[OE-core] [PATCH] gcc: upgrade to 4.9.2 and fix the bug #6824
Peter A. Bigot
pab at pabigot.com
Mon Jan 12 12:38:20 UTC 2015
On 01/11/2015 09:12 PM, Lei Maohui wrote:
> fix the yocto bug #6824
> backport from gcc
>
> Signed-off-by: Lei Maohui <leimaohui at cn.fujitsu.com>
> ---
> meta/recipes-devtools/gcc/gcc-4.9.inc | 12 +-
> ...e500-double-in-SPE_SIMD_REGNO_P-registers.patch | 13 ++
> .../0061-Fix-for-unwinder-aborts-on-e500.patch | 161 +++++++++++++++++++++
> 3 files changed, 180 insertions(+), 6 deletions(-)
> create mode 100644 meta/recipes-devtools/gcc/gcc-4.9/0060-Only-allow-e500-double-in-SPE_SIMD_REGNO_P-registers.patch
> create mode 100644 meta/recipes-devtools/gcc/gcc-4.9/0061-Fix-for-unwinder-aborts-on-e500.patch
>
> diff --git a/meta/recipes-devtools/gcc/gcc-4.9.inc b/meta/recipes-devtools/gcc/gcc-4.9.inc
> index 2568e99..c4ec40d 100644
> --- a/meta/recipes-devtools/gcc/gcc-4.9.inc
> +++ b/meta/recipes-devtools/gcc/gcc-4.9.inc
> @@ -2,11 +2,11 @@ require gcc-common.inc
>
> # Third digit in PV should be incremented after a minor release
>
> -PV = "4.9.1"
> +PV = "4.9.2"
>
> # BINV should be incremented to a revision after a minor gcc release
>
> -BINV = "4.9.1"
> +BINV = "4.9.2"
>
> FILESEXTRAPATHS =. "${FILE_DIRNAME}/gcc-4.9:"
>
> @@ -69,11 +69,11 @@ SRC_URI = "\
> file://0051-eabispe.patch \
> file://0053-gcc-fix-segfault-from-calling-free-on-non-malloc-d-a.patch \
> file://0054-gcc-Makefile.in-fix-parallel-building-failure.patch \
> - file://0055-PR-rtl-optimization-61801.patch \
> file://0056-top-level-reorder_gcc-bug-61144.patch \
> - file://0057-aarch64-config.patch \
> file://0058-gcc-r212171.patch \
> file://0059-gcc-PR-rtl-optimization-63348.patch \
> + file://0060-Only-allow-e500-double-in-SPE_SIMD_REGNO_P-registers.patch \
> + file://0061-Fix-for-unwinder-aborts-on-e500.patch \
> "
> SRC_URI[md5sum] = "fddf71348546af523353bd43d34919c1"
> SRC_URI[sha256sum] = "d334781a124ada6f38e63b545e2a3b8c2183049515a1abab6d513f109f1d717e"
> @@ -125,8 +125,8 @@ EXTRA_OECONF_INTERMEDIATE = "\
> EXTRA_OECONF_append_libc-uclibc = " --disable-decimal-float "
>
> EXTRA_OECONF_PATHS = "\
> - --with-gxx-include-dir=/not/exist{target_includedir}/c++/${BINV} \
> - --with-sysroot=/not/exist \
> + --with-gxx-include-dir=${STAGING_DIR_TARGET}${target_includedir}/c++/${BINV} \
> + --with-sysroot=${STAGING_DIR_TARGET} \
> --with-build-sysroot=${STAGING_DIR_TARGET} \
> "
>
> diff --git a/meta/recipes-devtools/gcc/gcc-4.9/0060-Only-allow-e500-double-in-SPE_SIMD_REGNO_P-registers.patch b/meta/recipes-devtools/gcc/gcc-4.9/0060-Only-allow-e500-double-in-SPE_SIMD_REGNO_P-registers.patch
As gcc is (IMO) one of the most critical components of Yocto I think we
need to be very clear about what changes are being made to it and why.
In addition to missing SOB lines, these patches at a minimum should
reference the upstream bug that they supposedly patch. That would be
https://gcc.gnu.org/bugzilla/show_bug.cgi?id=63908
Ideally, the patch would be taken from upstream, being created by
git-format-patch with the SHA1 of the relevant change as retrieved from:
git://gcc.gnu.org/git/gcc.git
It worries me that these patches, though present in upstream trunk, have
not been back-ported to gcc-4_9-branch. There may be a reason for that,
and the fact backports have not (yet) been done should be noted in the
OE patch header.
This first patch is apparently upstream commit
5c0092070253113cf0d9c45eacc884b3ecc34d81.
> new file mode 100644
> index 0000000..810ff90
> --- /dev/null
> +++ b/meta/recipes-devtools/gcc/gcc-4.9/0060-Only-allow-e500-double-in-SPE_SIMD_REGNO_P-registers.patch
> @@ -0,0 +1,13 @@
> +diff --git a/gcc/config/rs6000/rs6000.c b/gcc/config/rs6000/rs6000.c
> +index 512e7d6..292fb2c 100644
> +--- a/gcc/config/rs6000/rs6000.c
> ++++ b/gcc/config/rs6000/rs6000.c
> +@@ -1704,7 +1704,7 @@ rs6000_hard_regno_nregs_internal (int regno, enum machine_mode mode)
> + SCmode so as to pass the value correctly in a pair of
> + registers. */
> + else if (TARGET_E500_DOUBLE && FLOAT_MODE_P (mode) && mode != SCmode
> +- && !DECIMAL_FLOAT_MODE_P (mode))
> ++ && !DECIMAL_FLOAT_MODE_P (mode) && SPE_SIMD_REGNO_P (regno))
> + reg_size = UNITS_PER_FP_WORD;
> +
> + else
> diff --git a/meta/recipes-devtools/gcc/gcc-4.9/0061-Fix-for-unwinder-aborts-on-e500.patch b/meta/recipes-devtools/gcc/gcc-4.9/0061-Fix-for-unwinder-aborts-on-e500.patch
> new file mode 100644
This patch appears to combine c0235a33de8c4f78cce35b2a8c2035c83fe1bd14
with 4fd39f1329379e00f958394adde6be96f0caf21f.
> index 0000000..c37a4c3
> --- /dev/null
> +++ b/meta/recipes-devtools/gcc/gcc-4.9/0061-Fix-for-unwinder-aborts-on-e500.patch
> @@ -0,0 +1,161 @@
> +diff --git a/gcc/defaults.h b/gcc/defaults.h
> +index f94ae17..80a798f 100644
> +--- a/gcc/defaults.h
> ++++ b/gcc/defaults.h
> +@@ -438,6 +438,11 @@ see the files COPYING3 and COPYING.RUNTIME respectively. If not, see
> + #define DWARF_FRAME_REGNUM(REG) DBX_REGISTER_NUMBER (REG)
> + #endif
> +
> ++/* The mapping from dwarf CFA reg number to internal dwarf reg numbers. */
> ++#ifndef DWARF_REG_TO_UNWIND_COLUMN
> ++#define DWARF_REG_TO_UNWIND_COLUMN(REGNO) (REGNO)
> ++#endif
> ++
> + /* Map register numbers held in the call frame info that gcc has
> + collected using DWARF_FRAME_REGNUM to those that should be output in
> + .debug_frame and .eh_frame. */
> +diff --git a/gcc/dwarf2cfi.c b/gcc/dwarf2cfi.c
> +index abcdeb3..4e59dfe 100644
> +--- a/gcc/dwarf2cfi.c
> ++++ b/gcc/dwarf2cfi.c
> +@@ -252,7 +252,60 @@ init_return_column_size (enum machine_mode mode, rtx mem, unsigned int c)
> + gen_int_mode (size, mode));
> + }
> +
> +-/* Generate code to initialize the register size table. */
> ++/* Datastructure used by expand_builtin_init_dwarf_reg_sizes and
> ++ init_one_dwarf_reg_size to communicate on what has been done by the
> ++ latter. */
> ++
> ++typedef struct
> ++{
> ++ /* Whether the dwarf return column was initialized. */
> ++ bool wrote_return_column;
> ++
> ++ /* For each hard register REGNO, whether init_one_dwarf_reg_size
> ++ was given REGNO to process already. */
> ++ bool processed_regno [FIRST_PSEUDO_REGISTER];
> ++
> ++} init_one_dwarf_reg_state;
> ++
> ++/* Helper for expand_builtin_init_dwarf_reg_sizes. Generate code to
> ++ initialize the dwarf register size table entry corresponding to register
> ++ REGNO in REGMODE. TABLE is the table base address, SLOTMODE is the mode to
> ++ use for the size entry to initialize, and INIT_STATE is the communication
> ++ datastructure conveying what we're doing to our caller. */
> ++
> ++static
> ++void init_one_dwarf_reg_size (int regno, machine_mode regmode,
> ++ rtx table, machine_mode slotmode,
> ++ init_one_dwarf_reg_state *init_state)
> ++{
> ++ const unsigned int dnum = DWARF_FRAME_REGNUM (regno);
> ++ const unsigned int rnum = DWARF2_FRAME_REG_OUT (dnum, 1);
> ++ const unsigned int dcol = DWARF_REG_TO_UNWIND_COLUMN (rnum);
> ++
> ++ const HOST_WIDE_INT slotoffset = dcol * GET_MODE_SIZE (slotmode);
> ++ const HOST_WIDE_INT regsize = GET_MODE_SIZE (regmode);
> ++
> ++ init_state->processed_regno[regno] = true;
> ++
> ++ if (rnum >= DWARF_FRAME_REGISTERS)
> ++ return;
> ++
> ++ if (dnum == DWARF_FRAME_RETURN_COLUMN)
> ++ {
> ++ if (regmode == VOIDmode)
> ++ return;
> ++ init_state->wrote_return_column = true;
> ++ }
> ++
> ++ if (slotoffset < 0)
> ++ return;
> ++
> ++ emit_move_insn (adjust_address (table, slotmode, slotoffset),
> ++ gen_int_mode (regsize, slotmode));
> ++}
> ++
> ++/* Generate code to initialize the dwarf register size table located
> ++ at the provided ADDRESS. */
> +
> + void
> + expand_builtin_init_dwarf_reg_sizes (tree address)
> +@@ -261,37 +314,41 @@ expand_builtin_init_dwarf_reg_sizes (tree address)
> + enum machine_mode mode = TYPE_MODE (char_type_node);
> + rtx addr = expand_normal (address);
> + rtx mem = gen_rtx_MEM (BLKmode, addr);
> +- bool wrote_return_column = false;
> ++
> ++ init_one_dwarf_reg_state init_state;
> ++ memset ((char *)&init_state, 0, sizeof (init_state));
> +
> + for (i = 0; i < FIRST_PSEUDO_REGISTER; i++)
> + {
> +- unsigned int dnum = DWARF_FRAME_REGNUM (i);
> +- unsigned int rnum = DWARF2_FRAME_REG_OUT (dnum, 1);
> ++ machine_mode save_mode;
> ++ rtx span;
> +
> +- if (rnum < DWARF_FRAME_REGISTERS)
> +- {
> +- HOST_WIDE_INT offset = rnum * GET_MODE_SIZE (mode);
> +- enum machine_mode save_mode = reg_raw_mode[i];
> +- HOST_WIDE_INT size;
> ++ /* No point in processing a register multiple times. This could happen
> ++ with register spans, e.g. when a reg is first processed as a piece of
> ++ a span, then as a register on its own later on. */
> +
> +- if (HARD_REGNO_CALL_PART_CLOBBERED (i, save_mode))
> +- save_mode = choose_hard_reg_mode (i, 1, true);
> +- if (dnum == DWARF_FRAME_RETURN_COLUMN)
> +- {
> +- if (save_mode == VOIDmode)
> +- continue;
> +- wrote_return_column = true;
> +- }
> +- size = GET_MODE_SIZE (save_mode);
> +- if (offset < 0)
> +- continue;
> ++ if (init_state.processed_regno[i])
> ++ continue;
> +
> +- emit_move_insn (adjust_address (mem, mode, offset),
> +- gen_int_mode (size, mode));
> ++ save_mode = reg_raw_mode[i];
> ++ if (HARD_REGNO_CALL_PART_CLOBBERED (i, save_mode))
> ++ save_mode = choose_hard_reg_mode (i, 1, true);
> ++
> ++ span = targetm.dwarf_register_span (gen_rtx_REG (save_mode, i));
> ++ if (!span)
> ++ init_one_dwarf_reg_size (i, save_mode, mem, mode, &init_state);
> ++ else
> ++ {
> ++ for (int si = 0; si < XVECLEN (span, 0); si++)
> ++ {
> ++ rtx reg = XVECEXP (span, 0, si);
> ++ init_one_dwarf_reg_size
> ++ (REGNO (reg), GET_MODE (reg), mem, mode, &init_state);
> ++ }
> + }
> + }
> +
> +- if (!wrote_return_column)
> ++ if (!init_state.wrote_return_column)
> + init_return_column_size (mode, mem, DWARF_FRAME_RETURN_COLUMN);
> +
> + #ifdef DWARF_ALT_FRAME_RETURN_COLUMN
> +diff --git a/libgcc/unwind-dw2.c b/libgcc/unwind-dw2.c
> +index 55fc4bc..37f0ae2 100644
> +--- a/libgcc/unwind-dw2.c
> ++++ b/libgcc/unwind-dw2.c
> +@@ -55,10 +55,6 @@
> + #define PRE_GCC3_DWARF_FRAME_REGISTERS DWARF_FRAME_REGISTERS
> + #endif
> +
> +-#ifndef DWARF_REG_TO_UNWIND_COLUMN
> +-#define DWARF_REG_TO_UNWIND_COLUMN(REGNO) (REGNO)
> +-#endif
> +-
> + /* ??? For the public function interfaces, we tend to gcc_assert that the
> + column numbers are in range. For the dwarf2 unwind info this does happen,
> + although so far in a case that doesn't actually matter.
More information about the Openembedded-core
mailing list