On 29 September 2016 at 14:45, Kyrill Tkachov <kyrylo.tkac...@foss.arm.com> wrote: > > On 29/09/16 09:45, Matthew Wahab wrote: >> >> The patch at https://gcc.gnu.org/ml/gcc-patches/2016-09/msg01975.html >> added constraints and "arch" attributes to some data movement patterns, >> to fix wrongly generating MOVW instructions when not supported by the >> targets. This was needed to fix a broken build but also resulted in MOVW >> instructions not being generated when they should be. >> >> This patch fixes the code-gen problems by removing the new attributes >> from *arm_movsi_vfp and *thumb2_movsi_vfp since they are unnecessary. >> It changes the rest of the added arch attributes from "t2", which >> specifies a Thumb-2 target, to the weaker "v6t2" which specifies a >> target that supports Thumb-2. >> >> Tested for arm-none-linux-gnueabihf with native bootstrap and make check >> on ARMv8-A and for arm-none-eabi with cross-compiled check-gcc on an >> ARMv8.2-A emulator. >> >> There is one unrelated failure in gcc.target/arm/fp16-aapcs-3.c, which >> is a recently added test for the FP16 ARM alternative format. This has >> dg-require-effective-target arm_fp16_ok >> which is true for +fp16 because that implies mfp16-format=ieee. The test >> should instead be requiring >> dg-require-effective-target arm_fp16_alternative_ok >> I'll send a patch to fix this. >> >> Ok for trunk? > > > Ok. > Thanks for the fixup, > Kyrill >
I started validations before you committed, and the results came back after. I noticed no difference. Thanks > > >> Matthew >> >> gcc/ >> 2016-09-29 Matthew Wahab <matthew.wa...@arm.com> >> >> * config/arm/arm.md (*arm_movsi_insn): Replace "t2" arch attribute >> with "v6t2". Move "arch" attribute above "pool_range". >> * config/arm/vfp.md (*arm_movhi_vfp): Likewise. >> (*thumb2_movhi_vfp): Likewise. >> (*arm_movhi_fp16): Likewise. >> (*thumb2_movhi_fp16): Likewise. >> (*arm_movsi_vfp): Remove "arch" attribute. >> (*thumb2_movsi_vfp): Likewise. > >
