus"
with the aim that keeping the multiplication and addition next to each
other allows them to be recognized as fused-multiply-add on suitable
targets. This transformation requires a tweak to match.pd's
vec_same_elem_p predicate to handle CONSTRUCTOR_EXPRs using the same
SSA_NAME_DEF_STMT idiom used for constructors elsewhere in match.pd.
The net effect is that the following code example:
void foo(float * __restrict__ a, float b, float *c) {
a[0] = c[0]*b + a[0];
a[1] = c[2]*b + a[1];
a[2] = c[1]*b + a[2];
a[3] = c[3]*b + a[3];
}
when compiled on x86_64-pc-linux-gnu with -O2 -march=cascadelake
currently generates:
vbroadcastss %xmm0, %xmm0
vmulps (%rsi), %xmm0, %xmm0
vpermilps $216, %xmm0, %xmm0
vaddps (%rdi), %xmm0, %xmm0
vmovups %xmm0, (%rdi)
ret
but with this patch now generates the improved:
vpermilps $216, (%rsi), %xmm1
vbroadcastss %xmm0, %xmm0
vfmadd213ps (%rdi), %xmm0, %xmm1
vmovups %xmm1, (%rdi)
ret
This patch has been tested on x86_64-pc-linux-gnu with make bootstrap
and make -k check with no new failures. Ok for mainline?
2022-03-11 Roger Sayle <[email protected]>
gcc/ChangeLog
PR tree-optimization/101895
* match.pd (vec_same_elem_p): Handle CONSTRUCTOR_EXPR def.
(plus (vec_perm (mult ...) ...) ...): New reordering simplification.
gcc/testsuite/ChangeLog
* gcc.target/i386/pr101895.c: New test case.
Thanks in advance,
Roger
--
diff --git a/gcc/match.pd b/gcc/match.pd
index 97399e5..9184276 100644
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -7695,10 +7695,22 @@ and,
(match vec_same_elem_p
(vec_duplicate @0))
+(match vec_same_elem_p
+ CONSTRUCTOR@0
+ (if (uniform_vector_p (TREE_CODE (@0) == SSA_NAME
+ ? gimple_assign_rhs1 (SSA_NAME_DEF_STMT (@0)) : @0))))
+
(simplify
(vec_perm vec_same_elem_p@0 @0 @1)
@0)
+/* Push VEC_PERM earlier if that may help FMA perception (PR101895). */
+(for plusminus (plus minus)
+ (simplify
+ (plusminus (vec_perm (mult@0 @1 vec_same_elem_p@2) @0 @3) @4)
+ (plusminus (mult (vec_perm @1 @1 @3) @2) @4)))
+
+
/* Match count trailing zeroes for simplify_count_trailing_zeroes in fwprop.
The canonical form is array[((x & -x) * C) >> SHIFT] where C is a magic
constant which when multiplied by a power of 2 contains a unique value
diff --git a/gcc/testsuite/gcc.target/i386/pr101895.c
b/gcc/testsuite/gcc.target/i386/pr101895.c
new file mode 100644
index 0000000..4d0f1cb
--- /dev/null
+++ b/gcc/testsuite/gcc.target/i386/pr101895.c
@@ -0,0 +1,11 @@
+/* { dg-do compile } */
+/* { dg-options "-O2 -march=cascadelake" } */
+
+void foo(float * __restrict__ a, float b, float *c) {
+ a[0] = c[0]*b + a[0];
+ a[1] = c[2]*b + a[1];
+ a[2] = c[1]*b + a[2];
+ a[3] = c[3]*b + a[3];
+}
+
+/* { dg-final { scan-assembler "vfmadd" } } */
