This patch is adding an obvious missing mult_high auto-vectorization pattern.
Consider this following case:
void __attribute__ ((noipa)) \
mod_##TYPE (TYPE *__restrict dst, TYPE *__restrict src, int count) \
{ \
for (int i = 0; i < count; ++i) \
dst[i] = src[i] / 17; \
}
T (int32_t) \
TEST_ALL (DEF_LOOP)
Before this patch:
mod_int32_t:
ble a2,zero,.L5
li a5,17
vsetvli a3,zero,e32,m1,ta,ma
vmv.v.x v2,a5
.L3:
vsetvli a5,a2,e8,mf4,ta,ma
vle32.v v1,0(a1)
vsetvli a3,zero,e32,m1,ta,ma
slli a4,a5,2
vdiv.vv v1,v1,v2
sub a2,a2,a5
vsetvli zero,a5,e32,m1,ta,ma
vse32.v v1,0(a0)
add a1,a1,a4
add a0,a0,a4
bne a2,zero,.L3
.L5:
ret
After this patch:
mod_int32_t:
ble a2,zero,.L5
li a5,
2021163008
addiw a5,a5,-1927
vsetvli a3,zero,e32,m1,ta,ma
vmv.v.x v3,a5
.L3:
vsetvli a5,a2,e8,mf4,ta,ma
vle32.v v2,0(a1)
vsetvli a3,zero,e32,m1,ta,ma
slli a4,a5,2
vmulh.vv v1,v2,v3
sub a2,a2,a5
vsra.vi v2,v2,31
vsra.vi v1,v1,3
vsub.vv v1,v1,v2
vsetvli zero,a5,e32,m1,ta,ma
vse32.v v1,0(a0)
add a1,a1,a4
add a0,a0,a4
bne a2,zero,.L3
.L5:
ret
Even though a single "vdiv" is lower into "1 vmulh + 2 vsra + 1 vsub",
4 more instructions are generated, we belive it's much better than before
since division is very slow in the hardward.
gcc/ChangeLog:
* config/riscv/autovec.md (smul<mode>3_highpart): New pattern.
(umul<mode>3_highpart): Ditto.
gcc/testsuite/ChangeLog:
* gcc.target/riscv/rvv/autovec/binop/mulh-1.c: New test.
* gcc.target/riscv/rvv/autovec/binop/mulh-2.c: New test.
* gcc.target/riscv/rvv/autovec/binop/mulh_run-1.c: New test.
* gcc.target/riscv/rvv/autovec/binop/mulh_run-2.c: New test.
riscv_vector::RVV_BINOP, operands);
DONE;
})
+
+;; -------------------------------------------------------------------------
+;; ---- [INT] Highpart multiplication
+;; -------------------------------------------------------------------------
+;; Includes:
+;; - vmulh.vv
+;; - vmulhu.vv
+;; -------------------------------------------------------------------------
+
+(define_expand "smul<mode>3_highpart"
+ [(match_operand:VFULLI 0 "register_operand")
+ (match_operand:VFULLI 1 "register_operand")
+ (match_operand:VFULLI 2 "register_operand")]
+ "TARGET_VECTOR"
+{
+ insn_code icode = code_for_pred_mulh (UNSPEC_VMULHS, <MODE>mode);
+ riscv_vector::emit_vlmax_insn (icode, riscv_vector::RVV_BINOP, operands);
+ DONE;
+})
+
+(define_expand "umul<mode>3_highpart"
+ [(match_operand:VFULLI 0 "register_operand")
+ (match_operand:VFULLI 1 "register_operand")
+ (match_operand:VFULLI 2 "register_operand")]
+ "TARGET_VECTOR"
+{
+ insn_code icode = code_for_pred_mulh (UNSPEC_VMULHU, <MODE>mode);
+ riscv_vector::emit_vlmax_insn (icode, riscv_vector::RVV_BINOP, operands);
+ DONE;
+})
--- /dev/null
+/* { dg-do compile } */
+/* { dg-additional-options "-march=rv32gcv -mabi=ilp32d --param=riscv-autovec-preference=scalable -fno-vect-cost-model" } */
+
+#include <stdint-gcc.h>
+
+#define DEF_LOOP(TYPE) \
+ void __attribute__ ((noipa)) mod_##TYPE (TYPE *dst, TYPE *src, int count) \
+ { \
+ for (int i = 0; i < count; ++i) \
+ dst[i] = src[i] % 19; \
+ }
+
+#define TEST_ALL(T) \
+ T (int8_t) \
+ T (uint8_t) \
+ T (int16_t) \
+ T (uint16_t) \
+ T (int32_t) \
+ T (uint32_t) \
+ T (int64_t) \
+ T (uint64_t)
+
+TEST_ALL (DEF_LOOP)
+
+/* { dg-final { scan-assembler-times {\tvmulh\.vv} 4 } } */
+/* { dg-final { scan-assembler-times {\tvmulhu\.vv} 4 } } */
--- /dev/null
+/* { dg-do compile } */
+/* { dg-additional-options "-march=rv32gcv -mabi=ilp32d --param=riscv-autovec-preference=scalable -fno-vect-cost-model" } */
+
+#include <stdint-gcc.h>
+
+#define DEF_LOOP(TYPE) \
+void __attribute__ ((noipa)) \
+mod_##TYPE (TYPE *dst, TYPE *src, int count) \
+{ \
+ for (int i = 0; i < count; ++i) \
+ dst[i] = src[i] / 17; \
+}
+
+#define TEST_ALL(T) \
+ T (int8_t) \
+ T (uint8_t) \
+ T (int16_t) \
+ T (uint16_t) \
+ T (int32_t) \
+ T (uint32_t) \
+ T (int64_t) \
+ T (uint64_t)
+
+TEST_ALL (DEF_LOOP)
+
+/* { dg-final { scan-assembler-times {\tvmulh\.vv} 4 } } */
+/* { dg-final { scan-assembler-times {\tvmulhu\.vv} 4 } } */
--- /dev/null
+/* { dg-do run { target { riscv_vector } } } */
+/* { dg-additional-options "--param=riscv-autovec-preference=scalable" } */
+
+#include "mulh-1.c"
+
+#define N 79
+
+#define TEST_LOOP(TYPE) \
+ { \
+ TYPE dst[N], src[N]; \
+ for (int i = 0; i < N; ++i) \
+ { \
+ src[i] = i * 7 + i % 3; \
+ if (i % 11 > 7) \
+ src[i] = -src[i]; \
+ asm volatile ("" ::: "memory"); \
+ } \
+ mod_##TYPE (dst, src, N); \
+ for (int i = 0; i < N; ++i) \
+ if (dst[i] != src[i] % 19) \
+ __builtin_abort (); \
+ }
+
+int
+main (void)
+{
+ TEST_ALL (TEST_LOOP);
+ return 0;
+}
--- /dev/null
+/* { dg-do run { target { riscv_vector } } } */
+/* { dg-additional-options "--param=riscv-autovec-preference=scalable" } */
+
+#include "mulh-2.c"
+
+#define N 79
+
+#define TEST_LOOP(TYPE) \
+ { \
+ TYPE dst[N], src[N]; \
+ for (int i = 0; i < N; ++i) \
+ { \
+ src[i] = i * 7 + i % 3; \
+ if (i % 11 > 7) \
+ src[i] = -src[i]; \
+ asm volatile ("" ::: "memory"); \
+ } \
+ mod_##TYPE (dst, src, N); \
+ for (int i = 0; i < N; ++i) \
+ if (dst[i] != src[i] / 17) \
+ __builtin_abort (); \
+ }
+
+int
+main (void)
+{
+ TEST_ALL (TEST_LOOP);
+ return 0;
+}