| Message ID | 01fd01d8a5ac$bfe5dc60$3fb19520$@nextmovesoftware.com |
|---|---|
| State | New |
| Headers | show |
| Series | [x86_64] PR target/106481: Handle CONST_WIDE_INT in REG_EQUAL during STV. | expand |
On Mon, Aug 1, 2022 at 3:43 PM Roger Sayle <roger@nextmovesoftware.com> wrote: > > > This patch resolves PR target/106481, and is an oversight in my recent > battles with REG_EQUAL notes during TImode STV (see PR target/106278 > https://gcc.gnu.org/pipermail/gcc-patches/2022-July/598416.html). > > The patch above's/current behaviour is that we check that the mode of > the REG_EQUAL note is TImode before using PUT_MODE to set it to V1TImode. > However, the new test case reveals that this doesn't consider REG_EQUAL > notes that are CONST_INT or CONST_WIDE_INT, i.e. that are VOIDmode, > and so STV produces: > > (insn 85 84 86 2 (set (reg:V1TI 113) > (reg:V1TI 84)) "pr106481.c":13:3 1766 {movv1ti_internal} > (expr_list:REG_EQUAL (const_wide_int 0x0ffffffff00000004) > (nil))) > > which causes problems as the const_wide_int isn't a valid immediate > constant for V1TImode. With this patch, we now generate the correct: > > (insn 85 84 86 2 (set (reg:V1TI 113) > (reg:V1TI 84)) "pr106481.c":13:3 1766 {movv1ti_internal} > (expr_list:REG_EQUAL (const_vector:V1TI [ > (const_wide_int 0x0ffffffff00000004) > ]) > (nil))) > > > This patch has been tested on x86_64-pc-linux-gnu with make bootstrap > and make -k check, both with and without --target_board=unix{-m32}, > with no new failures. Ok for mainline? > > > 2022-08-01 Roger Sayle <roger@nextmovesoftware.com> > > gcc/ChangeLog > PR target/106481 > * config/i386/i386-features.cc (timode_scalar_chain::convert_insn): > Convert a CONST_SCALAR_INT_P in a REG_EQUAL note into a V1TImode > CONST_VECTOR. > > gcc/testsuite/ChangeLog > PR target/106481 > * gcc.target/i386/pr106481.c: New test case. OK, but how about we restructure the conditions a bit, like in the attached patch? Uros. diff --git a/gcc/config/i386/i386-features.cc b/gcc/config/i386/i386-features.cc index e4cc4a30590..e3ecc22580a 100644 --- a/gcc/config/i386/i386-features.cc +++ b/gcc/config/i386/i386-features.cc @@ -1353,8 +1353,15 @@ timode_scalar_chain::convert_insn (rtx_insn *insn) if (GET_MODE (dst) == V1TImode) { tmp = find_reg_equal_equiv_note (insn); - if (tmp && GET_MODE (XEXP (tmp, 0)) == TImode) - PUT_MODE (XEXP (tmp, 0), V1TImode); + if (tmp) + { + if (GET_MODE (XEXP (tmp, 0)) == TImode) + PUT_MODE (XEXP (tmp, 0), V1TImode); + else if (CONST_SCALAR_INT_P (XEXP (tmp, 0))) + XEXP (tmp, 0) + = gen_rtx_CONST_VECTOR (V1TImode, + gen_rtvec (1, XEXP (tmp, 0))); + } } break; case MEM:
diff --git a/gcc/config/i386/i386-features.cc b/gcc/config/i386/i386-features.cc index aa5de71..c1564c4 100644 --- a/gcc/config/i386/i386-features.cc +++ b/gcc/config/i386/i386-features.cc @@ -1355,6 +1355,10 @@ timode_scalar_chain::convert_insn (rtx_insn *insn) tmp = find_reg_equal_equiv_note (insn); if (tmp && GET_MODE (XEXP (tmp, 0)) == TImode) PUT_MODE (XEXP (tmp, 0), V1TImode); + else if (tmp && CONST_SCALAR_INT_P (XEXP (tmp, 0))) + XEXP (tmp, 0) = gen_rtx_CONST_VECTOR (V1TImode, + gen_rtvec (1, + XEXP (tmp, 0))); } break; case MEM: diff --git a/gcc/testsuite/gcc.target/i386/pr106481.c b/gcc/testsuite/gcc.target/i386/pr106481.c new file mode 100644 index 0000000..8cc7048 --- /dev/null +++ b/gcc/testsuite/gcc.target/i386/pr106481.c @@ -0,0 +1,17 @@ +/* { dg-do compile { target int128 } } */ +/* { dg-options "-O2 -fno-dce -fno-forward-propagate -fno-rerun-cse-after-loop -Wno-psabi" } */ + +typedef int V __attribute__((vector_size (64))); +typedef __int128 W __attribute__((vector_size (64))); + +W w; +V bar (void); + +void +foo (V v, W) +{ + foo ((V){4, ~0}, (W) v); + foo (v, w); + bar (); +} +
This patch resolves PR target/106481, and is an oversight in my recent battles with REG_EQUAL notes during TImode STV (see PR target/106278 https://gcc.gnu.org/pipermail/gcc-patches/2022-July/598416.html). The patch above's/current behaviour is that we check that the mode of the REG_EQUAL note is TImode before using PUT_MODE to set it to V1TImode. However, the new test case reveals that this doesn't consider REG_EQUAL notes that are CONST_INT or CONST_WIDE_INT, i.e. that are VOIDmode, and so STV produces: (insn 85 84 86 2 (set (reg:V1TI 113) (reg:V1TI 84)) "pr106481.c":13:3 1766 {movv1ti_internal} (expr_list:REG_EQUAL (const_wide_int 0x0ffffffff00000004) (nil))) which causes problems as the const_wide_int isn't a valid immediate constant for V1TImode. With this patch, we now generate the correct: (insn 85 84 86 2 (set (reg:V1TI 113) (reg:V1TI 84)) "pr106481.c":13:3 1766 {movv1ti_internal} (expr_list:REG_EQUAL (const_vector:V1TI [ (const_wide_int 0x0ffffffff00000004) ]) (nil))) This patch has been tested on x86_64-pc-linux-gnu with make bootstrap and make -k check, both with and without --target_board=unix{-m32}, with no new failures. Ok for mainline? 2022-08-01 Roger Sayle <roger@nextmovesoftware.com> gcc/ChangeLog PR target/106481 * config/i386/i386-features.cc (timode_scalar_chain::convert_insn): Convert a CONST_SCALAR_INT_P in a REG_EQUAL note into a V1TImode CONST_VECTOR. gcc/testsuite/ChangeLog PR target/106481 * gcc.target/i386/pr106481.c: New test case. Thanks again, Roger --