Patchwork ARM: Permit any ARMv6K CPU to read the MVFR0 and MVFR1 VFP registers.

login
register
mail settings
Submitter Andrew Towers
Date March 28, 2012, 2:37 p.m.
Message ID <1332945448-31266-1-git-send-email-atowers@gmail.com>
Download mbox | patch
Permalink /patch/149256/
State New
Headers show

Comments

Andrew Towers - March 28, 2012, 2:37 p.m.
From: Andrew Towers <atowers@gmail.com>

Replaces the ARM_FEATURE_VFP3 check when reading MVFR0/1 with a check for
ARM_FEATURE_V6K. Rationale: MVFR0/1 were introduced in the ARM1136 at the
same time as ARMv6K, and a survey of TRMs indicates support in later models.

According to reference documentation on arm.com, MVFR0 and MVFR1 were
introduced in ARM1136JF-S r1p1 (ARMv6K, VFPv2). They are also present in
ARM1156T2F-S and ARM1176JZF-S, which contain VFP11 r5, and in ARM11 MPCore
r1 which contains VFP11 r4.

Reference (ARM DDI 0211H, 0290G, 0301H, 0360E)
http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0211h/Ffbefjag.html

Without this change, the linux kernel will not boot with VFP support enabled
under ARM1176 system emulation, due to the unconditional use of MVFR1 at the
end of vfp_init() in arch/arm/vfp/vfpmodule.c in the kernel:

  VFP support v0.3: implemetor 41 architecture 1 part 20 variant b rev 5
  Internal error: Oops - undefined instruction: 0 [#1]

Yes, I am slightly abusing the versatilepb hw emulation, but then so is
everyone else in the world who is still waiting for their Raspberry Pi ;)

Signed-off-by: Andrew Towers <atowers@gmail.com>
---
 target-arm/translate.c |    2 +-
 1 files changed, 1 insertions(+), 1 deletions(-)
Peter Maydell - March 28, 2012, 3:24 p.m.
On 28 March 2012 15:37, Andrew Towers <atowers@gmail.com> wrote:
> From: Andrew Towers <atowers@gmail.com>
>
> Replaces the ARM_FEATURE_VFP3 check when reading MVFR0/1 with a check for
> ARM_FEATURE_V6K. Rationale: MVFR0/1 were introduced in the ARM1136 at the
> same time as ARMv6K, and a survey of TRMs indicates support in later models.
>
> According to reference documentation on arm.com, MVFR0 and MVFR1 were
> introduced in ARM1136JF-S r1p1 (ARMv6K, VFPv2). They are also present in
> ARM1156T2F-S and ARM1176JZF-S, which contain VFP11 r5, and in ARM11 MPCore
> r1 which contains VFP11 r4.
>
> Reference (ARM DDI 0211H, 0290G, 0301H, 0360E)
> http://infocenter.arm.com/help/index.jsp?topic=/com.arm.doc.ddi0211h/Ffbefjag.html

Unfortunately I think the ARM1156 is the exception which demonstrates that
this isn't a V6K feature. The 1156 isn't V6K (it doesn't have the TLS registers
or the byte/half/doubleword load/store exclusive instructions), but it does
have the MVFR* registers. So I think we need a new ARM_FEATURE_MVFR flag, and
in cpu_reset_model_id() have V6K imply MVFR.

> Yes, I am slightly abusing the versatilepb hw emulation, but then so is
> everyone else in the world who is still waiting for their Raspberry Pi ;)

...I've been wondering if anybody's going to submit patches for an actual
rPi board model :-)

-- PMM

Patch

diff --git a/target-arm/translate.c b/target-arm/translate.c
index 81725d1..b5861c8 100644
--- a/target-arm/translate.c
+++ b/target-arm/translate.c
@@ -2906,7 +2906,7 @@  static int disas_vfp_insn(CPUARMState * env, DisasContext *s, uint32_t insn)
                         case ARM_VFP_MVFR0:
                         case ARM_VFP_MVFR1:
                             if (IS_USER(s)
-                                || !arm_feature(env, ARM_FEATURE_VFP3))
+                                || !arm_feature(env, ARM_FEATURE_V6K))
                                 return 1;
                             tmp = load_cpu_field(vfp.xregs[rn]);
                             break;