Hi Segher, Following your review comments in [1], this patch is separated from Xionghu's patch v4 [2] and mainly targetted for 32-bit element size, it changes with the generic call altivec_vmrg*w in vec_widen_[su]mult_{hi,lo}* expanders as well. If this patch looks good to you, I'll post the others for 16 and 8 bits element size.
btw, there are still something that can be improved like the loose predicates pointed out by Peter, vsx_xxmrg[hl]w_* can be merged with altivec_vmrg*w etc., but I think we want this one to focus on fixing regression and those enhancements can be supported by some separated follow-up patches. Bootstrapped and regtested on powerpc64-linux-gnu P8/P9 and powerpc64le-linux-gnu P9 and P10. [1] https://gcc.gnu.org/pipermail/gcc-patches/2024-June/655019.html [2] https://gcc.gnu.org/pipermail/gcc-patches/2023-February/611683.html Is it ok for trunk? BR, Kewen ----- Commit r12-4496 changes some define_expands and define_insns for vector merge high/low word, which are altivec_vmrg[hl]w, vsx_xxmrg[hl]w_<VSX_W:mode>. These defines are mainly for built-in function vec_merge{h,l}, __builtin_vsx_xxmrghw, __builtin_vsx_xxmrghw_4si and some internal gen function needs. These functions should consider endianness, taking vec_mergeh as example, as PVIPR defines, vec_mergeh "Merges the first halves (in element order) of two vectors", it does note it's in element order. So it's mapped into vmrghw on BE while vmrglw on LE respectively. Although the mapped insns are different, as the discussion in PR106069, the RTL pattern should be still the same, it is conformed before commit r12-4496, define_expand altivec_vmrghw got expanded into: (vec_select:VSX_W (vec_concat:<VS_double> (match_operand:VSX_W 1 "register_operand" "wa,v") (match_operand:VSX_W 2 "register_operand" "wa,v")) (parallel [(const_int 0) (const_int 4) (const_int 1) (const_int 5)])))] on both BE and LE then. But commit r12-4496 changed it to expand into: (vec_select:VSX_W (vec_concat:<VS_double> (match_operand:VSX_W 1 "register_operand" "wa,v") (match_operand:VSX_W 2 "register_operand" "wa,v")) (parallel [(const_int 0) (const_int 4) (const_int 1) (const_int 5)])))] on BE, and (vec_select:VSX_W (vec_concat:<VS_double> (match_operand:VSX_W 1 "register_operand" "wa,v") (match_operand:VSX_W 2 "register_operand" "wa,v")) (parallel [(const_int 2) (const_int 6) (const_int 3) (const_int 7)])))] on LE, although the mapped insn are still vmrghw on BE and vmrglw on LE, the associated RTL pattern is completely wrong and inconsistent with the mapped insn. If optimization passes leave this pattern alone, even if its pattern doesn't represent its mapped insn, it's still fine, that's why simple testing on bif doesn't expose this issue. But once some optimization pass such as combine does some changes basing on this wrong pattern, because the pattern doesn't match the semantics that the expanded insn is intended to represent, it would cause the unexpected result. So this patch is to fix the wrong RTL pattern, ensure the associated RTL patterns become the same as before which can have the same semantic as their mapped insns. With the proposed patch, the expanders like altivec_vmrghw expands into altivec_vmrghb_direct_be or altivec_vmrglb_direct_le depending on endianness, "direct" can easily show which insn would be generated, _be and _le are mainly for the different RTL patterns as endianness. Co-authored-by: Xionghu Luo <xionghu...@tencent.com> PR target/106069 PR target/115355 gcc/ChangeLog: * config/rs6000/altivec.md (altivec_vmrghw_direct_<VSX_W:mode>): Rename to ... (altivec_vmrghw_direct_<VSX_W:mode>_be): ... this. Add the condition BYTES_BIG_ENDIAN. (altivec_vmrghw_direct_<VSX_W:mode>_le): New define_insn. (altivec_vmrglw_direct_<VSX_W:mode>): Rename to ... (altivec_vmrglw_direct_<VSX_W:mode>_be): ... this. Add the condition BYTES_BIG_ENDIAN. (altivec_vmrglw_direct_<VSX_W:mode>_le): New define_insn. (altivec_vmrghw): Adjust by calling gen_altivec_vmrghw_direct_v4si_be for BE and gen_altivec_vmrglw_direct_v4si_le for LE. (altivec_vmrglw): Adjust by calling gen_altivec_vmrglw_direct_v4si_be for BE and gen_altivec_vmrghw_direct_v4si_le for LE. (vec_widen_umult_hi_v8hi): Adjust the call to gen_altivec_vmrghw_direct_v4si by gen_altivec_vmrghw for BE and by gen_altivec_vmrglw for LE. (vec_widen_smult_hi_v8hi): Likewise. (vec_widen_umult_lo_v8hi): Adjust the call to gen_altivec_vmrglw_direct_v4si by gen_altivec_vmrglw for BE and by gen_altivec_vmrghw for LE (vec_widen_smult_lo_v8hi): Likewise. * config/rs6000/rs6000.cc (altivec_expand_vec_perm_const): Replace CODE_FOR_altivec_vmrghw_direct_v4si by CODE_FOR_altivec_vmrghw_direct_v4si_be for BE and CODE_FOR_altivec_vmrghw_direct_v4si_le for LE. And replace CODE_FOR_altivec_vmrglw_direct_v4si by CODE_FOR_altivec_vmrglw_direct_v4si_be for BE and CODE_FOR_altivec_vmrglw_direct_v4si_le for LE. * config/rs6000/vsx.md (vsx_xxmrghw_<VSX_W:mode>): Adjust by calling gen_altivec_vmrghw_direct_v4si_be for BE and gen_altivec_vmrglw_direct_v4si_le for LE. (vsx_xxmrglw_<VSX_W:mode>): Adjust by calling gen_altivec_vmrglw_direct_v4si_be for BE and gen_altivec_vmrghw_direct_v4si_le for LE. gcc/testsuite/ChangeLog: * g++.target/powerpc/pr106069.C: New test. * gcc.target/powerpc/pr115355.c: New test. --- gcc/config/rs6000/altivec.md | 80 +++++++++---- gcc/config/rs6000/rs6000.cc | 8 +- gcc/config/rs6000/vsx.md | 28 +++-- gcc/testsuite/g++.target/powerpc/pr106069.C | 119 ++++++++++++++++++++ gcc/testsuite/gcc.target/powerpc/pr115355.c | 37 ++++++ 5 files changed, 232 insertions(+), 40 deletions(-) create mode 100644 gcc/testsuite/g++.target/powerpc/pr106069.C create mode 100644 gcc/testsuite/gcc.target/powerpc/pr115355.c diff --git a/gcc/config/rs6000/altivec.md b/gcc/config/rs6000/altivec.md index bb20441c096..dcc71cc0f52 100644 --- a/gcc/config/rs6000/altivec.md +++ b/gcc/config/rs6000/altivec.md @@ -1212,16 +1212,18 @@ (define_expand "altivec_vmrghw" (use (match_operand:V4SI 2 "register_operand"))] "VECTOR_MEM_ALTIVEC_P (V4SImode)" { - rtx (*fun) (rtx, rtx, rtx); - fun = BYTES_BIG_ENDIAN ? gen_altivec_vmrghw_direct_v4si - : gen_altivec_vmrglw_direct_v4si; - if (!BYTES_BIG_ENDIAN) - std::swap (operands[1], operands[2]); - emit_insn (fun (operands[0], operands[1], operands[2])); + if (BYTES_BIG_ENDIAN) + emit_insn (gen_altivec_vmrghw_direct_v4si_be (operands[0], + operands[1], + operands[2])); + else + emit_insn (gen_altivec_vmrglw_direct_v4si_le (operands[0], + operands[2], + operands[1])); DONE; }) -(define_insn "altivec_vmrghw_direct_<mode>" +(define_insn "altivec_vmrghw_direct_<mode>_be" [(set (match_operand:VSX_W 0 "register_operand" "=wa,v") (vec_select:VSX_W (vec_concat:<VS_double> @@ -1229,7 +1231,21 @@ (define_insn "altivec_vmrghw_direct_<mode>" (match_operand:VSX_W 2 "register_operand" "wa,v")) (parallel [(const_int 0) (const_int 4) (const_int 1) (const_int 5)])))] - "TARGET_ALTIVEC" + "TARGET_ALTIVEC && BYTES_BIG_ENDIAN" + "@ + xxmrghw %x0,%x1,%x2 + vmrghw %0,%1,%2" + [(set_attr "type" "vecperm")]) + +(define_insn "altivec_vmrghw_direct_<mode>_le" + [(set (match_operand:VSX_W 0 "register_operand" "=wa,v") + (vec_select:VSX_W + (vec_concat:<VS_double> + (match_operand:VSX_W 2 "register_operand" "wa,v") + (match_operand:VSX_W 1 "register_operand" "wa,v")) + (parallel [(const_int 2) (const_int 6) + (const_int 3) (const_int 7)])))] + "TARGET_ALTIVEC && !BYTES_BIG_ENDIAN" "@ xxmrghw %x0,%x1,%x2 vmrghw %0,%1,%2" @@ -1318,16 +1334,18 @@ (define_expand "altivec_vmrglw" (use (match_operand:V4SI 2 "register_operand"))] "VECTOR_MEM_ALTIVEC_P (V4SImode)" { - rtx (*fun) (rtx, rtx, rtx); - fun = BYTES_BIG_ENDIAN ? gen_altivec_vmrglw_direct_v4si - : gen_altivec_vmrghw_direct_v4si; - if (!BYTES_BIG_ENDIAN) - std::swap (operands[1], operands[2]); - emit_insn (fun (operands[0], operands[1], operands[2])); + if (BYTES_BIG_ENDIAN) + emit_insn (gen_altivec_vmrglw_direct_v4si_be (operands[0], + operands[1], + operands[2])); + else + emit_insn (gen_altivec_vmrghw_direct_v4si_le (operands[0], + operands[2], + operands[1])); DONE; }) -(define_insn "altivec_vmrglw_direct_<mode>" +(define_insn "altivec_vmrglw_direct_<mode>_be" [(set (match_operand:VSX_W 0 "register_operand" "=wa,v") (vec_select:VSX_W (vec_concat:<VS_double> @@ -1335,7 +1353,21 @@ (define_insn "altivec_vmrglw_direct_<mode>" (match_operand:VSX_W 2 "register_operand" "wa,v")) (parallel [(const_int 2) (const_int 6) (const_int 3) (const_int 7)])))] - "TARGET_ALTIVEC" + "TARGET_ALTIVEC && BYTES_BIG_ENDIAN" + "@ + xxmrglw %x0,%x1,%x2 + vmrglw %0,%1,%2" + [(set_attr "type" "vecperm")]) + +(define_insn "altivec_vmrglw_direct_<mode>_le" + [(set (match_operand:VSX_W 0 "register_operand" "=wa,v") + (vec_select:VSX_W + (vec_concat:<VS_double> + (match_operand:VSX_W 2 "register_operand" "wa,v") + (match_operand:VSX_W 1 "register_operand" "wa,v")) + (parallel [(const_int 0) (const_int 4) + (const_int 1) (const_int 5)])))] + "TARGET_ALTIVEC && !BYTES_BIG_ENDIAN" "@ xxmrglw %x0,%x1,%x2 vmrglw %0,%1,%2" @@ -3861,13 +3893,13 @@ (define_expand "vec_widen_umult_hi_v8hi" { emit_insn (gen_altivec_vmuleuh (ve, operands[1], operands[2])); emit_insn (gen_altivec_vmulouh (vo, operands[1], operands[2])); - emit_insn (gen_altivec_vmrghw_direct_v4si (operands[0], ve, vo)); + emit_insn (gen_altivec_vmrghw (operands[0], ve, vo)); } else { emit_insn (gen_altivec_vmulouh (ve, operands[1], operands[2])); emit_insn (gen_altivec_vmuleuh (vo, operands[1], operands[2])); - emit_insn (gen_altivec_vmrghw_direct_v4si (operands[0], vo, ve)); + emit_insn (gen_altivec_vmrglw (operands[0], ve, vo)); } DONE; }) @@ -3886,13 +3918,13 @@ (define_expand "vec_widen_umult_lo_v8hi" { emit_insn (gen_altivec_vmuleuh (ve, operands[1], operands[2])); emit_insn (gen_altivec_vmulouh (vo, operands[1], operands[2])); - emit_insn (gen_altivec_vmrglw_direct_v4si (operands[0], ve, vo)); + emit_insn (gen_altivec_vmrglw (operands[0], ve, vo)); } else { emit_insn (gen_altivec_vmulouh (ve, operands[1], operands[2])); emit_insn (gen_altivec_vmuleuh (vo, operands[1], operands[2])); - emit_insn (gen_altivec_vmrglw_direct_v4si (operands[0], vo, ve)); + emit_insn (gen_altivec_vmrghw (operands[0], ve, vo)); } DONE; }) @@ -3911,13 +3943,13 @@ (define_expand "vec_widen_smult_hi_v8hi" { emit_insn (gen_altivec_vmulesh (ve, operands[1], operands[2])); emit_insn (gen_altivec_vmulosh (vo, operands[1], operands[2])); - emit_insn (gen_altivec_vmrghw_direct_v4si (operands[0], ve, vo)); + emit_insn (gen_altivec_vmrghw (operands[0], ve, vo)); } else { emit_insn (gen_altivec_vmulosh (ve, operands[1], operands[2])); emit_insn (gen_altivec_vmulesh (vo, operands[1], operands[2])); - emit_insn (gen_altivec_vmrghw_direct_v4si (operands[0], vo, ve)); + emit_insn (gen_altivec_vmrglw (operands[0], ve, vo)); } DONE; }) @@ -3936,13 +3968,13 @@ (define_expand "vec_widen_smult_lo_v8hi" { emit_insn (gen_altivec_vmulesh (ve, operands[1], operands[2])); emit_insn (gen_altivec_vmulosh (vo, operands[1], operands[2])); - emit_insn (gen_altivec_vmrglw_direct_v4si (operands[0], ve, vo)); + emit_insn (gen_altivec_vmrglw (operands[0], ve, vo)); } else { emit_insn (gen_altivec_vmulosh (ve, operands[1], operands[2])); emit_insn (gen_altivec_vmulesh (vo, operands[1], operands[2])); - emit_insn (gen_altivec_vmrglw_direct_v4si (operands[0], vo, ve)); + emit_insn (gen_altivec_vmrghw (operands[0], ve, vo)); } DONE; }) diff --git a/gcc/config/rs6000/rs6000.cc b/gcc/config/rs6000/rs6000.cc index e4dc629ddcc..2046a831938 100644 --- a/gcc/config/rs6000/rs6000.cc +++ b/gcc/config/rs6000/rs6000.cc @@ -23450,8 +23450,8 @@ altivec_expand_vec_perm_const (rtx target, rtx op0, rtx op1, : CODE_FOR_altivec_vmrglh_direct, {0, 1, 16, 17, 2, 3, 18, 19, 4, 5, 20, 21, 6, 7, 22, 23}}, {OPTION_MASK_ALTIVEC, - BYTES_BIG_ENDIAN ? CODE_FOR_altivec_vmrghw_direct_v4si - : CODE_FOR_altivec_vmrglw_direct_v4si, + BYTES_BIG_ENDIAN ? CODE_FOR_altivec_vmrghw_direct_v4si_be + : CODE_FOR_altivec_vmrglw_direct_v4si_le, {0, 1, 2, 3, 16, 17, 18, 19, 4, 5, 6, 7, 20, 21, 22, 23}}, {OPTION_MASK_ALTIVEC, BYTES_BIG_ENDIAN ? CODE_FOR_altivec_vmrglb_direct @@ -23462,8 +23462,8 @@ altivec_expand_vec_perm_const (rtx target, rtx op0, rtx op1, : CODE_FOR_altivec_vmrghh_direct, {8, 9, 24, 25, 10, 11, 26, 27, 12, 13, 28, 29, 14, 15, 30, 31}}, {OPTION_MASK_ALTIVEC, - BYTES_BIG_ENDIAN ? CODE_FOR_altivec_vmrglw_direct_v4si - : CODE_FOR_altivec_vmrghw_direct_v4si, + BYTES_BIG_ENDIAN ? CODE_FOR_altivec_vmrglw_direct_v4si_be + : CODE_FOR_altivec_vmrghw_direct_v4si_le, {8, 9, 10, 11, 24, 25, 26, 27, 12, 13, 14, 15, 28, 29, 30, 31}}, {OPTION_MASK_P8_VECTOR, BYTES_BIG_ENDIAN ? CODE_FOR_p8_vmrgew_v4sf_direct diff --git a/gcc/config/rs6000/vsx.md b/gcc/config/rs6000/vsx.md index f135fa079bd..7a9c19ac903 100644 --- a/gcc/config/rs6000/vsx.md +++ b/gcc/config/rs6000/vsx.md @@ -4798,12 +4798,14 @@ (define_expand "vsx_xxmrghw_<mode>" (const_int 1) (const_int 5)])))] "VECTOR_MEM_VSX_P (<MODE>mode)" { - rtx (*fun) (rtx, rtx, rtx); - fun = BYTES_BIG_ENDIAN ? gen_altivec_vmrghw_direct_<mode> - : gen_altivec_vmrglw_direct_<mode>; - if (!BYTES_BIG_ENDIAN) - std::swap (operands[1], operands[2]); - emit_insn (fun (operands[0], operands[1], operands[2])); + if (BYTES_BIG_ENDIAN) + emit_insn (gen_altivec_vmrghw_direct_v4si_be (operands[0], + operands[1], + operands[2])); + else + emit_insn (gen_altivec_vmrglw_direct_v4si_le (operands[0], + operands[2], + operands[1])); DONE; } [(set_attr "type" "vecperm")]) @@ -4818,12 +4820,14 @@ (define_expand "vsx_xxmrglw_<mode>" (const_int 3) (const_int 7)])))] "VECTOR_MEM_VSX_P (<MODE>mode)" { - rtx (*fun) (rtx, rtx, rtx); - fun = BYTES_BIG_ENDIAN ? gen_altivec_vmrglw_direct_<mode> - : gen_altivec_vmrghw_direct_<mode>; - if (!BYTES_BIG_ENDIAN) - std::swap (operands[1], operands[2]); - emit_insn (fun (operands[0], operands[1], operands[2])); + if (BYTES_BIG_ENDIAN) + emit_insn (gen_altivec_vmrglw_direct_v4si_be (operands[0], + operands[1], + operands[2])); + else + emit_insn (gen_altivec_vmrghw_direct_v4si_le (operands[0], + operands[2], + operands[1])); DONE; } [(set_attr "type" "vecperm")]) diff --git a/gcc/testsuite/g++.target/powerpc/pr106069.C b/gcc/testsuite/g++.target/powerpc/pr106069.C new file mode 100644 index 00000000000..537207d2fe8 --- /dev/null +++ b/gcc/testsuite/g++.target/powerpc/pr106069.C @@ -0,0 +1,119 @@ +/* { dg-options "-O -fno-tree-forwprop -maltivec" } */ +/* { dg-require-effective-target vmx_hw } */ +/* { dg-do run } */ + +typedef __attribute__ ((altivec (vector__))) unsigned native_simd_type; + +union +{ + native_simd_type V; + int R[4]; +} store_le_vec; + +struct S +{ + S () = default; + S (unsigned B0) + { + native_simd_type val{B0}; + m_simd = val; + } + void store_le (unsigned int out[]) + { + store_le_vec.V = m_simd; + unsigned int x0 = store_le_vec.R[0]; + __builtin_memcpy (out, &x0, 4); + } + S rotl (unsigned int r) + { + native_simd_type rot{r}; + return __builtin_vec_rl (m_simd, rot); + } + void operator+= (S other) + { + m_simd = __builtin_vec_add (m_simd, other.m_simd); + } + void operator^= (S other) + { + m_simd = __builtin_vec_xor (m_simd, other.m_simd); + } + static void transpose (S &B0, S B1, S B2, S B3) + { + native_simd_type T0 = __builtin_vec_mergeh (B0.m_simd, B2.m_simd); + native_simd_type T1 = __builtin_vec_mergeh (B1.m_simd, B3.m_simd); + native_simd_type T2 = __builtin_vec_mergel (B0.m_simd, B2.m_simd); + native_simd_type T3 = __builtin_vec_mergel (B1.m_simd, B3.m_simd); + B0 = __builtin_vec_mergeh (T0, T1); + B3 = __builtin_vec_mergel (T2, T3); + } + S (native_simd_type x) : m_simd (x) {} + native_simd_type m_simd; +}; + +void +foo (unsigned int output[], unsigned state[]) +{ + S R00 = state[0]; + S R01 = state[0]; + S R02 = state[2]; + S R03 = state[0]; + S R05 = state[5]; + S R06 = state[6]; + S R07 = state[7]; + S R08 = state[8]; + S R09 = state[9]; + S R10 = state[10]; + S R11 = state[11]; + S R12 = state[12]; + S R13 = state[13]; + S R14 = state[4]; + S R15 = state[15]; + for (int r = 0; r != 10; ++r) + { + R09 += R13; + R11 += R15; + R05 ^= R09; + R06 ^= R10; + R07 ^= R11; + R07 = R07.rotl (7); + R00 += R05; + R01 += R06; + R02 += R07; + R15 ^= R00; + R12 ^= R01; + R13 ^= R02; + R00 += R05; + R01 += R06; + R02 += R07; + R15 ^= R00; + R12 = R12.rotl (8); + R13 = R13.rotl (8); + R10 += R15; + R11 += R12; + R08 += R13; + R09 += R14; + R05 ^= R10; + R06 ^= R11; + R07 ^= R08; + R05 = R05.rotl (7); + R06 = R06.rotl (7); + R07 = R07.rotl (7); + } + R00 += state[0]; + S::transpose (R00, R01, R02, R03); + R00.store_le (output); +} + +unsigned int res[1]; +unsigned main_state[]{1634760805, 60878, 2036477234, 6, + 0, 825562964, 1471091955, 1346092787, + 506976774, 4197066702, 518848283, 118491664, + 0, 0, 0, 0}; +int +main () +{ + foo (res, main_state); + if (res[0] != 0x41fcef98) + __builtin_abort (); + return 0; +} diff --git a/gcc/testsuite/gcc.target/powerpc/pr115355.c b/gcc/testsuite/gcc.target/powerpc/pr115355.c new file mode 100644 index 00000000000..8955126b808 --- /dev/null +++ b/gcc/testsuite/gcc.target/powerpc/pr115355.c @@ -0,0 +1,37 @@ +/* { dg-do run } */ +/* { dg-require-effective-target p9vector_hw } */ +/* Force vectorization with -fno-vect-cost-model to have vector unpack + which exposes the issue in PR115355. */ +/* { dg-options "-O2 -mdejagnu-cpu=power9 -fno-vect-cost-model" } */ + +/* Verify it runs successfully. */ + +__attribute__((noipa)) +void setToIdentityGOOD(unsigned long long *mVec, unsigned int mLen) +{ + #pragma GCC novector + for (unsigned int i = 0; i < mLen; i++) + mVec[i] = i; +} + +__attribute__((noipa)) +void setToIdentityBAD(unsigned long long *mVec, unsigned int mLen) +{ + for (unsigned int i = 0; i < mLen; i++) + mVec[i] = i; +} + +unsigned long long vec1[100]; +unsigned long long vec2[100]; + +int main() +{ + unsigned int l = 29; + setToIdentityGOOD (vec1, 29); + setToIdentityBAD (vec2, 29); + + if (__builtin_memcmp (vec1, vec2, l * sizeof (vec1[0])) != 0) + __builtin_abort (); + + return 0; +} -- 2.43.0