Re: [PATCH v4 09/12] x86: Enable SSE in 64-bit mode

On Mon, Nov 13, 2023 at 03:28:13PM -0700, Simon Glass wrote:

Hi Bin,

On Mon, 13 Nov 2023 at 15:08, Bin Meng bmeng.cn@gmail.com wrote:

...

Hi Simon,

On Mon, Nov 13, 2023 at 4:03 AM Simon Glass sjg@chromium.org wrote:

...

This is needed to support Truetype fonts. In any case, the compiler expects SSE to be available in 64-bit mode. Provide an option to enable SSE so that hardware floating-point arithmetic works.

Signed-off-by: Simon Glass sjg@chromium.org Suggested-by: Bin Meng bmeng.cn@gmail.com

---

Changes in v4:

• Use a Kconfig option

arch/x86/Kconfig | 8 ++++++++ arch/x86/config.mk | 4 ++++ arch/x86/cpu/x86_64/cpu.c | 12 ++++++++++++ drivers/video/Kconfig | 1 + 4 files changed, 25 insertions(+)

diff --git a/arch/x86/Kconfig b/arch/x86/Kconfig index 99e59d94c606..6b532d712ee8 100644 --- a/arch/x86/Kconfig +++ b/arch/x86/Kconfig @@ -723,6 +723,14 @@ config ROM_TABLE_SIZE hex default 0x10000

+config X86_HARDFP

•   bool "Support hardware floating point"
  

•   help
  

•     U-Boot generally does not make use of floating point. Where this is
  

•     needed, it can be enabled using this option. This adjusts the
  

•     start-up code for 64-bit mode and changes the compiler options for
  

•     64-bit to enable SSE.
  

As discussed in another thread, this option should be made global to all architectures and by default no.

...

config HAVE_ITSS bool "Enable ITSS" help diff --git a/arch/x86/config.mk b/arch/x86/config.mk index 26ec1af2f0b0..2e3a7119e798 100644 --- a/arch/x86/config.mk +++ b/arch/x86/config.mk @@ -27,9 +27,13 @@ ifeq ($(IS_32BIT),y) PLATFORM_CPPFLAGS += -march=i386 -m32 else PLATFORM_CPPFLAGS += $(if $(CONFIG_SPL_BUILD),,-fpic) -fno-common -march=core2 -m64

+ifndef CONFIG_X86_HARDFP PLATFORM_CPPFLAGS += -mno-mmx -mno-sse endif

+endif # IS_32BIT

PLATFORM_RELFLAGS += -fdata-sections -ffunction-sections -fvisibility=hidden

KBUILD_LDFLAGS += -Bsymbolic -Bsymbolic-functions diff --git a/arch/x86/cpu/x86_64/cpu.c b/arch/x86/cpu/x86_64/cpu.c index 2647bff891f8..5ea746ecce4d 100644 --- a/arch/x86/cpu/x86_64/cpu.c +++ b/arch/x86/cpu/x86_64/cpu.c @@ -10,6 +10,7 @@ #include <init.h> #include <asm/cpu.h> #include <asm/global_data.h> +#include <asm/processor-flags.h>

DECLARE_GLOBAL_DATA_PTR;

@@ -39,11 +40,22 @@ int x86_mp_init(void) return 0; }

+/* enable SSE features for hardware floating point */ +static void setup_sse_features(void) +{

•   asm ("mov %%cr4, %%rax\n" \
  

•   "or  %0, %%rax\n" \
  

•   "mov %%rax, %%cr4\n" \
  

•   : : "i" (X86_CR4_OSFXSR | X86_CR4_OSXMMEXCPT) : "eax");
  

+}

int x86_cpu_reinit_f(void) { /* set the vendor to Intel so that native_calibrate_tsc() works */ gd->arch.x86_vendor = X86_VENDOR_INTEL; gd->arch.has_mtrr = true;

•   if (IS_ENABLED(CONFIG_X86_HARDFP))
  

•           setup_sse_features();
  
    return 0;
  

} diff --git a/drivers/video/Kconfig b/drivers/video/Kconfig index 6f319ba0d544..39c82521be16 100644 --- a/drivers/video/Kconfig +++ b/drivers/video/Kconfig @@ -180,6 +180,7 @@ config CONSOLE_ROTATION

config CONSOLE_TRUETYPE bool "Support a console that uses TrueType fonts"

•   select X86_HARDFP if X86
  

This should be "depends on HARDFP", indicating that the TrueType library is using hardware fp itself, and user has to explicitly turn the hardware fp Kconfig option on.

So you mean 'depends on HARDFP if X86' ? After all, this is only for X86 - other archs can use softfp which is already enabled, as I understand it.

...

"Select" does not work for architectures that does not have the "enabling hardware fp" logic in place.

...

    help
      TrueTrype fonts can provide outline-drawing capability rather than
      needing to provide a bitmap for each font and size that is needed.

--

I still don't think we are on the same page here. I would prefer to just enable the options without any option. I really don't want to get into RISC-V stuff - that is a separate concern.

From my POV it seems that x86 is special in that:

• it uses hardfp
• hardfp is always available in any CPU with 64-bit support (I think?)

Maybe the issue even is that on x86 we're being too imprecise in our build rules (and also on RISC-V, another issue). Today on x86 this fails because we say -mno-mmx -mno-sse and not also -msoft-float. I can just turn that on, on all x86 targets today and things build. Would that not also fix the truetype issue?