PR tree-optimization/101895: Fold VEC_PERM to help recognize FMA.

This patch resolves PR tree-optimization/101895 a missed optimization
regression, by adding a costant 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

2022-03-15  Roger Sayle  <roger@nextmovesoftware.com>
	    Marc Glisse  <marc.glisse@inria.fr>
	    Richard Biener  <rguenther@suse.de>

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
	PR tree-optimization/101895
	* gcc.target/i386/pr101895.c: New test case.

diff --git a/gcc/match.pd b/gcc/match.pd
index 8b44b5c..7d619f4 100644 (file)
--- a/gcc/match.pd
+++ b/gcc/match.pd
@@ -7692,16 +7692,33 @@ and,
 /* VEC_PERM_EXPR (v, v, mask) -> v where v contains same element.  */
 
 (match vec_same_elem_p
+ (vec_duplicate @0))
+
+(match vec_same_elem_p
+ CONSTRUCTOR@0
+ (if (TREE_CODE (@0) == SSA_NAME
+      && uniform_vector_p (gimple_assign_rhs1 (SSA_NAME_DEF_STMT (@0))))))
+
+(match vec_same_elem_p
  @0
  (if (uniform_vector_p (@0))))
 
-(match vec_same_elem_p
- (vec_duplicate @0))
 
 (simplify
  (vec_perm vec_same_elem_p@0 @0 @1)
  @0)
 
+/* Push VEC_PERM earlier if that may help FMA perception (PR101895).  */
+(simplify
+ (plus:c (vec_perm:s (mult:c@0 @1 vec_same_elem_p@2) @0 @3) @4)
+ (if (TREE_CODE (@0) == SSA_NAME && num_imm_uses (@0) == 2)
+  (plus (mult (vec_perm @1 @1 @3) @2) @4)))
+(simplify
+ (minus (vec_perm:s (mult:c@0 @1 vec_same_elem_p@2) @0 @3) @4)
+ (if (TREE_CODE (@0) == SSA_NAME && num_imm_uses (@0) == 2)
+  (minus (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 (file)
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" } } */