From: Pan Li <pan2.li@intel.com>
This patch try to legitimise the const0_rtx (aka zero register)
as the base register for the RVV load/store instructions.
For example:
vint32m1_t test_vle32_v_i32m1_shortcut (size_t vl)
{
return __riscv_vle32_v_i32m1 ((int32_t *)0, vl);
}
Before this patch:
li a5,0
vsetvli zero,a1,e32,m1,ta,ma
vle32.v v24,0(a5) <- can propagate the const 0 to a5 here
vs1r.v v24,0(a0)
After this patch:
vsetvli zero,a1,e32,m1,ta,ma
vle32.v v24,0(zero)
vs1r.v v24,0(a0)
As above, this patch allow you to propagaate the const 0 (aka zero
register) to the base register of the RVV Unit-Stride load in the
combine pass. This may benefit the underlying RVV auto-vectorization.
However, the indexed load failed to perform the optimization and it
will be token care of in another PATCH.
gcc/ChangeLog:
* config/riscv/riscv.cc (riscv_classify_address): Allow
const0_rtx for the RVV load/store.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/base/zero_base_load_store_optimization.c: New test.
Signed-off-by: Pan Li <pan2.li@intel.com>
Co-authored-by: Ju-Zhe Zhong <juzhe.zhong@rivai.ai>
---
gcc/config/riscv/riscv.cc | 17 ++-
.../base/zero_base_load_store_optimization.c | 135 ++++++++++++++++++
2 files changed, 150 insertions(+), 2 deletions(-)
create mode 100644 gcc/testsuite/gcc.target/riscv/rvv/base/zero_base_load_store_optimization.c
@@ -1088,9 +1088,22 @@ riscv_classify_address (struct riscv_address_info *info, rtx x,
&& riscv_valid_lo_sum_p (info->symbol_type, mode, info->offset));
case CONST_INT:
- /* RVV load/store disallow CONST_INT. */
+ /* We only allow the const0_rtx for the RVV load/store. For example:
+ +----------------------------------------------------------+
+ | li a5,0 |
+ | vsetvli zero,a1,e32,m1,ta,ma |
+ | vle32.v v24,0(a5) <- propagate the const 0 to a5 here. |
+ | vs1r.v v24,0(a0) |
+ +----------------------------------------------------------+
+ It can be folded to:
+ +----------------------------------------------------------+
+ | vsetvli zero,a1,e32,m1,ta,ma |
+ | vle32.v v24,0(zero) |
+ | vs1r.v v24,0(a0) |
+ +----------------------------------------------------------+
+ This behavior will benefit the underlying RVV auto vectorization. */
if (riscv_v_ext_vector_mode_p (mode))
- return false;
+ return x == const0_rtx;
/* Small-integer addresses don't occur very often, but they
are legitimate if x0 is a valid base register. */
new file mode 100644
@@ -0,0 +1,135 @@
+/* { dg-do compile } */
+/* { dg-options "-march=rv64gcv -mabi=lp64 -O3" } */
+
+// #include <stdint-gcc.h>
+#include "riscv_vector.h"
+
+#define float32_t float
+
+// Unit-Stride Load/Store
+vint32m1_t test_vle32_v_i32m1_shortcut (size_t vl)
+{
+ return __riscv_vle32_v_i32m1 ((int32_t *)0, vl);
+}
+
+vuint32m1_t test_vle32_v_u32m1_shortcut (size_t vl)
+{
+ return __riscv_vle32_v_u32m1 ((int32_t *)0, vl);
+}
+
+vfloat32m1_t test_vle32_v_f32m1_shortcut (size_t vl)
+{
+ return __riscv_vle32_v_f32m1 ((float32_t *)0, vl);
+}
+
+void test_vse32_v_i32m1_shortcut (vint32m1_t val, size_t vl)
+{
+ __riscv_vse32_v_i32m1 ((int32_t *)0, val, vl);
+}
+
+void test_vse32_v_u32m1_shortcut (vuint32m1_t val, size_t vl)
+{
+ __riscv_vse32_v_u32m1 ((uint32_t *)0, val, vl);
+}
+
+void test_vse32_v_f32m1_shortcut (vfloat32m1_t val, size_t vl)
+{
+ __riscv_vse32_v_f32m1 ((float32_t *)0, val, vl);
+}
+
+// Stride Load/Store
+vint32m1_t test_vlse32_v_i32m1_shortcut (ptrdiff_t bstride, size_t vl)
+{
+ return __riscv_vlse32_v_i32m1 ((int32_t *)0, bstride, vl);
+}
+
+vuint32m1_t test_vlse32_v_u32m1_shortcut (ptrdiff_t bstride, size_t vl)
+{
+ return __riscv_vlse32_v_u32m1 ((uint32_t *)0, bstride, vl);
+}
+
+vfloat32m1_t test_vlse32_v_f32m1_shortcut (ptrdiff_t bstride, size_t vl)
+{
+ return __riscv_vlse32_v_f32m1 ((float32_t *)0, bstride, vl);
+}
+
+void test_vsse32_v_i32m1_shortcut (ptrdiff_t bstride, vint32m1_t val, size_t vl)
+{
+ __riscv_vsse32_v_i32m1 ((int32_t *)0, bstride, val, vl);
+}
+
+void test_vsse32_v_u32m1_shortcut (ptrdiff_t bstride, vuint32m1_t val, size_t vl)
+{
+ __riscv_vsse32_v_u32m1 ((uint32_t *)0, bstride, val, vl);
+}
+
+void test_vsse32_v_f32m1_shortcut (ptrdiff_t bstride, vfloat32m1_t val,size_t vl)
+{
+ __riscv_vsse32_v_f32m1 ((float32_t *)0, bstride, val, vl);
+}
+
+// Indexed-Unordered Load/Store
+vint32m1_t test_vluxei32_v_i32m1_shortcut (vuint32m1_t bindex, size_t vl)
+{
+ return __riscv_vluxei32_v_i32m1 ((int32_t *)0, bindex, vl);
+}
+
+vuint32m1_t test_vluxei32_v_u32m1_shortcut (vuint32m1_t bindex, size_t vl)
+{
+ return __riscv_vluxei32_v_u32m1 ((uint32_t *)0, bindex, vl);
+}
+
+vfloat32m1_t test_vluxei32_v_f32m1_shortcut (vuint32m1_t bindex, size_t vl)
+{
+ return __riscv_vluxei32_v_f32m1 ((float32_t *)0, bindex, vl);
+}
+
+void test_vsuxei32_v_i32m1_shortcut (vuint32m1_t bindex, vint32m1_t val, size_t vl)
+{
+ __riscv_vsuxei32_v_i32m1 ((int32_t *)0, bindex, val, vl);
+}
+
+void test_vsuxei32_v_u32m1_shortcut (vuint32m1_t bindex, vuint32m1_t val, size_t vl)
+{
+ __riscv_vsuxei32_v_u32m1 ((uint32_t *)0, bindex, val, vl);
+}
+
+void test_vsuxei32_v_f32m1_shortcut (vuint32m1_t bindex, vfloat32m1_t val, size_t vl)
+{
+ __riscv_vsuxei32_v_f32m1 ((float32_t *)0, bindex, val, vl);
+}
+
+// Indexed-Ordered Load/Store
+vint32m1_t test_vloxei32_v_i32m1_shortcut (vuint32m1_t bindex, size_t vl)
+{
+ return __riscv_vloxei32_v_i32m1 ((int32_t *)0, bindex, vl);
+}
+
+vuint32m1_t test_vloxei32_v_u32m1_shortcut (vuint32m1_t bindex, size_t vl)
+{
+ return __riscv_vloxei32_v_u32m1 ((uint32_t *)0, bindex, vl);
+}
+
+vfloat32m1_t test_vloxei32_v_f32m1_shortcut (vuint32m1_t bindex, size_t vl)
+{
+ return __riscv_vloxei32_v_f32m1 ((float32_t *)0, bindex, vl);
+}
+
+void test_vsoxei32_v_i32m1_shortcut (vuint32m1_t bindex, vint32m1_t val, size_t vl)
+{
+ __riscv_vsoxei32_v_i32m1 ((int32_t *)0, bindex, val, vl);
+}
+
+void test_vsoxei32_v_u32m1_shortcut (vuint32m1_t bindex, vuint32m1_t val, size_t vl)
+{
+ __riscv_vsoxei32_v_u32m1 ((uint32_t *)0, bindex, val, vl);
+}
+
+void test_vsoxei32_v_f32m1_shortcut (vuint32m1_t bindex, vfloat32m1_t val, size_t vl)
+{
+ __riscv_vsoxei32_v_f32m1 ((float32_t *)0, bindex, val, vl);
+}
+
+/* { dg-final { scan-assembler-times {v[ls]e[0-9]+\.v\s+v[0-9]+,\s*0\(zero\)} 6 } } */
+/* { dg-final { scan-assembler-times {v[ls]se[0-9]+\.v\s+v[0-9]+,\s*0\(zero\),\s*[ax][0-9]+} 6 } } */
+/* { dg-final { scan-assembler-times {li\s+[a-x][0-9]+,\s*0} 12 } } */