RISC-V: Lower vmv.v.x (avl = 1) into vmv.s.x

Notice there is a AI benchmark, GCC vs Clang has 3% performance drop.

It's because Clang/LLVM has a simplification transform vmv.v.x (avl = 1) into vmv.s.x.

Since vmv.s.x has more flexible vsetvl demand than vmv.v.x that can allow us to have
better chances to fuse vsetvl.

Consider this following case:

void
foo (uint32_t *outputMat, uint32_t *inputMat)
{
  vuint32m1_t matRegIn0 = __riscv_vle32_v_u32m1 (inputMat, 4);
  vuint32m1_t matRegIn1 = __riscv_vle32_v_u32m1 (inputMat + 4, 4);
  vuint32m1_t matRegIn2 = __riscv_vle32_v_u32m1 (inputMat + 8, 4);
  vuint32m1_t matRegIn3 = __riscv_vle32_v_u32m1 (inputMat + 12, 4);

  vbool32_t oddMask
    = __riscv_vreinterpret_v_u32m1_b32 (__riscv_vmv_v_x_u32m1 (0xaaaa, 1));

  vuint32m1_t smallTransposeMat0
    = __riscv_vslideup_vx_u32m1_tumu (oddMask, matRegIn0, matRegIn1, 1, 4);
  vuint32m1_t smallTransposeMat2
    = __riscv_vslideup_vx_u32m1_tumu (oddMask, matRegIn2, matRegIn3, 1, 4);

  vuint32m1_t outMat0 = __riscv_vslideup_vx_u32m1_tu (smallTransposeMat0,
						      smallTransposeMat2, 2, 4);

  __riscv_vse32_v_u32m1 (outputMat, outMat0, 4);
}

Before this patch:

        vsetivli        zero,4,e32,m1,ta,ma
        li      a5,45056
        addi    a2,a1,16
        addi    a3,a1,32
        addi    a4,a1,48
        vle32.v v1,0(a1)
        vle32.v v4,0(a2)
        vle32.v v2,0(a3)
        vle32.v v3,0(a4)
        addiw   a5,a5,-1366
        vsetivli        zero,1,e32,m1,ta,ma
        vmv.v.x v0,a5                         ---> Since it avl = 1, we can transform it into vmv.s.x
        vsetivli        zero,4,e32,m1,tu,mu
        vslideup.vi     v1,v4,1,v0.t
        vslideup.vi     v2,v3,1,v0.t
        vslideup.vi     v1,v2,2
        vse32.v v1,0(a0)
        ret

After this patch:

	li	a5,45056
	addi	a2,a1,16
	vsetivli	zero,4,e32,m1,tu,mu
	addiw	a5,a5,-1366
	vle32.v	v3,0(a2)
	addi	a3,a1,32
	addi	a4,a1,48
	vle32.v	v1,0(a1)
	vmv.s.x	v0,a5
	vle32.v	v2,0(a3)
	vslideup.vi	v1,v3,1,v0.t
	vle32.v	v3,0(a4)
	vslideup.vi	v2,v3,1,v0.t
	vslideup.vi	v1,v2,2
	vse32.v	v1,0(a0)
	ret

Tested on both RV32 and RV64 no regression.

gcc/ChangeLog:

	* config/riscv/riscv-protos.h (splat_to_scalar_move_p): New function.
	* config/riscv/riscv-v.cc (splat_to_scalar_move_p): Ditto.
	* config/riscv/vector.md: Simplify vmv.v.x. into vmv.s.x.

gcc/testsuite/ChangeLog:

	* gcc.target/riscv/rvv/vsetvl/attribute-2.c: New test.
	* gcc.target/riscv/rvv/vsetvl/attribute-3.c: New test.

5 files changed, 94 insertions, 1 deletions

diff --git a/gcc/config/riscv/riscv-protos.h b/gcc/config/riscv/riscv-protos.h
index 7fe26fc..b3f0bdb 100644
--- a/gcc/config/riscv/riscv-protos.h
+++ b/gcc/config/riscv/riscv-protos.h
@@ -708,6 +708,7 @@ bool can_be_broadcasted_p (rtx);
 bool gather_scatter_valid_offset_p (machine_mode);
 HOST_WIDE_INT estimated_poly_value (poly_int64, unsigned int);
 bool whole_reg_to_reg_move_p (rtx *, machine_mode, int);
+bool splat_to_scalar_move_p (rtx *);
 }
 
 /* We classify builtin types into two classes:
diff --git a/gcc/config/riscv/riscv-v.cc b/gcc/config/riscv/riscv-v.cc
index 93a1238..4bacb7f 100644
--- a/gcc/config/riscv/riscv-v.cc
+++ b/gcc/config/riscv/riscv-v.cc
@@ -5151,4 +5151,16 @@ whole_reg_to_reg_move_p (rtx *ops, machine_mode mode, int avl_type_index)
   return false;
 }
 
+/* Return true if we can transform vmv.v.x/vfmv.v.f to vmv.s.x/vfmv.s.f.  */
+bool
+splat_to_scalar_move_p (rtx *ops)
+{
+  return satisfies_constraint_Wc1 (ops[1])
+	 && satisfies_constraint_vu (ops[2])
+	 && !MEM_P (ops[3])
+	 && satisfies_constraint_c01 (ops[4])
+	 && INTVAL (ops[7]) == NONVLMAX
+	 && known_ge (GET_MODE_SIZE (Pmode), GET_MODE_SIZE (GET_MODE (ops[3])));
+}
+
 } // namespace riscv_vector
diff --git a/gcc/config/riscv/vector.md b/gcc/config/riscv/vector.md
index 307d9a8..ab6e099 100644
--- a/gcc/config/riscv/vector.md
+++ b/gcc/config/riscv/vector.md
@@ -1977,8 +1977,15 @@
 	  (match_operand:V_VLS 2 "vector_merge_operand")))]
   "TARGET_VECTOR"
 {
+  /* Transform vmv.v.x/vfmv.v.f (avl = 1) into vmv.s.x since vmv.s.x/vfmv.s.f
+     has better chances to do vsetvl fusion in vsetvl pass.  */
+  if (riscv_vector::splat_to_scalar_move_p (operands))
+    {
+      operands[1] = riscv_vector::gen_scalar_move_mask (<VM>mode);
+      operands[3] = force_reg (<VEL>mode, operands[3]);
+    }
   /* Handle vmv.s.x instruction (Wb1 mask) which has memory scalar.  */
-  if (satisfies_constraint_Wdm (operands[3]))
+  else if (satisfies_constraint_Wdm (operands[3]))
     {
       if (satisfies_constraint_Wb1 (operands[1]))
 	{
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/vsetvl/attribute-2.c b/gcc/testsuite/gcc.target/riscv/rvv/vsetvl/attribute-2.c
new file mode 100644
index 0000000..b3fec26
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/rvv/vsetvl/attribute-2.c
@@ -0,0 +1,37 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gcv -mabi=lp64d -O3" } */
+
+#include "riscv_vector.h"
+
+void
+foo (uint32_t *outputMat, uint32_t *inputMat)
+{
+  vuint32m1_t matRegIn0 = __riscv_vle32_v_u32m1 (inputMat, 4);
+  vuint32m1_t matRegIn1 = __riscv_vle32_v_u32m1 (inputMat + 4, 4);
+  vuint32m1_t matRegIn2 = __riscv_vle32_v_u32m1 (inputMat + 8, 4);
+  vuint32m1_t matRegIn3 = __riscv_vle32_v_u32m1 (inputMat + 12, 4);
+
+  vbool32_t oddMask
+    = __riscv_vreinterpret_v_u32m1_b32 (__riscv_vmv_v_x_u32m1 (0xaaaa, 1));
+
+  vuint32m1_t smallTransposeMat0
+    = __riscv_vslideup_vx_u32m1_tumu (oddMask, matRegIn0, matRegIn1, 1, 4);
+  vuint32m1_t smallTransposeMat2
+    = __riscv_vslideup_vx_u32m1_tumu (oddMask, matRegIn2, matRegIn3, 1, 4);
+
+  vuint32m1_t outMat0 = __riscv_vslideup_vx_u32m1_tu (smallTransposeMat0,
+						      smallTransposeMat2, 2, 4);
+
+  __riscv_vse32_v_u32m1 (outputMat, outMat0, 4);
+}
+
+void
+foo2 (void *outputMat, void *inputMat)
+{
+  vfloat32m1_t v = __riscv_vfmv_v_f_f32m1 (0xaaaa, 1);
+  __riscv_vse32_v_f32m1 (outputMat, v, 4);
+}
+
+/* { dg-final { scan-assembler-times {vsetivli\s+zero,\s*4,\s*e32,\s*m1,\s*t[au],\s*m[au]} 2 } } */
+/* { dg-final { scan-assembler-times {vsetivli} 2 } } */
+/* { dg-final { scan-assembler-not {vsetvli} } } */
diff --git a/gcc/testsuite/gcc.target/riscv/rvv/vsetvl/attribute-3.c b/gcc/testsuite/gcc.target/riscv/rvv/vsetvl/attribute-3.c
new file mode 100644
index 0000000..643f6a9
--- /dev/null
+++ b/gcc/testsuite/gcc.target/riscv/rvv/vsetvl/attribute-3.c
@@ -0,0 +1,36 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gcv -mabi=lp64d -O3 -fno-schedule-insns -fno-schedule-insns2" } */
+
+#include "riscv_vector.h"
+
+void matrix_transpose_in_register(uint32_t* outputMat, uint32_t* inputMat) {
+    vuint32m1_t matRegIn0 = __riscv_vle32_v_u32m1(inputMat, 4);
+    vuint32m1_t matRegIn1 = __riscv_vle32_v_u32m1(inputMat + 4, 4);
+    vuint32m1_t matRegIn2 = __riscv_vle32_v_u32m1(inputMat + 8, 4);
+    vuint32m1_t matRegIn3 = __riscv_vle32_v_u32m1(inputMat + 12, 4);
+
+    vbool32_t oddMask = __riscv_vreinterpret_v_u32m1_b32(__riscv_vmv_v_x_u32m1(0xaaaa, 1));
+
+    vuint32m1_t smallTransposeMat0 = __riscv_vslideup_vx_u32m1_tumu(oddMask, matRegIn0, matRegIn1, 1, 4);
+    vuint32m1_t smallTransposeMat2 = __riscv_vslideup_vx_u32m1_tumu(oddMask, matRegIn2, matRegIn3, 1, 4);
+
+    vbool32_t evenMask = __riscv_vreinterpret_v_u32m1_b32(__riscv_vmv_v_x_u32m1(0x5555, 1));
+
+    vuint32m1_t smallTransposeMat1 = __riscv_vslidedown_vx_u32m1_tumu(evenMask, matRegIn1, matRegIn0, 1, 4);
+    vuint32m1_t smallTransposeMat3 = __riscv_vslidedown_vx_u32m1_tumu(evenMask, matRegIn3, matRegIn2, 1, 4);
+
+    vuint32m1_t outMat0 = __riscv_vslideup_vx_u32m1_tu(smallTransposeMat0, smallTransposeMat2, 2, 4);
+    vuint32m1_t outMat1 = __riscv_vslideup_vx_u32m1_tu(smallTransposeMat1, smallTransposeMat3, 2, 4);
+
+    vuint32m1_t outMat2 = __riscv_vslidedown_vx_u32m1_tu(smallTransposeMat2, smallTransposeMat0, 2, 2);
+    vuint32m1_t outMat3 = __riscv_vslidedown_vx_u32m1_tu(smallTransposeMat3, smallTransposeMat1, 2, 2);
+    __riscv_vse32_v_u32m1(outputMat, outMat0, 4);
+    __riscv_vse32_v_u32m1(outputMat + 4, outMat1, 4);
+    __riscv_vse32_v_u32m1(outputMat + 8, outMat2, 4);
+    __riscv_vse32_v_u32m1(outputMat + 12, outMat3, 4);
+}
+
+/* { dg-final { scan-assembler-times {vsetivli\s+zero,\s*4,\s*e32,\s*m1,\s*t[au],\s*m[au]} 2 } } */
+/* { dg-final { scan-assembler-times {vsetivli\s+zero,\s*2,\s*e32,\s*m1,\s*t[au],\s*m[au]} 1 } } */
+/* { dg-final { scan-assembler-times {vsetivli} 3 } } */
+/* { dg-final { scan-assembler-not {vsetvli} } } */