[gcc r17-3247] RISC-V: Add test cases for vwmacc.vv reg overlap
Pan Li via Gcc-cvs <[email protected]>
| Newsgroups | gmane.comp.gcc.cvs |
|---|---|
| Message-ID | <[email protected]> |
https://gcc.gnu.org/g:55afdf36b7a2acd881c93485ce2af6d1cf006e18 commit r17-3247-g55afdf36b7a2acd881c93485ce2af6d1cf006e18 Author: Pan Li <[email protected]> Date: Wed Aug 12 10:42:17 2026 +0800 RISC-V: Add test cases for vwmacc.vv reg overlap Add test cases for register group overlap, please note it is not overlap as much as possible. gcc/testsuite/ChangeLog: * gcc.target/riscv/rvv/autovec/group_overlap/group_overlap.h: Add test helper macros. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m1.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m2.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m4.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-mf2.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-mf4.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m1.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m2.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m4.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-mf2.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m1.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m2.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m4.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf2.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf4.c: New test. * gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf8.c: New test. Signed-off-by: Pan Li <[email protected]> Diff: --- .../rvv/autovec/group_overlap/group_overlap.h | 230 +++++++++++++++++++++ .../rvv/autovec/group_overlap/vwmacc_vv-i16-m1.c | 57 +++++ .../rvv/autovec/group_overlap/vwmacc_vv-i16-m2.c | 49 +++++ .../rvv/autovec/group_overlap/vwmacc_vv-i16-m4.c | 45 ++++ .../rvv/autovec/group_overlap/vwmacc_vv-i16-mf2.c | 21 ++ .../rvv/autovec/group_overlap/vwmacc_vv-i16-mf4.c | 21 ++ .../rvv/autovec/group_overlap/vwmacc_vv-i32-m1.c | 57 +++++ .../rvv/autovec/group_overlap/vwmacc_vv-i32-m2.c | 49 +++++ .../rvv/autovec/group_overlap/vwmacc_vv-i32-m4.c | 45 ++++ .../rvv/autovec/group_overlap/vwmacc_vv-i32-mf2.c | 21 ++ .../rvv/autovec/group_overlap/vwmacc_vv-i8-m1.c | 57 +++++ .../rvv/autovec/group_overlap/vwmacc_vv-i8-m2.c | 49 +++++ .../rvv/autovec/group_overlap/vwmacc_vv-i8-m4.c | 45 ++++ .../rvv/autovec/group_overlap/vwmacc_vv-i8-mf2.c | 21 ++ .../rvv/autovec/group_overlap/vwmacc_vv-i8-mf4.c | 21 ++ .../rvv/autovec/group_overlap/vwmacc_vv-i8-mf8.c | 21 ++ 16 files changed, 809 insertions(+) diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/group_overlap.h b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/group_overlap.h index 9d6e5f2b8049..ac2abed02ba8 100644 --- a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/group_overlap.h +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/group_overlap.h @@ -533,6 +533,205 @@ ST_F ((void *)out, vd14, VL); OUT += VL; \ ST_F ((void *)out, vd15, VL); OUT += VL; \ +#define LOOP_DUAL_WIDEN_TERNARY_BODY_X4(NT, WT, LD_NF, LD_WF, OUT_F, ST_F, \ + OUT, START, VL) \ + NT vs0 = LD_NF ((void *)START, VL); START += VL; \ + NT vs1 = LD_NF ((void *)START, VL); START += VL; \ + NT vs2 = LD_NF ((void *)START, VL); START += VL; \ + NT vs3 = LD_NF ((void *)START, VL); START += VL; \ + NT vt0 = LD_NF ((void *)START, VL); START += VL; \ + NT vt1 = LD_NF ((void *)START, VL); START += VL; \ + NT vt2 = LD_NF ((void *)START, VL); START += VL; \ + NT vt3 = LD_NF ((void *)START, VL); START += VL; \ + WT vw0 = LD_WF ((void *)START, VL); START += VL; \ + WT vw1 = LD_WF ((void *)START, VL); START += VL; \ + WT vw2 = LD_WF ((void *)START, VL); START += VL; \ + WT vw3 = LD_WF ((void *)START, VL); START += VL; \ + \ + asm volatile("nop" ::: "memory"); \ + \ + WT vd0 = OUT_F (vw0, vs0, vt0, VL); \ + WT vd1 = OUT_F (vw1, vs1, vt1, VL); \ + WT vd2 = OUT_F (vw2, vs2, vt2, VL); \ + WT vd3 = OUT_F (vw3, vs3, vt3, VL); \ + \ + asm volatile("nop" ::: "memory"); \ + \ + ST_F ((void *)out, vd0, VL); OUT += VL; \ + ST_F ((void *)out, vd1, VL); OUT += VL; \ + ST_F ((void *)out, vd2, VL); OUT += VL; \ + ST_F ((void *)out, vd3, VL); OUT += VL; \ + +#define LOOP_DUAL_WIDEN_TERNARY_BODY_X8(NT, WT, LD_NF, LD_WF, OUT_F, ST_F, \ + OUT, START, VL) \ + NT vs0 = LD_NF ((void *)START, VL); START += VL; \ + NT vs1 = LD_NF ((void *)START, VL); START += VL; \ + NT vs2 = LD_NF ((void *)START, VL); START += VL; \ + NT vs3 = LD_NF ((void *)START, VL); START += VL; \ + NT vs4 = LD_NF ((void *)START, VL); START += VL; \ + NT vs5 = LD_NF ((void *)START, VL); START += VL; \ + NT vs6 = LD_NF ((void *)START, VL); START += VL; \ + NT vs7 = LD_NF ((void *)START, VL); START += VL; \ + NT vt0 = LD_NF ((void *)START, VL); START += VL; \ + NT vt1 = LD_NF ((void *)START, VL); START += VL; \ + NT vt2 = LD_NF ((void *)START, VL); START += VL; \ + NT vt3 = LD_NF ((void *)START, VL); START += VL; \ + NT vt4 = LD_NF ((void *)START, VL); START += VL; \ + NT vt5 = LD_NF ((void *)START, VL); START += VL; \ + NT vt6 = LD_NF ((void *)START, VL); START += VL; \ + NT vt7 = LD_NF ((void *)START, VL); START += VL; \ + WT vw0 = LD_WF ((void *)START, VL); START += VL; \ + WT vw1 = LD_WF ((void *)START, VL); START += VL; \ + WT vw2 = LD_WF ((void *)START, VL); START += VL; \ + WT vw3 = LD_WF ((void *)START, VL); START += VL; \ + WT vw4 = LD_WF ((void *)START, VL); START += VL; \ + WT vw5 = LD_WF ((void *)START, VL); START += VL; \ + WT vw6 = LD_WF ((void *)START, VL); START += VL; \ + WT vw7 = LD_WF ((void *)START, VL); START += VL; \ + \ + asm volatile("nop" ::: "memory"); \ + \ + WT vd0 = OUT_F (vw0, vs0, vt0, VL); \ + WT vd1 = OUT_F (vw1, vs1, vt1, VL); \ + WT vd2 = OUT_F (vw2, vs2, vt2, VL); \ + WT vd3 = OUT_F (vw3, vs3, vt3, VL); \ + WT vd4 = OUT_F (vw4, vs4, vt4, VL); \ + WT vd5 = OUT_F (vw5, vs5, vt5, VL); \ + WT vd6 = OUT_F (vw6, vs6, vt6, VL); \ + WT vd7 = OUT_F (vw7, vs7, vt7, VL); \ + \ + asm volatile("nop" ::: "memory"); \ + \ + ST_F ((void *)out, vd0, VL); OUT += VL; \ + ST_F ((void *)out, vd1, VL); OUT += VL; \ + ST_F ((void *)out, vd2, VL); OUT += VL; \ + ST_F ((void *)out, vd3, VL); OUT += VL; \ + ST_F ((void *)out, vd4, VL); OUT += VL; \ + ST_F ((void *)out, vd5, VL); OUT += VL; \ + ST_F ((void *)out, vd6, VL); OUT += VL; \ + ST_F ((void *)out, vd7, VL); OUT += VL; \ + +#define LOOP_DUAL_WIDEN_TERNARY_BODY_X16(NT, WT, LD_NF, LD_WF, OUT_F, ST_F, \ + OUT, START, VL) \ + NT vs0 = LD_NF ((void *)START, VL); START += VL; \ + NT vs1 = LD_NF ((void *)START, VL); START += VL; \ + NT vs2 = LD_NF ((void *)START, VL); START += VL; \ + NT vs3 = LD_NF ((void *)START, VL); START += VL; \ + NT vs4 = LD_NF ((void *)START, VL); START += VL; \ + NT vs5 = LD_NF ((void *)START, VL); START += VL; \ + NT vs6 = LD_NF ((void *)START, VL); START += VL; \ + NT vs7 = LD_NF ((void *)START, VL); START += VL; \ + NT vs8 = LD_NF ((void *)START, VL); START += VL; \ + NT vs9 = LD_NF ((void *)START, VL); START += VL; \ + NT vs10 = LD_NF ((void *)START, VL); START += VL; \ + NT vs11 = LD_NF ((void *)START, VL); START += VL; \ + NT vs12 = LD_NF ((void *)START, VL); START += VL; \ + NT vs13 = LD_NF ((void *)START, VL); START += VL; \ + NT vs14 = LD_NF ((void *)START, VL); START += VL; \ + NT vs15 = LD_NF ((void *)START, VL); START += VL; \ + NT vt0 = LD_NF ((void *)START, VL); START += VL; \ + NT vt1 = LD_NF ((void *)START, VL); START += VL; \ + NT vt2 = LD_NF ((void *)START, VL); START += VL; \ + NT vt3 = LD_NF ((void *)START, VL); START += VL; \ + NT vt4 = LD_NF ((void *)START, VL); START += VL; \ + NT vt5 = LD_NF ((void *)START, VL); START += VL; \ + NT vt6 = LD_NF ((void *)START, VL); START += VL; \ + NT vt7 = LD_NF ((void *)START, VL); START += VL; \ + NT vt8 = LD_NF ((void *)START, VL); START += VL; \ + NT vt9 = LD_NF ((void *)START, VL); START += VL; \ + NT vt10 = LD_NF ((void *)START, VL); START += VL; \ + NT vt11 = LD_NF ((void *)START, VL); START += VL; \ + NT vt12 = LD_NF ((void *)START, VL); START += VL; \ + NT vt13 = LD_NF ((void *)START, VL); START += VL; \ + NT vt14 = LD_NF ((void *)START, VL); START += VL; \ + NT vt15 = LD_NF ((void *)START, VL); START += VL; \ + WT vw0 = LD_WF ((void *)START, VL); START += VL; \ + WT vw1 = LD_WF ((void *)START, VL); START += VL; \ + WT vw2 = LD_WF ((void *)START, VL); START += VL; \ + WT vw3 = LD_WF ((void *)START, VL); START += VL; \ + WT vw4 = LD_WF ((void *)START, VL); START += VL; \ + WT vw5 = LD_WF ((void *)START, VL); START += VL; \ + WT vw6 = LD_WF ((void *)START, VL); START += VL; \ + WT vw7 = LD_WF ((void *)START, VL); START += VL; \ + WT vw8 = LD_WF ((void *)START, VL); START += VL; \ + WT vw9 = LD_WF ((void *)START, VL); START += VL; \ + WT vw10 = LD_WF ((void *)START, VL); START += VL; \ + WT vw11 = LD_WF ((void *)START, VL); START += VL; \ + WT vw12 = LD_WF ((void *)START, VL); START += VL; \ + WT vw13 = LD_WF ((void *)START, VL); START += VL; \ + WT vw14 = LD_WF ((void *)START, VL); START += VL; \ + WT vw15 = LD_WF ((void *)START, VL); START += VL; \ + \ + asm volatile("nop" ::: "memory"); \ + \ + WT vd0 = OUT_F (vw0, vs0, vt0, VL); \ + WT vd1 = OUT_F (vw1, vs1, vt1, VL); \ + WT vd2 = OUT_F (vw2, vs2, vt2, VL); \ + WT vd3 = OUT_F (vw3, vs3, vt3, VL); \ + WT vd4 = OUT_F (vw4, vs4, vt4, VL); \ + WT vd5 = OUT_F (vw5, vs5, vt5, VL); \ + WT vd6 = OUT_F (vw6, vs6, vt6, VL); \ + WT vd7 = OUT_F (vw7, vs7, vt7, VL); \ + WT vd8 = OUT_F (vw8, vs8, vt8, VL); \ + WT vd9 = OUT_F (vw9, vs9, vt9, VL); \ + WT vd10 = OUT_F (vw10, vs10, vt10, VL); \ + WT vd11 = OUT_F (vw11, vs11, vt11, VL); \ + WT vd12 = OUT_F (vw12, vs12, vt12, VL); \ + WT vd13 = OUT_F (vw13, vs13, vt13, VL); \ + WT vd14 = OUT_F (vw14, vs14, vt14, VL); \ + WT vd15 = OUT_F (vw15, vs15, vt15, VL); \ + \ + asm volatile("nop" ::: "memory"); \ + \ + ST_F ((void *)out, vd0, VL); OUT += VL; \ + ST_F ((void *)out, vd1, VL); OUT += VL; \ + ST_F ((void *)out, vd2, VL); OUT += VL; \ + ST_F ((void *)out, vd3, VL); OUT += VL; \ + ST_F ((void *)out, vd4, VL); OUT += VL; \ + ST_F ((void *)out, vd5, VL); OUT += VL; \ + ST_F ((void *)out, vd6, VL); OUT += VL; \ + ST_F ((void *)out, vd7, VL); OUT += VL; \ + ST_F ((void *)out, vd8, VL); OUT += VL; \ + ST_F ((void *)out, vd9, VL); OUT += VL; \ + ST_F ((void *)out, vd10, VL); OUT += VL; \ + ST_F ((void *)out, vd11, VL); OUT += VL; \ + ST_F ((void *)out, vd12, VL); OUT += VL; \ + ST_F ((void *)out, vd13, VL); OUT += VL; \ + ST_F ((void *)out, vd14, VL); OUT += VL; \ + ST_F ((void *)out, vd15, VL); OUT += VL; \ + +/* The widened destination register group of a dual widen ternary insn is tied + to the accumulator, which is still live when the narrowed sources are read. + Feeding one narrowed source from the highest-numbered half of the + accumulator's own register group is thus the only way the sources can + legally overlap the destination register group: the shared registers hold + one single value, seen as the accumulator's high part and as the narrowed + source at the same time. RI_F reinterprets the widened type as the + narrowed element type, GET_F extracts the highest-numbered half of it. */ +#define LOOP_DUAL_WIDEN_TERNARY_BODY_OVERLAP_X2(NT, WT, WNT, LD_NF, LD_WF, \ + RI_F, GET_F, OUT_F, ST_F, \ + OUT, START, VL) \ + NT vt0 = LD_NF ((void *)START, VL); START += VL; \ + NT vt1 = LD_NF ((void *)START, VL); START += VL; \ + WT vw0 = LD_WF ((void *)START, VL); START += VL; \ + WT vw1 = LD_WF ((void *)START, VL); START += VL; \ + \ + asm volatile("nop" ::: "memory"); \ + \ + WNT vr0 = RI_F (vw0); \ + WNT vr1 = RI_F (vw1); \ + \ + NT vs0 = GET_F (vr0, 1); \ + NT vs1 = GET_F (vr1, 1); \ + \ + WT vd0 = OUT_F (vw0, vs0, vt0, VL); \ + WT vd1 = OUT_F (vw1, vs1, vt1, VL); \ + \ + asm volatile("nop" ::: "memory"); \ + \ + ST_F ((void *)out, vd0, VL); OUT += VL; \ + ST_F ((void *)out, vd1, VL); OUT += VL; \ + #define DEF_GROUP_OVERLAP_UNARY_0(VL_F, NT, WT, LD_F, OUT_F, ST_F, NAME, \ LOOP_BODY) \ void test_group_overlap_##NAME##_##NT##_unary_0(uint8_t *data, \ @@ -594,4 +793,35 @@ } \ } +#define DEF_GROUP_OVERLAP_TERNARY_0(VL_F, NT, WT, LD_NF, LD_WF, OUT_F, \ + ST_F, NAME, LOOP_BODY) \ + void test_group_overlap_##NAME##_##NT##_ternary_0(uint8_t *data, \ + uint8_t *out, \ + size_t limit) \ + { \ + uint8_t *start = data; \ + uint8_t *end = data + limit; \ + size_t vl = VL_F (); \ + \ + while (start < end) { \ + LOOP_BODY (NT, WT, LD_NF, LD_WF, OUT_F, ST_F, out, start, vl); \ + } \ + } + +#define DEF_GROUP_OVERLAP_TERNARY_1(VL_F, NT, WT, WNT, LD_NF, LD_WF, RI_F, \ + GET_F, OUT_F, ST_F, NAME, LOOP_BODY) \ + void test_group_overlap_##NAME##_##NT##_ternary_1(uint8_t *data, \ + uint8_t *out, \ + size_t limit) \ + { \ + uint8_t *start = data; \ + uint8_t *end = data + limit; \ + size_t vl = VL_F (); \ + \ + while (start < end) { \ + LOOP_BODY (NT, WT, WNT, LD_NF, LD_WF, RI_F, GET_F, OUT_F, ST_F, out, \ + start, vl); \ + } \ + } + #endif diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m1.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m1.c new file mode 100644 index 000000000000..79f34b2aa0bc --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m1.c @@ -0,0 +1,57 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e16m1, + vint16m1_t, + vint32m2_t, + __riscv_vle16_v_i16m1, + __riscv_vle32_v_i32m2, + __riscv_vwmacc_vv_i32m2, + __riscv_vse32_v_i32m2, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X16) + +DEF_GROUP_OVERLAP_TERNARY_1( + __riscv_vsetvlmax_e16m1, + vint16m1_t, + vint32m2_t, + vint16m2_t, + __riscv_vle16_v_i16m1, + __riscv_vle32_v_i32m2, + __riscv_vreinterpret_v_i32m2_i16m2, + __riscv_vget_v_i16m2_i16m1, + __riscv_vwmacc_vv_i32m2, + __riscv_vse32_v_i32m2, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_OVERLAP_X2) + +/* ternary_0: the accumulator occupies the whole destination register group and + is live when the narrowed sources are read, so no source can be allocated + inside the destination register group. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v12,v30,v29([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v4,v16,v15([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v2,v0,v31([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v14,v1([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v30,v28,v27([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v28,v26,v25([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v6,v20,v19([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v26,v24,v23([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v2,v18,v17([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v10,v1,v15([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v24,v22,v21([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v14,v19,v17([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v21,v23([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v18,v20,v22([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v20,v0,v1([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v22,v0,v1([^0-9]|$)} 1 } } */ + +/* ternary_1: each narrowed source is the highest-numbered half of its own + accumulator, thus it overlaps the highest-numbered part of the destination + register group. Without the group overlap the sources would have to be + copied out to a disjoint register group first. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v4,v5,v8([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v2,v3,v1([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-not {vmv[0-9]+r\.v} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m2.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m2.c new file mode 100644 index 000000000000..0d6660b5cb6e --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m2.c @@ -0,0 +1,49 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e16m2, + vint16m2_t, + vint32m4_t, + __riscv_vle16_v_i16m2, + __riscv_vle32_v_i32m4, + __riscv_vwmacc_vv_i32m4, + __riscv_vse32_v_i32m4, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X8) + +DEF_GROUP_OVERLAP_TERNARY_1( + __riscv_vsetvlmax_e16m2, + vint16m2_t, + vint32m4_t, + vint16m4_t, + __riscv_vle16_v_i16m2, + __riscv_vle32_v_i32m4, + __riscv_vreinterpret_v_i32m4_i16m4, + __riscv_vget_v_i16m4_i16m2, + __riscv_vwmacc_vv_i32m4, + __riscv_vse32_v_i32m4, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_OVERLAP_X2) + +/* ternary_0: the accumulator occupies the whole destination register group and + is live when the narrowed sources are read, so no source can be allocated + inside the destination register group. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v0,v30([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v0,v28,v26([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v28,v24,v22([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v24,v20,v18([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v20,v16,v14([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v12,v10([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v12,v8,v6([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v4,v2([^0-9]|$)} 1 } } */ + +/* ternary_1: each narrowed source is the highest-numbered half of its own + accumulator, thus it overlaps the highest-numbered part of the destination + register group. Without the group overlap the sources would have to be + copied out to a disjoint register group first. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v10,v16([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v4,v6,v2([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-not {vmv[0-9]+r\.v} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m4.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m4.c new file mode 100644 index 000000000000..7568c6a871ab --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-m4.c @@ -0,0 +1,45 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e16m4, + vint16m4_t, + vint32m8_t, + __riscv_vle16_v_i16m4, + __riscv_vle32_v_i32m8, + __riscv_vwmacc_vv_i32m8, + __riscv_vse32_v_i32m8, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X4) + +DEF_GROUP_OVERLAP_TERNARY_1( + __riscv_vsetvlmax_e16m4, + vint16m4_t, + vint32m8_t, + vint16m8_t, + __riscv_vle16_v_i16m4, + __riscv_vle32_v_i32m8, + __riscv_vreinterpret_v_i32m8_i16m8, + __riscv_vget_v_i16m8_i16m4, + __riscv_vwmacc_vv_i32m8, + __riscv_vse32_v_i32m8, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_OVERLAP_X2) + +/* ternary_0: the accumulator occupies the whole destination register group and + is live when the narrowed sources are read, so no source can be allocated + inside the destination register group. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v0,v4([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v0,v28,v20([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v24,v16,v12([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v8,v12([^0-9]|$)} 1 } } */ + +/* ternary_1: each narrowed source is the highest-numbered half of its own + accumulator, thus it overlaps the highest-numbered part of the destination + register group. Without the group overlap the sources would have to be + copied out to a disjoint register group first. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v20,v0([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v12,v4([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-not {vmv[0-9]+r\.v} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-mf2.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-mf2.c new file mode 100644 index 000000000000..f90ed36ffa6d --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-mf2.c @@ -0,0 +1,21 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e16m1, + vint16mf2_t, + vint32m1_t, + __riscv_vle16_v_i16mf2, + __riscv_vle32_v_i32m1, + __riscv_vwmacc_vv_i32m1, + __riscv_vse32_v_i32m1, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X16) + +/* The fractional LMUL source has EMUL < 1, thus the widened destination + register group must not overlap either source at all. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv} 16 } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),\1,} } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),v[0-9]+,\1([^0-9]|$)} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-mf4.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-mf4.c new file mode 100644 index 000000000000..6f7544fd19d9 --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i16-mf4.c @@ -0,0 +1,21 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e16m1, + vint16mf4_t, + vint32mf2_t, + __riscv_vle16_v_i16mf4, + __riscv_vle32_v_i32mf2, + __riscv_vwmacc_vv_i32mf2, + __riscv_vse32_v_i32mf2, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X16) + +/* The fractional LMUL source has EMUL < 1, thus the widened destination + register group must not overlap either source at all. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv} 16 } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),\1,} } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),v[0-9]+,\1([^0-9]|$)} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m1.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m1.c new file mode 100644 index 000000000000..bfd25e177d8a --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m1.c @@ -0,0 +1,57 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e32m1, + vint32m1_t, + vint64m2_t, + __riscv_vle32_v_i32m1, + __riscv_vle64_v_i64m2, + __riscv_vwmacc_vv_i64m2, + __riscv_vse64_v_i64m2, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X16) + +DEF_GROUP_OVERLAP_TERNARY_1( + __riscv_vsetvlmax_e32m1, + vint32m1_t, + vint64m2_t, + vint32m2_t, + __riscv_vle32_v_i32m1, + __riscv_vle64_v_i64m2, + __riscv_vreinterpret_v_i64m2_i32m2, + __riscv_vget_v_i32m2_i32m1, + __riscv_vwmacc_vv_i64m2, + __riscv_vse64_v_i64m2, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_OVERLAP_X2) + +/* ternary_0: the accumulator occupies the whole destination register group and + is live when the narrowed sources are read, so no source can be allocated + inside the destination register group. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v12,v30,v29([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v4,v16,v15([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v2,v0,v31([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v14,v1([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v30,v28,v27([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v28,v26,v25([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v6,v20,v19([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v26,v24,v23([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v2,v18,v17([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v10,v1,v15([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v24,v22,v21([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v14,v19,v17([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v21,v23([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v18,v20,v22([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v20,v0,v1([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v22,v0,v1([^0-9]|$)} 1 } } */ + +/* ternary_1: each narrowed source is the highest-numbered half of its own + accumulator, thus it overlaps the highest-numbered part of the destination + register group. Without the group overlap the sources would have to be + copied out to a disjoint register group first. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v4,v5,v8([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v2,v3,v1([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-not {vmv[0-9]+r\.v} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m2.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m2.c new file mode 100644 index 000000000000..a24c7410bc02 --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m2.c @@ -0,0 +1,49 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e32m2, + vint32m2_t, + vint64m4_t, + __riscv_vle32_v_i32m2, + __riscv_vle64_v_i64m4, + __riscv_vwmacc_vv_i64m4, + __riscv_vse64_v_i64m4, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X8) + +DEF_GROUP_OVERLAP_TERNARY_1( + __riscv_vsetvlmax_e32m2, + vint32m2_t, + vint64m4_t, + vint32m4_t, + __riscv_vle32_v_i32m2, + __riscv_vle64_v_i64m4, + __riscv_vreinterpret_v_i64m4_i32m4, + __riscv_vget_v_i32m4_i32m2, + __riscv_vwmacc_vv_i64m4, + __riscv_vse64_v_i64m4, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_OVERLAP_X2) + +/* ternary_0: the accumulator occupies the whole destination register group and + is live when the narrowed sources are read, so no source can be allocated + inside the destination register group. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v0,v30([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v0,v28,v26([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v28,v24,v22([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v24,v20,v18([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v20,v16,v14([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v12,v10([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v12,v8,v6([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v4,v2([^0-9]|$)} 1 } } */ + +/* ternary_1: each narrowed source is the highest-numbered half of its own + accumulator, thus it overlaps the highest-numbered part of the destination + register group. Without the group overlap the sources would have to be + copied out to a disjoint register group first. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v10,v16([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v4,v6,v2([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-not {vmv[0-9]+r\.v} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m4.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m4.c new file mode 100644 index 000000000000..17490d331e0c --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-m4.c @@ -0,0 +1,45 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e32m4, + vint32m4_t, + vint64m8_t, + __riscv_vle32_v_i32m4, + __riscv_vle64_v_i64m8, + __riscv_vwmacc_vv_i64m8, + __riscv_vse64_v_i64m8, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X4) + +DEF_GROUP_OVERLAP_TERNARY_1( + __riscv_vsetvlmax_e32m4, + vint32m4_t, + vint64m8_t, + vint32m8_t, + __riscv_vle32_v_i32m4, + __riscv_vle64_v_i64m8, + __riscv_vreinterpret_v_i64m8_i32m8, + __riscv_vget_v_i32m8_i32m4, + __riscv_vwmacc_vv_i64m8, + __riscv_vse64_v_i64m8, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_OVERLAP_X2) + +/* ternary_0: the accumulator occupies the whole destination register group and + is live when the narrowed sources are read, so no source can be allocated + inside the destination register group. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v0,v4([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v0,v28,v20([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v24,v16,v12([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v8,v12([^0-9]|$)} 1 } } */ + +/* ternary_1: each narrowed source is the highest-numbered half of its own + accumulator, thus it overlaps the highest-numbered part of the destination + register group. Without the group overlap the sources would have to be + copied out to a disjoint register group first. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v20,v0([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v12,v4([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-not {vmv[0-9]+r\.v} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-mf2.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-mf2.c new file mode 100644 index 000000000000..2b7c19cd5372 --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i32-mf2.c @@ -0,0 +1,21 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e32m1, + vint32mf2_t, + vint64m1_t, + __riscv_vle32_v_i32mf2, + __riscv_vle64_v_i64m1, + __riscv_vwmacc_vv_i64m1, + __riscv_vse64_v_i64m1, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X16) + +/* The fractional LMUL source has EMUL < 1, thus the widened destination + register group must not overlap either source at all. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv} 16 } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),\1,} } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),v[0-9]+,\1([^0-9]|$)} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m1.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m1.c new file mode 100644 index 000000000000..a6399818a70d --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m1.c @@ -0,0 +1,57 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e8m1, + vint8m1_t, + vint16m2_t, + __riscv_vle8_v_i8m1, + __riscv_vle16_v_i16m2, + __riscv_vwmacc_vv_i16m2, + __riscv_vse16_v_i16m2, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X16) + +DEF_GROUP_OVERLAP_TERNARY_1( + __riscv_vsetvlmax_e8m1, + vint8m1_t, + vint16m2_t, + vint8m2_t, + __riscv_vle8_v_i8m1, + __riscv_vle16_v_i16m2, + __riscv_vreinterpret_v_i16m2_i8m2, + __riscv_vget_v_i8m2_i8m1, + __riscv_vwmacc_vv_i16m2, + __riscv_vse16_v_i16m2, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_OVERLAP_X2) + +/* ternary_0: the accumulator occupies the whole destination register group and + is live when the narrowed sources are read, so no source can be allocated + inside the destination register group. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v12,v30,v29([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v4,v16,v15([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v2,v0,v31([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v14,v1([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v30,v28,v27([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v28,v26,v25([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v6,v20,v19([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v26,v24,v23([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v2,v18,v17([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v10,v1,v15([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v24,v22,v21([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v14,v19,v17([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v21,v23([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v18,v20,v22([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v20,v0,v1([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v22,v0,v1([^0-9]|$)} 1 } } */ + +/* ternary_1: each narrowed source is the highest-numbered half of its own + accumulator, thus it overlaps the highest-numbered part of the destination + register group. Without the group overlap the sources would have to be + copied out to a disjoint register group first. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v4,v5,v8([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v2,v3,v1([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-not {vmv[0-9]+r\.v} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m2.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m2.c new file mode 100644 index 000000000000..a8e4628a8c5e --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m2.c @@ -0,0 +1,49 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e8m2, + vint8m2_t, + vint16m4_t, + __riscv_vle8_v_i8m2, + __riscv_vle16_v_i16m4, + __riscv_vwmacc_vv_i16m4, + __riscv_vse16_v_i16m4, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X8) + +DEF_GROUP_OVERLAP_TERNARY_1( + __riscv_vsetvlmax_e8m2, + vint8m2_t, + vint16m4_t, + vint8m4_t, + __riscv_vle8_v_i8m2, + __riscv_vle16_v_i16m4, + __riscv_vreinterpret_v_i16m4_i8m4, + __riscv_vget_v_i8m4_i8m2, + __riscv_vwmacc_vv_i16m4, + __riscv_vse16_v_i16m4, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_OVERLAP_X2) + +/* ternary_0: the accumulator occupies the whole destination register group and + is live when the narrowed sources are read, so no source can be allocated + inside the destination register group. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v0,v30([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v0,v28,v26([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v28,v24,v22([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v24,v20,v18([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v20,v16,v14([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v12,v10([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v12,v8,v6([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v4,v2([^0-9]|$)} 1 } } */ + +/* ternary_1: each narrowed source is the highest-numbered half of its own + accumulator, thus it overlaps the highest-numbered part of the destination + register group. Without the group overlap the sources would have to be + copied out to a disjoint register group first. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v10,v16([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v4,v6,v2([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-not {vmv[0-9]+r\.v} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m4.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m4.c new file mode 100644 index 000000000000..67c647f4ca80 --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-m4.c @@ -0,0 +1,45 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e8m4, + vint8m4_t, + vint16m8_t, + __riscv_vle8_v_i8m4, + __riscv_vle16_v_i16m8, + __riscv_vwmacc_vv_i16m8, + __riscv_vse16_v_i16m8, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X4) + +DEF_GROUP_OVERLAP_TERNARY_1( + __riscv_vsetvlmax_e8m4, + vint8m4_t, + vint16m8_t, + vint8m8_t, + __riscv_vle8_v_i8m4, + __riscv_vle16_v_i16m8, + __riscv_vreinterpret_v_i16m8_i8m8, + __riscv_vget_v_i8m8_i8m4, + __riscv_vwmacc_vv_i16m8, + __riscv_vse16_v_i16m8, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_OVERLAP_X2) + +/* ternary_0: the accumulator occupies the whole destination register group and + is live when the narrowed sources are read, so no source can be allocated + inside the destination register group. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v0,v4([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v0,v28,v20([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v24,v16,v12([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v8,v12([^0-9]|$)} 1 } } */ + +/* ternary_1: each narrowed source is the highest-numbered half of its own + accumulator, thus it overlaps the highest-numbered part of the destination + register group. Without the group overlap the sources would have to be + copied out to a disjoint register group first. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v16,v20,v0([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-times {vwmacc\.vv\s+v8,v12,v4([^0-9]|$)} 1 } } */ +/* { dg-final { scan-assembler-not {vmv[0-9]+r\.v} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf2.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf2.c new file mode 100644 index 000000000000..c0404d74b4aa --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf2.c @@ -0,0 +1,21 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e8m1, + vint8mf2_t, + vint16m1_t, + __riscv_vle8_v_i8mf2, + __riscv_vle16_v_i16m1, + __riscv_vwmacc_vv_i16m1, + __riscv_vse16_v_i16m1, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X16) + +/* The fractional LMUL source has EMUL < 1, thus the widened destination + register group must not overlap either source at all. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv} 16 } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),\1,} } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),v[0-9]+,\1([^0-9]|$)} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf4.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf4.c new file mode 100644 index 000000000000..11bce6b79a96 --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf4.c @@ -0,0 +1,21 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e8m1, + vint8mf4_t, + vint16mf2_t, + __riscv_vle8_v_i8mf4, + __riscv_vle16_v_i16mf2, + __riscv_vwmacc_vv_i16mf2, + __riscv_vse16_v_i16mf2, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X16) + +/* The fractional LMUL source has EMUL < 1, thus the widened destination + register group must not overlap either source at all. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv} 16 } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),\1,} } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),v[0-9]+,\1([^0-9]|$)} } } */ diff --git a/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf8.c b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf8.c new file mode 100644 index 000000000000..f0704ada8050 --- /dev/null +++ b/gcc/testsuite/gcc.target/riscv/rvv/autovec/group_overlap/vwmacc_vv-i8-mf8.c @@ -0,0 +1,21 @@ +/* { dg-do compile } */ +/* { dg-options "-march=rv64gcv -mabi=lp64d" } */ + +#include "group_overlap.h" + +DEF_GROUP_OVERLAP_TERNARY_0( + __riscv_vsetvlmax_e8m1, + vint8mf8_t, + vint16mf4_t, + __riscv_vle8_v_i8mf8, + __riscv_vle16_v_i16mf4, + __riscv_vwmacc_vv_i16mf4, + __riscv_vse16_v_i16mf4, + vwmacc_vv, + LOOP_DUAL_WIDEN_TERNARY_BODY_X16) + +/* The fractional LMUL source has EMUL < 1, thus the widened destination + register group must not overlap either source at all. */ +/* { dg-final { scan-assembler-times {vwmacc\.vv} 16 } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),\1,} } } */ +/* { dg-final { scan-assembler-not {vwmacc\.vv\s+(v[0-9]+),v[0-9]+,\1([^0-9]|$)} } } */