public inbox for kvm@vger.kernel.org
 help / color / mirror / Atom feed
* [PATCH kvm-unit-tests] xsave: add testcase for emulation of AVX instructions
@ 2025-11-14  0:32 Paolo Bonzini
  2025-11-21 18:05 ` Sean Christopherson
  0 siblings, 1 reply; 2+ messages in thread
From: Paolo Bonzini @ 2025-11-14  0:32 UTC (permalink / raw)
  To: kvm; +Cc: kbusch

Companion patch to the emulator changes in KVM.  Funnily enough,
no valid combination involving AVX was tried.

Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
---
 x86/xsave.c | 45 ++++++++++++++++++++++++++++++++++++++++++++-
 1 file changed, 44 insertions(+), 1 deletion(-)

diff --git a/x86/xsave.c b/x86/xsave.c
index cc8e3a0a..e6d15938 100644
--- a/x86/xsave.c
+++ b/x86/xsave.c
@@ -15,6 +15,34 @@
 #define XSTATE_SSE      0x2
 #define XSTATE_YMM      0x4
 
+char __attribute__((aligned(32))) v32_1[32] = {
+    1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
+    128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143,
+};
+char __attribute__((aligned(32))) v32_2[32] = { 0 };
+
+static __attribute__((target("avx"))) void
+test_avx_fep(void)
+{
+	asm volatile("vzeroall\n"
+	    KVM_FEP "vmovdqa v32_1, %%ymm0\n"
+	    KVM_FEP "vmovdqa %%ymm0, v32_2\n" : : :
+	    "memory",
+	    "%ymm0", "%ymm1", "%ymm2", "%ymm3", "%ymm4", "%ymm5", "%ymm6", "%ymm7",
+	    "%ymm8", "%ymm9", "%ymm10", "%ymm11", "%ymm12", "%ymm13", "%ymm14", "%ymm15");
+}
+
+static __attribute__((target("avx"))) void
+test_avx(void)
+{
+	asm volatile("vzeroall\n"
+	    "vmovdqa v32_1, %%ymm0\n"
+	    "vmovdqa %%ymm0, v32_2\n" : : :
+	    "memory",
+	    "%ymm0", "%ymm1", "%ymm2", "%ymm3", "%ymm4", "%ymm5", "%ymm6", "%ymm7",
+	    "%ymm8", "%ymm9", "%ymm10", "%ymm11", "%ymm12", "%ymm13", "%ymm14", "%ymm15");
+}
+
 static void test_xsave(void)
 {
     unsigned long cr4;
@@ -45,7 +73,22 @@ static void test_xsave(void)
     report(xsetbv_safe(XCR_XFEATURE_ENABLED_MASK, test_bits) == 0,
            "\t\txsetbv(XCR_XFEATURE_ENABLED_MASK, XSTATE_FP | XSTATE_SSE)");
     report(xgetbv_safe(XCR_XFEATURE_ENABLED_MASK, &xcr0) == 0,
-           "        xgetbv(XCR_XFEATURE_ENABLED_MASK)");
+           "\t\txgetbv(XCR_XFEATURE_ENABLED_MASK)");
+
+    if (supported_xcr0 & XSTATE_YMM) {
+        test_bits = XSTATE_FP | XSTATE_SSE | XSTATE_YMM;
+        report(xsetbv_safe(XCR_XFEATURE_ENABLED_MASK, test_bits) == 0,
+               "\t\txsetbv(XCR_XFEATURE_ENABLED_MASK, XSTATE_FS | XSTATE_SSE | XSTATE_YMM)");
+        xgetbv_safe(XCR_XFEATURE_ENABLED_MASK, &xcr0);
+        if (xcr0 == test_bits) {
+            if (is_fep_available)
+                test_avx_fep();
+            else
+                test_avx();
+            report(memcmp(v32_1, v32_2, 32) == 0,
+                   "vmovdqa emulation");
+        }
+    }
 
     printf("\tIllegal tests\n");
     test_bits = 0;
-- 
2.43.5


^ permalink raw reply related	[flat|nested] 2+ messages in thread

* Re: [PATCH kvm-unit-tests] xsave: add testcase for emulation of AVX instructions
  2025-11-14  0:32 [PATCH kvm-unit-tests] xsave: add testcase for emulation of AVX instructions Paolo Bonzini
@ 2025-11-21 18:05 ` Sean Christopherson
  0 siblings, 0 replies; 2+ messages in thread
From: Sean Christopherson @ 2025-11-21 18:05 UTC (permalink / raw)
  To: Paolo Bonzini; +Cc: kvm, kbusch

On Thu, Nov 13, 2025, Paolo Bonzini wrote:
> Companion patch to the emulator changes in KVM.  Funnily enough,
> no valid combination involving AVX was tried.
> 
> Signed-off-by: Paolo Bonzini <pbonzini@redhat.com>
> ---
>  x86/xsave.c | 45 ++++++++++++++++++++++++++++++++++++++++++++-
>  1 file changed, 44 insertions(+), 1 deletion(-)
> 
> diff --git a/x86/xsave.c b/x86/xsave.c
> index cc8e3a0a..e6d15938 100644
> --- a/x86/xsave.c
> +++ b/x86/xsave.c
> @@ -15,6 +15,34 @@
>  #define XSTATE_SSE      0x2
>  #define XSTATE_YMM      0x4
>  
> +char __attribute__((aligned(32))) v32_1[32] = {
> +    1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
> +    128, 129, 130, 131, 132, 133, 134, 135, 136, 137, 138, 139, 140, 141, 142, 143,
> +};
> +char __attribute__((aligned(32))) v32_2[32] = { 0 };
> +
> +static __attribute__((target("avx"))) void
> +test_avx_fep(void)

Bizarre wrap.

> +{
> +	asm volatile("vzeroall\n"

Why zero all registers?  I can't think of any reason why VZEROALL is needed, and
clobbering everything for no apparent reason is weird/confusing.

> +	    KVM_FEP "vmovdqa v32_1, %%ymm0\n"

These need to use RIP-relative addressing, otherwise EFI builds fail.

> +	    KVM_FEP "vmovdqa %%ymm0, v32_2\n" : : :
> +	    "memory",
> +	    "%ymm0", "%ymm1", "%ymm2", "%ymm3", "%ymm4", "%ymm5", "%ymm6", "%ymm7",
> +	    "%ymm8", "%ymm9", "%ymm10", "%ymm11", "%ymm12", "%ymm13", "%ymm14", "%ymm15");
> +}
> +
> +static __attribute__((target("avx"))) void
> +test_avx(void)
> +{
> +	asm volatile("vzeroall\n"
> +	    "vmovdqa v32_1, %%ymm0\n"
> +	    "vmovdqa %%ymm0, v32_2\n" : : :

We should also test reg=>reg, registers other than ymm0, and that emulated accesses
are correctly propagated to hardware.  E.g. with macro shenanigans, it's not too
hard:

	asm volatile(FEP "vmovdqa v32_1(%%rip), %%" #reg1 "\n"		\
		     FEP "vmovdqa %%" #reg1 ", %%" #reg2 "\n"		\
		     FEP "vmovdqa %%" #reg2 ", v32_2(%%rip)\n"		\
		     "vmovdqa %%" #reg2 ", v32_3(%%rip)\n"		\
		     ::: "memory", #reg1, #reg2);			\

> +	    "memory",
> +	    "%ymm0", "%ymm1", "%ymm2", "%ymm3", "%ymm4", "%ymm5", "%ymm6", "%ymm7",
> +	    "%ymm8", "%ymm9", "%ymm10", "%ymm11", "%ymm12", "%ymm13", "%ymm14", "%ymm15");
> +}
> +
>  static void test_xsave(void)
>  {
>      unsigned long cr4;
> @@ -45,7 +73,22 @@ static void test_xsave(void)
>      report(xsetbv_safe(XCR_XFEATURE_ENABLED_MASK, test_bits) == 0,
>             "\t\txsetbv(XCR_XFEATURE_ENABLED_MASK, XSTATE_FP | XSTATE_SSE)");
>      report(xgetbv_safe(XCR_XFEATURE_ENABLED_MASK, &xcr0) == 0,
> -           "        xgetbv(XCR_XFEATURE_ENABLED_MASK)");
> +           "\t\txgetbv(XCR_XFEATURE_ENABLED_MASK)");

Ugh, this test is ancient and crusty.  We really should have dedicated helpers
for accessing XCR0, it uses spaces instead of tabs, there's unnecessary code
duplication, and a handful of other minor issues.

I'll send a series with a pile of cleanups, and better validation of VMOVDQA
emulation.

^ permalink raw reply	[flat|nested] 2+ messages in thread

end of thread, other threads:[~2025-11-21 18:05 UTC | newest]

Thread overview: 2+ messages (download: mbox.gz follow: Atom feed
-- links below jump to the message on this page --
2025-11-14  0:32 [PATCH kvm-unit-tests] xsave: add testcase for emulation of AVX instructions Paolo Bonzini
2025-11-21 18:05 ` Sean Christopherson

This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox