https://bugs.kde.org/show_bug.cgi?id=281975
--- Comment #7 from Markus Gothe <markus.go...@genexis.eu> --- Comment on attachment 167661 --> https://bugs.kde.org/attachment.cgi?id=167661 Support for ARMv7 w/o NEON and VFP >From 8fbd691e7f8a79983b8dbec1ee89cdf32d6b3c41 Mon Sep 17 00:00:00 2001 From: Dave Lerner <dave.ler...@windriver.com> Date: Fri, 22 Mar 2024 17:34:15 +0100 Subject: [PATCH] Support ARMv7 w/o NEON and VFP Don't try to cleanup the fpscr register that is not available and has no opcode support for armv7 processors that are without the SIMD and VFP extensions, since the fmxr/fmrx instructions will throw a SIGILL, illegal opcode, killing valgrind before it executes the first instruction of the application under test. Make it possible to disable NEON and VFP auto-detection on ARMv7. There are asymmetrical devices where one or more cores doesn't support the NEON and VFP extensions. These devices needs a special workaround. Signed-off-by: Dave Lerner <dave.ler...@windriver.com> Co-authored-by: Markus Gothe <markus.go...@genexis.eu> --- configure.ac | 9 +++++++++ coregrind/m_dispatch/dispatch-arm-linux.S | 11 +++++++++++ coregrind/m_machine.c | 13 ++++++++++--- coregrind/pub_core_machine.h | 1 + 4 files changed, 31 insertions(+), 3 deletions(-) diff --git a/configure.ac b/configure.ac index 8279683ac..0d0a35089 100755 --- a/configure.ac +++ b/configure.ac @@ -3789,6 +3789,15 @@ CFLAGS="$save_CFLAGS" AM_CONDITIONAL(BUILD_ARMV82_DOTPROD_TESTS, test x$ac_have_armv82_dotprod_feature = xyes) +# Enable VFP and NEON auto-detection on ARM +AC_CACHE_CHECK([for VFP auto-detection on ARM], vg_cv_arm_vfp_autodetection, [vg_cv_arm_vfp_autodetection=yes]) +AC_CACHE_CHECK([for NEON auto-detection on ARM], vg_cv_arm_neon_autodetection, [vg_cv_arm_neon_autodetection=yes]) +if test "x${vg_cv_arm_vfp_autodetection}" = "xyes"; then + AC_DEFINE(ENABLE_ARM_VFP, 1, [Define to 1 if auto-detection of ARM VFP capabilities should be used.]) +fi +if test "x${vg_cv_arm_neon_autodetection}" = "xyes"; then + AC_DEFINE(ENABLE_ARM_NEON, 1, [Define to 1 if auto-detection of ARM NEON capabilities should be used.]) +fi # XXX JRS 2010 Oct 13: what is this for? For sure, we don't need this # when building the tool executables. I think we should get rid of it. diff --git a/coregrind/m_dispatch/dispatch-arm-linux.S b/coregrind/m_dispatch/dispatch-arm-linux.S index 40629017c..284df0513 100644 --- a/coregrind/m_dispatch/dispatch-arm-linux.S +++ b/coregrind/m_dispatch/dispatch-arm-linux.S @@ -65,10 +65,16 @@ VG_(disp_run_translations): order to keep the stack 8-aligned. */ push {r0, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, lr} + ldr r4, =VG_(machine_arm_has_fpscr) + ldr r4, [r4] + cmp r4, #0 + beq no_fpscr_setup + /* set FPSCR to vex-required default value */ mov r4, #0 fmxr fpscr, r4 +no_fpscr_setup: /* Set up the guest state pointer */ mov r8, r1 @@ -93,6 +99,11 @@ postamble: /* We're leaving. Check that nobody messed with FPSCR in ways we don't expect. */ + ldr r4, =VG_(machine_arm_has_fpscr) + ldr r4, [r4] + cmp r4, #0 + beq remove_frame + fmrx r4, fpscr bic r4, #0xF8000000 /* mask out NZCV and QC */ bic r4, #0x0000009F /* mask out IDC,IXC,UFC,OFC,DZC,IOC */ diff --git a/coregrind/m_machine.c b/coregrind/m_machine.c index a4c2218bf..8fc82eb81 100644 --- a/coregrind/m_machine.c +++ b/coregrind/m_machine.c @@ -472,6 +472,7 @@ ULong VG_(machine_ppc64_has_VMX) = 0; #endif #if defined(VGA_arm) Int VG_(machine_arm_archlevel) = 4; +UInt VG_(machine_arm_has_fpscr) = 0; #endif @@ -1673,7 +1674,7 @@ Bool VG_(machine_get_hwcaps)( void ) tmp_sigfpe_act.sa_flags |= VKI_SA_NODEFER; tmp_sigfpe_act.ksa_handler = handler_unsup_insn; VG_(sigaction)(VKI_SIGFPE, &tmp_sigfpe_act, NULL); - +#ifdef ENABLE_ARM_VFP /* VFP insns */ have_VFP = True; if (VG_MINIMAL_SETJMP(env_unsup_insn)) { @@ -1681,11 +1682,14 @@ Bool VG_(machine_get_hwcaps)( void ) } else { __asm__ __volatile__(".word 0xEEB02B42"); /* VMOV.F64 d2, d2 */ } +#else + have_VFP = False; +#endif /* There are several generation of VFP extension but they differs very little so for now we will not distinguish them. */ have_VFP2 = have_VFP; have_VFP3 = have_VFP; - +#ifdef ENABLE_ARM_NEON /* NEON insns */ have_NEON = True; if (VG_MINIMAL_SETJMP(env_unsup_insn)) { @@ -1693,7 +1697,9 @@ Bool VG_(machine_get_hwcaps)( void ) } else { __asm__ __volatile__(".word 0xF2244154"); /* VMOV q2, q2 */ } - +#else + have_NEON = False; +#endif /* ARM architecture level */ archlevel = 5; /* v5 will be base level */ if (archlevel < 7) { @@ -1737,6 +1743,7 @@ Bool VG_(machine_get_hwcaps)( void ) (Int)have_NEON); VG_(machine_arm_archlevel) = archlevel; + VG_(machine_arm_has_fpscr) = (have_VFP || have_NEON) ? 1 : 0; va = VexArchARM; vai.endness = VexEndnessLE; diff --git a/coregrind/pub_core_machine.h b/coregrind/pub_core_machine.h index a9b7dd8b1..c8085df86 100644 --- a/coregrind/pub_core_machine.h +++ b/coregrind/pub_core_machine.h @@ -293,6 +293,7 @@ extern ULong VG_(machine_ppc64_has_VMX); #if defined(VGA_arm) extern Int VG_(machine_arm_archlevel); +extern UInt VG_(machine_arm_has_fpscr); #endif #endif // __PUB_CORE_MACHINE_H -- 2.43.2 -- You are receiving this mail because: You are watching all bug changes.
