Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[De Lara Guarch, Pablo]

Hi Neil,



>> What was the value of RTE_COMPILE_TIME_CPUFLAGS when you built your library?  If you added defines for flag support that are not tested for yet in the cpu_feature_table, that would be the error you would see I think



>RTE_COMPILE_TIME_CPUFLAGS= >RTE_CPUFLAG_SSE,RTE_CPUFLAG_SSE2,RTE_CPUFLAG_SSE3,RTE_CPUFLAG_SSSE3,RTE_CPUFLAG_SSE4_1,RTE_CPUFLAG_SSE4_2,RTE_CPUFLAG_AES,RTE_CPUFLAG_PCLMULQDQ,RTE_CPUFLAG_AVX



>For gcc_version >=4.6, that value is the same, of course (and all those flags are supported). So it has to be something else...



Silly me, I just saw what was hapenning. There are two missing brackets in the loop:



for (i = 0; i < sizeof(compile_time_flags)/sizeof(compile_time_flags[0]); i++) {

ret = rte_cpu_get_flag_enabled(compile_time_flags[i]);



if (ret < 0) {

fprintf(stderr,

  "ERROR: CPU feature flag lookup failed with error %d\n",

  ret);

exit(1);

}

if (!ret) {

fprintf(stderr,

  "ERROR: This system does not support \"%s\".\n"

  "Please check that RTE_MACHINE is set correctly.\n",

  cpu_feature_table[compile_time_flags[i]].name);

exit(1);

}

}



So, basically, "i" was size of the array +1 after the loop and only the return value was being checked for the last item. Not sure why higher versions of gcc do not catch this.



Thanks,



Pablo de Lara

DPDK SW Engineer



--------------------------------------------------------------

Intel Shannon Limited

Registered in Ireland

Registered Office: Collinstown Industrial Park, Leixlip, County Kildare Registered Number: 308263 Business address: Dromore House, East Park, Shannon, Co. Clare

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[Thomas Monjalon]

Hi Pablo,

2014-04-16 08:58, De Lara Guarch, Pablo:
> There are two missing brackets in the loop:

Good catch!
Do you plan to send a patch to fix this?

Thank you
-- 
Thomas

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[Neil Horman]

On Wed, Apr 16, 2014 at 08:58:31AM +0000, De Lara Guarch, Pablo wrote:
> Hi Neil,
> 
> 
> 
> >> What was the value of RTE_COMPILE_TIME_CPUFLAGS when you built your library?  If you added defines for flag support that are not tested for yet in the cpu_feature_table, that would be the error you would see I think
> 
> 
> 
> >RTE_COMPILE_TIME_CPUFLAGS= >RTE_CPUFLAG_SSE,RTE_CPUFLAG_SSE2,RTE_CPUFLAG_SSE3,RTE_CPUFLAG_SSSE3,RTE_CPUFLAG_SSE4_1,RTE_CPUFLAG_SSE4_2,RTE_CPUFLAG_AES,RTE_CPUFLAG_PCLMULQDQ,RTE_CPUFLAG_AVX
> 
> 
> 
> >For gcc_version >=4.6, that value is the same, of course (and all those flags are supported). So it has to be something else...
> 
> 
> 
> Silly me, I just saw what was hapenning. There are two missing brackets in the loop:
> 
> 
> 
> for (i = 0; i < sizeof(compile_time_flags)/sizeof(compile_time_flags[0]); i++) {
> 
> ret = rte_cpu_get_flag_enabled(compile_time_flags[i]);
> 
> 
> 
> if (ret < 0) {
> 
> fprintf(stderr,
> 
>   "ERROR: CPU feature flag lookup failed with error %d\n",
> 
>   ret);
> 
> exit(1);
> 
> }
> 
> if (!ret) {
> 
> fprintf(stderr,
> 
>   "ERROR: This system does not support \"%s\".\n"
> 
>   "Please check that RTE_MACHINE is set correctly.\n",
> 
>   cpu_feature_table[compile_time_flags[i]].name);
> 
> exit(1);
> 
> }
> 
> }
> 
> 
> 
> So, basically, "i" was size of the array +1 after the loop and only the return value was being checked for the last item. Not sure why higher versions of gcc do not catch this.
> 
Oh wow, yes, the if conditionals definately should be checked for each iteration
of the for loop. Good eye.

Still though, seems like a bug in gcc to check the state of the loop index on
exit, when its never used to index the array at that value.  Seems a bit like
this bug:
http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45180

Thomas I'll send a patch to fix this directly
Neil

> 
> 
> Thanks,
> 
> 
> 
> Pablo de Lara
> 
> DPDK SW Engineer
> 
> 
> 
> --------------------------------------------------------------
> 
> Intel Shannon Limited
> 
> Registered in Ireland
> 
> Registered Office: Collinstown Industrial Park, Leixlip, County Kildare Registered Number: 308263 Business address: Dromore House, East Park, Shannon, Co. Clare
> 
> 
> 
> 
> 
> 
> 

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[De Lara Guarch, Pablo]

Hi Neil,

>Oh wow, yes, the if conditionals definately should be checked for each iteration of the for loop. Good eye.

>Still though, seems like a bug in gcc to check the state of the loop index on exit, when its never used to index the array at that value.  Seems a bit like this bug:
>http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45180

>Thomas I'll send a patch to fix this directly Neil

Actually, that index may be used in cause last "ret" value is 0:

if (!ret) {
                        fprintf(stderr,
                                "ERROR: This system does not support \"%s\".\n"
                                "Please check that RTE_MACHINE is set correctly.\n",
                                cpu_feature_table[compile_time_flags[i]].name);
                        exit(1);
                }

Thanks,

Pablo de Lara
DPDK SW Engineer

--------------------------------------------------------------
Intel Shannon Limited
Registered in Ireland
Registered Office: Collinstown Industrial Park, Leixlip, County Kildare Registered Number: 308263 Business address: Dromore House, East Park, Shannon, Co. Clare



 

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[Neil Horman]

On Wed, Apr 16, 2014 at 11:09:46AM +0000, De Lara Guarch, Pablo wrote:
> Hi Neil,
> 
> >Oh wow, yes, the if conditionals definately should be checked for each iteration of the for loop. Good eye.
> 
> >Still though, seems like a bug in gcc to check the state of the loop index on exit, when its never used to index the array at that value.  Seems a bit like this bug:
> >http://gcc.gnu.org/bugzilla/show_bug.cgi?id=45180
> 
> >Thomas I'll send a patch to fix this directly Neil
> 
> Actually, that index may be used in cause last "ret" value is 0:
> 
> if (!ret) {
>                         fprintf(stderr,
>                                 "ERROR: This system does not support \"%s\".\n"
>                                 "Please check that RTE_MACHINE is set correctly.\n",
>                                 cpu_feature_table[compile_time_flags[i]].name);
>                         exit(1);
>                 }
> 
Ah, well there you have it.  Thank you.  Either way though, those two error
checks need to be encased in brackets.  It appears I missed it in my previous
patch as the for loop only had one if, and so didn't need additional brackets,
and I extended the check, requiring brackets that I never added.  apologies.

I've sent a patch for review
Neil

> Thanks,
> 
> Pablo de Lara
> DPDK SW Engineer
> 
> --------------------------------------------------------------
> Intel Shannon Limited
> Registered in Ireland
> Registered Office: Collinstown Industrial Park, Leixlip, County Kildare Registered Number: 308263 Business address: Dromore House, East Park, Shannon, Co. Clare
> 
> 
> 
>  
> 

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[De Lara Guarch, Pablo]

Hi Neil,

> What was the value of RTE_COMPILE_TIME_CPUFLAGS when you built your library?  If you added defines for flag support that are not tested for yet in the cpu_feature_table, that would be the error you would see I think

RTE_COMPILE_TIME_CPUFLAGS= RTE_CPUFLAG_SSE,RTE_CPUFLAG_SSE2,RTE_CPUFLAG_SSE3,RTE_CPUFLAG_SSSE3,RTE_CPUFLAG_SSE4_1,RTE_CPUFLAG_SSE4_2,RTE_CPUFLAG_AES,RTE_CPUFLAG_PCLMULQDQ,RTE_CPUFLAG_AVX

For gcc_version >=4.6, that value is the same, of course (and all those flags are supported). So it has to be something else... 

Thanks,

Pablo de Lara
DPDK SW Engineer

--------------------------------------------------------------
Intel Shannon Limited
Registered in Ireland
Registered Office: Collinstown Industrial Park, Leixlip, County Kildare Registered Number: 308263 Business address: Dromore House, East Park, Shannon, Co. Clare



 

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[Neil Horman]

On Tue, Apr 15, 2014 at 02:09:34PM +0000, De Lara Guarch, Pablo wrote:
> Hi,
> 
> I have checked that on gcc-4.5, I get a compilation error due to the following patch:
> 
> > > Neil Horman reported that on x86-64 the upper half of %rbx would get 
> > > clobbered when the code was compiled PIC or PIE, because the 
> > > i386-specific code to preserve %ebx was incorrectly compiled.
> > > 
> > > However, the code is really way more complex than it needs to be.  
> > > For one thing, the CPUID instruction only needs %eax (leaf) and %ecx
> > > (subleaf) as parameters, and since we are testing for bits, we might 
> > > as well list the bits explicitly.  Furthermore, we can use an array 
> > > rather than doing a switch statement inside a structure.
> > > 
> > > Reported-by: Neil Horman <nhorman@tuxdriver.com>
> > > Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
> > > Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
> 
> The compilation error is:
> 
> /root/dpdk/lib/librte_eal/common/eal_common_cpuflags.c: In function 'rte_cpu_check_supported':
> /root/dpdk/lib/librte_eal/common/eal_common_cpuflags.c:271:48: error: array subscript is above array bounds
> make[7]: *** [eal_common_cpuflags.o] Error 1
> 
> This error is not present in newer gcc versions (4.6 and above).
> 
> Thanks,
> 
> Pablo de Lara
> DPDK SW Engineer
> 

What was the value of RTE_COMPILE_TIME_CPUFLAGS when you built your library?  If
you added defines for flag support that are not tested for yet in the
cpu_feature_table, that would be the error you would see I think

Regards
Neil


> --------------------------------------------------------------
> Intel Shannon Limited
> Registered in Ireland
> Registered Office: Collinstown Industrial Park, Leixlip, County Kildare Registered Number: 308263 Business address: Dromore House, East Park, Shannon, Co. Clare
> 
> 
> 

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[De Lara Guarch, Pablo]

Hi,

I have checked that on gcc-4.5, I get a compilation error due to the following patch:

> > Neil Horman reported that on x86-64 the upper half of %rbx would get 
> > clobbered when the code was compiled PIC or PIE, because the 
> > i386-specific code to preserve %ebx was incorrectly compiled.
> > 
> > However, the code is really way more complex than it needs to be.  
> > For one thing, the CPUID instruction only needs %eax (leaf) and %ecx
> > (subleaf) as parameters, and since we are testing for bits, we might 
> > as well list the bits explicitly.  Furthermore, we can use an array 
> > rather than doing a switch statement inside a structure.
> > 
> > Reported-by: Neil Horman <nhorman@tuxdriver.com>
> > Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
> > Signed-off-by: Neil Horman <nhorman@tuxdriver.com>

The compilation error is:

/root/dpdk/lib/librte_eal/common/eal_common_cpuflags.c: In function 'rte_cpu_check_supported':
/root/dpdk/lib/librte_eal/common/eal_common_cpuflags.c:271:48: error: array subscript is above array bounds
make[7]: *** [eal_common_cpuflags.o] Error 1

This error is not present in newer gcc versions (4.6 and above).

Thanks,

Pablo de Lara
DPDK SW Engineer

--------------------------------------------------------------
Intel Shannon Limited
Registered in Ireland
Registered Office: Collinstown Industrial Park, Leixlip, County Kildare Registered Number: 308263 Business address: Dromore House, East Park, Shannon, Co. Clare

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[Thomas Monjalon]

2014-03-25 13:51, H. Peter Anvin:
> On 03/25/2014 12:52 PM, Neil Horman wrote:
> > Neil Horman reported that on x86-64 the upper half of %rbx would get
> > clobbered when the code was compiled PIC or PIE, because the
> > i386-specific code to preserve %ebx was incorrectly compiled.
> > 
> > However, the code is really way more complex than it needs to be.  For
> > one thing, the CPUID instruction only needs %eax (leaf) and %ecx
> > (subleaf) as parameters, and since we are testing for bits, we might
> > as well list the bits explicitly.  Furthermore, we can use an array
> > rather than doing a switch statement inside a structure.
> > 
> > Reported-by: Neil Horman <nhorman@tuxdriver.com>
> > Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
> > Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
> 
> Looks good to me.
> 
> Reviewed-by: H. Peter Anvin <hpa@linux.intel.com>

Applied for version 1.6.0r2.

Thanks for this difficult fix.
-- 
Thomas

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[Thomas Monjalon]

2014-04-02 07:16, Neil Horman:
> On Tue, Mar 25, 2014 at 01:51:04PM -0700, H. Peter Anvin wrote:
> > On 03/25/2014 12:52 PM, Neil Horman wrote:
> > > Neil Horman reported that on x86-64 the upper half of %rbx would get
> > > clobbered when the code was compiled PIC or PIE, because the
> > > i386-specific code to preserve %ebx was incorrectly compiled.
> > > 
> > > However, the code is really way more complex than it needs to be.  For
> > > one thing, the CPUID instruction only needs %eax (leaf) and %ecx
> > > (subleaf) as parameters, and since we are testing for bits, we might
> > > as well list the bits explicitly.  Furthermore, we can use an array
> > > rather than doing a switch statement inside a structure.
> > > 
> > > Reported-by: Neil Horman <nhorman@tuxdriver.com>
> > > Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
> > > Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
> > 
> > Looks good to me.
> > 
> > Reviewed-by: H. Peter Anvin <hpa@linux.intel.com>
> > 
> > 	-hpa
> 
> Bump, did this get lost somewhere?  Its been over a week and I don't see it
> in the tree

No, it's not lost. But this patch is not trivial and there were 5 versions 
with acknowledgements in the middle. So I think it was not a bad idea to wait 
few days in order to be sure this version is OK :)

-- 
Thomas

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[Neil Horman]

On Tue, Mar 25, 2014 at 01:51:04PM -0700, H. Peter Anvin wrote:
> On 03/25/2014 12:52 PM, Neil Horman wrote:
> > Neil Horman reported that on x86-64 the upper half of %rbx would get
> > clobbered when the code was compiled PIC or PIE, because the
> > i386-specific code to preserve %ebx was incorrectly compiled.
> > 
> > However, the code is really way more complex than it needs to be.  For
> > one thing, the CPUID instruction only needs %eax (leaf) and %ecx
> > (subleaf) as parameters, and since we are testing for bits, we might
> > as well list the bits explicitly.  Furthermore, we can use an array
> > rather than doing a switch statement inside a structure.
> > 
> > Reported-by: Neil Horman <nhorman@tuxdriver.com>
> > Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
> > Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
> > 
> 
> Looks good to me.
> 
> Reviewed-by: H. Peter Anvin <hpa@linux.intel.com>
> 
> 	-hpa
> 
> 
> 
Bump, did this get lost somewhere?  Its been over a week and I don't see it in
the tree

Regards
Neil

Re: [dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[H. Peter Anvin]

On 03/25/2014 12:52 PM, Neil Horman wrote:
> Neil Horman reported that on x86-64 the upper half of %rbx would get
> clobbered when the code was compiled PIC or PIE, because the
> i386-specific code to preserve %ebx was incorrectly compiled.
> 
> However, the code is really way more complex than it needs to be.  For
> one thing, the CPUID instruction only needs %eax (leaf) and %ecx
> (subleaf) as parameters, and since we are testing for bits, we might
> as well list the bits explicitly.  Furthermore, we can use an array
> rather than doing a switch statement inside a structure.
> 
> Reported-by: Neil Horman <nhorman@tuxdriver.com>
> Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
> Signed-off-by: Neil Horman <nhorman@tuxdriver.com>
> 

Looks good to me.

Reviewed-by: H. Peter Anvin <hpa@linux.intel.com>

	-hpa

[dpdk-dev] [PATCH v5] eal_common_cpuflags: Fix %rbx corruption, and simplify the code

[Neil Horman]

Neil Horman reported that on x86-64 the upper half of %rbx would get
clobbered when the code was compiled PIC or PIE, because the
i386-specific code to preserve %ebx was incorrectly compiled.

However, the code is really way more complex than it needs to be.  For
one thing, the CPUID instruction only needs %eax (leaf) and %ecx
(subleaf) as parameters, and since we are testing for bits, we might
as well list the bits explicitly.  Furthermore, we can use an array
rather than doing a switch statement inside a structure.

Reported-by: Neil Horman <nhorman@tuxdriver.com>
Signed-off-by: H. Peter Anvin <hpa@linux.intel.com>
Signed-off-by: Neil Horman <nhorman@tuxdriver.com>

---
Change notes:
v2) Corrected build errors
Fixed cpuid_register_t reference passing
Fixed typedef name typo

v3)
* Modified feature_entry struct to drop the name field, as its unused
* Modified cpu_feature_table to use C99 initalizers
* Updated FEAT_DEF macro to include all feature_entry fields
* Modified cpuid_reg enum to start at 1 rather than zero
* Added CPUID_REG macro to drop enum value by 1 during access
* Added check on feat->reg use to detect missing entries
* Fixed a bug in rte_cpu_check_supported in which negative errors are ignored

v4)
* Fixed sanity checks to not offset feat->reg and just check !feat->reg
* Added a check for the sanity of the leaf node

v5)
* Fixed max leaf check to just return not supported rather than error
---
 lib/librte_eal/common/eal_common_cpuflags.c | 281 ++++++++++++++--------------
 1 file changed, 136 insertions(+), 145 deletions(-)

diff --git a/lib/librte_eal/common/eal_common_cpuflags.c b/lib/librte_eal/common/eal_common_cpuflags.c
index 1ebf78c..f9c1840 100644
--- a/lib/librte_eal/common/eal_common_cpuflags.c
+++ b/lib/librte_eal/common/eal_common_cpuflags.c
@@ -59,16 +59,7 @@ enum cpu_register_t {
 	REG_EDX,
 };
 
-/**
- * Parameters for CPUID instruction
- */
-struct cpuid_parameters_t {
-	uint32_t eax;
-	uint32_t ebx;
-	uint32_t ecx;
-	uint32_t edx;
-	enum cpu_register_t return_register;
-};
+typedef uint32_t cpuid_registers_t[4];
 
 #define CPU_FLAG_NAME_MAX_LEN 64
 
@@ -76,109 +67,111 @@ struct cpuid_parameters_t {
  * Struct to hold a processor feature entry
  */
 struct feature_entry {
-	enum rte_cpu_flag_t feature;            /**< feature name */
+	uint32_t leaf;				/**< cpuid leaf */
+	uint32_t subleaf;			/**< cpuid subleaf */
+	uint32_t reg;				/**< cpuid register */
+	uint32_t bit;				/**< cpuid register bit */
 	char name[CPU_FLAG_NAME_MAX_LEN];       /**< String for printing */
-	struct cpuid_parameters_t params;       /**< cpuid parameters */
-	uint32_t feature_mask;                  /**< bitmask for feature */
 };
 
-#define FEAT_DEF(f) RTE_CPUFLAG_##f, #f
+#define FEAT_DEF(name, leaf, subleaf, reg, bit) \
+	[RTE_CPUFLAG_##name] = {leaf, subleaf, reg, bit, #name },
 
 /**
  * An array that holds feature entries
  */
 static const struct feature_entry cpu_feature_table[] = {
-	{FEAT_DEF(SSE3),              {0x1, 0, 0, 0, REG_ECX}, 0x00000001},
-	{FEAT_DEF(PCLMULQDQ),         {0x1, 0, 0, 0, REG_ECX}, 0x00000002},
-	{FEAT_DEF(DTES64),            {0x1, 0, 0, 0, REG_ECX}, 0x00000004},
-	{FEAT_DEF(MONITOR),           {0x1, 0, 0, 0, REG_ECX}, 0x00000008},
-	{FEAT_DEF(DS_CPL),            {0x1, 0, 0, 0, REG_ECX}, 0x00000010},
-	{FEAT_DEF(VMX),               {0x1, 0, 0, 0, REG_ECX}, 0x00000020},
-	{FEAT_DEF(SMX),               {0x1, 0, 0, 0, REG_ECX}, 0x00000040},
-	{FEAT_DEF(EIST),              {0x1, 0, 0, 0, REG_ECX}, 0x00000080},
-	{FEAT_DEF(TM2),               {0x1, 0, 0, 0, REG_ECX}, 0x00000100},
-	{FEAT_DEF(SSSE3),             {0x1, 0, 0, 0, REG_ECX}, 0x00000200},
-	{FEAT_DEF(CNXT_ID),           {0x1, 0, 0, 0, REG_ECX}, 0x00000400},
-	{FEAT_DEF(FMA),               {0x1, 0, 0, 0, REG_ECX}, 0x00001000},
-	{FEAT_DEF(CMPXCHG16B),        {0x1, 0, 0, 0, REG_ECX}, 0x00002000},
-	{FEAT_DEF(XTPR),              {0x1, 0, 0, 0, REG_ECX}, 0x00004000},
-	{FEAT_DEF(PDCM),              {0x1, 0, 0, 0, REG_ECX}, 0x00008000},
-	{FEAT_DEF(PCID),              {0x1, 0, 0, 0, REG_ECX}, 0x00020000},
-	{FEAT_DEF(DCA),               {0x1, 0, 0, 0, REG_ECX}, 0x00040000},
-	{FEAT_DEF(SSE4_1),            {0x1, 0, 0, 0, REG_ECX}, 0x00080000},
-	{FEAT_DEF(SSE4_2),            {0x1, 0, 0, 0, REG_ECX}, 0x00100000},
-	{FEAT_DEF(X2APIC),            {0x1, 0, 0, 0, REG_ECX}, 0x00200000},
-	{FEAT_DEF(MOVBE),             {0x1, 0, 0, 0, REG_ECX}, 0x00400000},
-	{FEAT_DEF(POPCNT),            {0x1, 0, 0, 0, REG_ECX}, 0x00800000},
-	{FEAT_DEF(TSC_DEADLINE),      {0x1, 0, 0, 0, REG_ECX}, 0x01000000},
-	{FEAT_DEF(AES),               {0x1, 0, 0, 0, REG_ECX}, 0x02000000},
-	{FEAT_DEF(XSAVE),             {0x1, 0, 0, 0, REG_ECX}, 0x04000000},
-	{FEAT_DEF(OSXSAVE),           {0x1, 0, 0, 0, REG_ECX}, 0x08000000},
-	{FEAT_DEF(AVX),               {0x1, 0, 0, 0, REG_ECX}, 0x10000000},
-	{FEAT_DEF(F16C),              {0x1, 0, 0, 0, REG_ECX}, 0x20000000},
-	{FEAT_DEF(RDRAND),            {0x1, 0, 0, 0, REG_ECX}, 0x40000000},
+	FEAT_DEF(SSE3, 0x00000001, 0, REG_ECX,  0)
+	FEAT_DEF(PCLMULQDQ, 0x00000001, 0, REG_ECX,  1)
+	FEAT_DEF(DTES64, 0x00000001, 0, REG_ECX,  2)
+	FEAT_DEF(MONITOR, 0x00000001, 0, REG_ECX,  3)
+	FEAT_DEF(DS_CPL, 0x00000001, 0, REG_ECX,  4)
+	FEAT_DEF(VMX, 0x00000001, 0, REG_ECX,  5)
+	FEAT_DEF(SMX, 0x00000001, 0, REG_ECX,  6)
+	FEAT_DEF(EIST, 0x00000001, 0, REG_ECX,  7)
+	FEAT_DEF(TM2, 0x00000001, 0, REG_ECX,  8)
+	FEAT_DEF(SSSE3, 0x00000001, 0, REG_ECX,  9)
+	FEAT_DEF(CNXT_ID, 0x00000001, 0, REG_ECX, 10)
+	FEAT_DEF(FMA, 0x00000001, 0, REG_ECX, 12)
+	FEAT_DEF(CMPXCHG16B, 0x00000001, 0, REG_ECX, 13)
+	FEAT_DEF(XTPR, 0x00000001, 0, REG_ECX, 14)
+	FEAT_DEF(PDCM, 0x00000001, 0, REG_ECX, 15)
+	FEAT_DEF(PCID, 0x00000001, 0, REG_ECX, 17)
+	FEAT_DEF(DCA, 0x00000001, 0, REG_ECX, 18)
+	FEAT_DEF(SSE4_1, 0x00000001, 0, REG_ECX, 19)
+	FEAT_DEF(SSE4_2, 0x00000001, 0, REG_ECX, 20)
+	FEAT_DEF(X2APIC, 0x00000001, 0, REG_ECX, 21)
+	FEAT_DEF(MOVBE, 0x00000001, 0, REG_ECX, 22)
+	FEAT_DEF(POPCNT, 0x00000001, 0, REG_ECX, 23)
+	FEAT_DEF(TSC_DEADLINE, 0x00000001, 0, REG_ECX, 24)
+	FEAT_DEF(AES, 0x00000001, 0, REG_ECX, 25)
+	FEAT_DEF(XSAVE, 0x00000001, 0, REG_ECX, 26)
+	FEAT_DEF(OSXSAVE, 0x00000001, 0, REG_ECX, 27)
+	FEAT_DEF(AVX, 0x00000001, 0, REG_ECX, 28)
+	FEAT_DEF(F16C, 0x00000001, 0, REG_ECX, 29)
+	FEAT_DEF(RDRAND, 0x00000001, 0, REG_ECX, 30)
 
-	{FEAT_DEF(FPU),               {0x1, 0, 0, 0, REG_EDX}, 0x00000001},
-	{FEAT_DEF(VME),               {0x1, 0, 0, 0, REG_EDX}, 0x00000002},
-	{FEAT_DEF(DE),                {0x1, 0, 0, 0, REG_EDX}, 0x00000004},
-	{FEAT_DEF(PSE),               {0x1, 0, 0, 0, REG_EDX}, 0x00000008},
-	{FEAT_DEF(TSC),               {0x1, 0, 0, 0, REG_EDX}, 0x00000010},
-	{FEAT_DEF(MSR),               {0x1, 0, 0, 0, REG_EDX}, 0x00000020},
-	{FEAT_DEF(PAE),               {0x1, 0, 0, 0, REG_EDX}, 0x00000040},
-	{FEAT_DEF(MCE),               {0x1, 0, 0, 0, REG_EDX}, 0x00000080},
-	{FEAT_DEF(CX8),               {0x1, 0, 0, 0, REG_EDX}, 0x00000100},
-	{FEAT_DEF(APIC),              {0x1, 0, 0, 0, REG_EDX}, 0x00000200},
-	{FEAT_DEF(SEP),               {0x1, 0, 0, 0, REG_EDX}, 0x00000800},
-	{FEAT_DEF(MTRR),              {0x1, 0, 0, 0, REG_EDX}, 0x00001000},
-	{FEAT_DEF(PGE),               {0x1, 0, 0, 0, REG_EDX}, 0x00002000},
-	{FEAT_DEF(MCA),               {0x1, 0, 0, 0, REG_EDX}, 0x00004000},
-	{FEAT_DEF(CMOV),              {0x1, 0, 0, 0, REG_EDX}, 0x00008000},
-	{FEAT_DEF(PAT),               {0x1, 0, 0, 0, REG_EDX}, 0x00010000},
-	{FEAT_DEF(PSE36),             {0x1, 0, 0, 0, REG_EDX}, 0x00020000},
-	{FEAT_DEF(PSN),               {0x1, 0, 0, 0, REG_EDX}, 0x00040000},
-	{FEAT_DEF(CLFSH),             {0x1, 0, 0, 0, REG_EDX}, 0x00080000},
-	{FEAT_DEF(DS),                {0x1, 0, 0, 0, REG_EDX}, 0x00200000},
-	{FEAT_DEF(ACPI),              {0x1, 0, 0, 0, REG_EDX}, 0x00400000},
-	{FEAT_DEF(MMX),               {0x1, 0, 0, 0, REG_EDX}, 0x00800000},
-	{FEAT_DEF(FXSR),              {0x1, 0, 0, 0, REG_EDX}, 0x01000000},
-	{FEAT_DEF(SSE),               {0x1, 0, 0, 0, REG_EDX}, 0x02000000},
-	{FEAT_DEF(SSE2),              {0x1, 0, 0, 0, REG_EDX}, 0x04000000},
-	{FEAT_DEF(SS),                {0x1, 0, 0, 0, REG_EDX}, 0x08000000},
-	{FEAT_DEF(HTT),               {0x1, 0, 0, 0, REG_EDX}, 0x10000000},
-	{FEAT_DEF(TM),                {0x1, 0, 0, 0, REG_EDX}, 0x20000000},
-	{FEAT_DEF(PBE),               {0x1, 0, 0, 0, REG_EDX}, 0x80000000},
+	FEAT_DEF(FPU, 0x00000001, 0, REG_EDX,  0)
+	FEAT_DEF(VME, 0x00000001, 0, REG_EDX,  1)
+	FEAT_DEF(DE, 0x00000001, 0, REG_EDX,  2)
+	FEAT_DEF(PSE, 0x00000001, 0, REG_EDX,  3)
+	FEAT_DEF(TSC, 0x00000001, 0, REG_EDX,  4)
+	FEAT_DEF(MSR, 0x00000001, 0, REG_EDX,  5)
+	FEAT_DEF(PAE, 0x00000001, 0, REG_EDX,  6)
+	FEAT_DEF(MCE, 0x00000001, 0, REG_EDX,  7)
+	FEAT_DEF(CX8, 0x00000001, 0, REG_EDX,  8)
+	FEAT_DEF(APIC, 0x00000001, 0, REG_EDX,  9)
+	FEAT_DEF(SEP, 0x00000001, 0, REG_EDX, 11)
+	FEAT_DEF(MTRR, 0x00000001, 0, REG_EDX, 12)
+	FEAT_DEF(PGE, 0x00000001, 0, REG_EDX, 13)
+	FEAT_DEF(MCA, 0x00000001, 0, REG_EDX, 14)
+	FEAT_DEF(CMOV, 0x00000001, 0, REG_EDX, 15)
+	FEAT_DEF(PAT, 0x00000001, 0, REG_EDX, 16)
+	FEAT_DEF(PSE36, 0x00000001, 0, REG_EDX, 17)
+	FEAT_DEF(PSN, 0x00000001, 0, REG_EDX, 18)
+	FEAT_DEF(CLFSH, 0x00000001, 0, REG_EDX, 19)
+	FEAT_DEF(DS, 0x00000001, 0, REG_EDX, 21)
+	FEAT_DEF(ACPI, 0x00000001, 0, REG_EDX, 22)
+	FEAT_DEF(MMX, 0x00000001, 0, REG_EDX, 23)
+	FEAT_DEF(FXSR, 0x00000001, 0, REG_EDX, 24)
+	FEAT_DEF(SSE, 0x00000001, 0, REG_EDX, 25)
+	FEAT_DEF(SSE2, 0x00000001, 0, REG_EDX, 26)
+	FEAT_DEF(SS, 0x00000001, 0, REG_EDX, 27)
+	FEAT_DEF(HTT, 0x00000001, 0, REG_EDX, 28)
+	FEAT_DEF(TM, 0x00000001, 0, REG_EDX, 29)
+	FEAT_DEF(PBE, 0x00000001, 0, REG_EDX, 31)
 
-	{FEAT_DEF(DIGTEMP),           {0x6, 0, 0, 0, REG_EAX}, 0x00000001},
-	{FEAT_DEF(TRBOBST),           {0x6, 0, 0, 0, REG_EAX}, 0x00000002},
-	{FEAT_DEF(ARAT),              {0x6, 0, 0, 0, REG_EAX}, 0x00000004},
-	{FEAT_DEF(PLN),               {0x6, 0, 0, 0, REG_EAX}, 0x00000010},
-	{FEAT_DEF(ECMD),              {0x6, 0, 0, 0, REG_EAX}, 0x00000020},
-	{FEAT_DEF(PTM),               {0x6, 0, 0, 0, REG_EAX}, 0x00000040},
+	FEAT_DEF(DIGTEMP, 0x00000006, 0, REG_EAX,  0)
+	FEAT_DEF(TRBOBST, 0x00000006, 0, REG_EAX,  1)
+	FEAT_DEF(ARAT, 0x00000006, 0, REG_EAX,  2)
+	FEAT_DEF(PLN, 0x00000006, 0, REG_EAX,  4)
+	FEAT_DEF(ECMD, 0x00000006, 0, REG_EAX,  5)
+	FEAT_DEF(PTM, 0x00000006, 0, REG_EAX,  6)
 
-	{FEAT_DEF(MPERF_APERF_MSR),   {0x6, 0, 0, 0, REG_ECX}, 0x00000001},
-	{FEAT_DEF(ACNT2),             {0x6, 0, 0, 0, REG_ECX}, 0x00000002},
-	{FEAT_DEF(ENERGY_EFF),        {0x6, 0, 0, 0, REG_ECX}, 0x00000008},
+	FEAT_DEF(MPERF_APERF_MSR, 0x00000006, 0, REG_ECX,  0)
+	FEAT_DEF(ACNT2, 0x00000006, 0, REG_ECX,  1)
+	FEAT_DEF(ENERGY_EFF, 0x00000006, 0, REG_ECX,  3)
 
-	{FEAT_DEF(FSGSBASE),          {0x7, 0, 0, 0, REG_EBX}, 0x00000001},
-	{FEAT_DEF(BMI1),              {0x7, 0, 0, 0, REG_EBX}, 0x00000004},
-	{FEAT_DEF(HLE),               {0x7, 0, 0, 0, REG_EBX}, 0x00000010},
-	{FEAT_DEF(AVX2),              {0x7, 0, 0, 0, REG_EBX}, 0x00000020},
-	{FEAT_DEF(SMEP),              {0x7, 0, 0, 0, REG_EBX}, 0x00000040},
-	{FEAT_DEF(BMI2),              {0x7, 0, 0, 0, REG_EBX}, 0x00000080},
-	{FEAT_DEF(ERMS),              {0x7, 0, 0, 0, REG_EBX}, 0x00000100},
-	{FEAT_DEF(INVPCID),           {0x7, 0, 0, 0, REG_EBX}, 0x00000400},
-	{FEAT_DEF(RTM),               {0x7, 0, 0, 0, REG_EBX}, 0x00000800},
+	FEAT_DEF(FSGSBASE, 0x00000007, 0, REG_EBX,  0)
+	FEAT_DEF(BMI1, 0x00000007, 0, REG_EBX,  2)
+	FEAT_DEF(HLE, 0x00000007, 0, REG_EBX,  4)
+	FEAT_DEF(AVX2, 0x00000007, 0, REG_EBX,  5)
+	FEAT_DEF(SMEP, 0x00000007, 0, REG_EBX,  6)
+	FEAT_DEF(BMI2, 0x00000007, 0, REG_EBX,  7)
+	FEAT_DEF(ERMS, 0x00000007, 0, REG_EBX,  8)
+	FEAT_DEF(INVPCID, 0x00000007, 0, REG_EBX, 10)
+	FEAT_DEF(RTM, 0x00000007, 0, REG_EBX, 11)
 
-	{FEAT_DEF(LAHF_SAHF),  {0x80000001, 0, 0, 0, REG_ECX}, 0x00000001},
-	{FEAT_DEF(LZCNT),      {0x80000001, 0, 0, 0, REG_ECX}, 0x00000010},
+	FEAT_DEF(LAHF_SAHF, 0x80000001, 0, REG_ECX,  0)
+	FEAT_DEF(LZCNT, 0x80000001, 0, REG_ECX,  4)
 
-	{FEAT_DEF(SYSCALL),    {0x80000001, 0, 0, 0, REG_EDX}, 0x00000800},
-	{FEAT_DEF(XD),         {0x80000001, 0, 0, 0, REG_EDX}, 0x00100000},
-	{FEAT_DEF(1GB_PG),     {0x80000001, 0, 0, 0, REG_EDX}, 0x04000000},
-	{FEAT_DEF(RDTSCP),     {0x80000001, 0, 0, 0, REG_EDX}, 0x08000000},
-	{FEAT_DEF(EM64T),      {0x80000001, 0, 0, 0, REG_EDX}, 0x20000000},
+	FEAT_DEF(SYSCALL, 0x80000001, 0, REG_EDX, 11)
+	FEAT_DEF(XD, 0x80000001, 0, REG_EDX, 20)
+	FEAT_DEF(1GB_PG, 0x80000001, 0, REG_EDX, 26)
+	FEAT_DEF(RDTSCP, 0x80000001, 0, REG_EDX, 27)
+	FEAT_DEF(EM64T, 0x80000001, 0, REG_EDX, 29)
 
-	{FEAT_DEF(INVTSC),     {0x80000007, 0, 0, 0, REG_EDX}, 0x00000100},
+	FEAT_DEF(INVTSC, 0x80000007, 0, REG_EDX,  8)
 };
 
 /*
@@ -187,51 +180,27 @@ static const struct feature_entry cpu_feature_table[] = {
  * This function, when compiled with GCC, will generate architecture-neutral
  * code, as per GCC manual.
  */
-static inline int
-rte_cpu_get_features(struct cpuid_parameters_t params)
+static inline void
+rte_cpu_get_features(uint32_t leaf, uint32_t subleaf, cpuid_registers_t out)
 {
-	int eax, ebx, ecx, edx;            /* registers */
-
-#ifndef __PIC__
-   asm volatile ("cpuid"
-                 /* output */
-                 : "=a" (eax),
-                   "=b" (ebx),
-                   "=c" (ecx),
-                   "=d" (edx)
-                 /* input */
-                 : "a" (params.eax),
-                   "b" (params.ebx),
-                   "c" (params.ecx),
-                   "d" (params.edx));
+#if defined(__i386__) && defined(__PIC__)
+    /* %ebx is a forbidden register if we compile with -fPIC or -fPIE */
+    asm volatile("movl %%ebx,%0 ; cpuid ; xchgl %%ebx,%0"
+		 : "=r" (out[REG_EBX]),
+		   "=a" (out[REG_EAX]),
+		   "=c" (out[REG_ECX]),
+		   "=d" (out[REG_EDX])
+		 : "a" (leaf), "c" (subleaf));
 #else
-	asm volatile ( 
-            "mov %%ebx, %%edi\n"
-            "cpuid\n"
-            "xchgl %%ebx, %%edi;\n"
-            : "=a" (eax),
-              "=D" (ebx),
-              "=c" (ecx),
-              "=d" (edx)
-            /* input */
-            : "a" (params.eax),
-              "D" (params.ebx),
-              "c" (params.ecx),
-              "d" (params.edx));
-#endif
 
-	switch (params.return_register) {
-	case REG_EAX:
-		return eax;
-	case REG_EBX:
-		return ebx;
-	case REG_ECX:
-		return ecx;
-	case REG_EDX:
-		return edx;
-	default:
-		return 0;
-	}
+    asm volatile("cpuid"
+		 : "=a" (out[REG_EAX]),
+		   "=b" (out[REG_EBX]),
+		   "=c" (out[REG_ECX]),
+		   "=d" (out[REG_EDX])
+		 : "a" (leaf), "c" (subleaf));
+
+#endif
 }
 
 /*
@@ -240,17 +209,30 @@ rte_cpu_get_features(struct cpuid_parameters_t params)
 int
 rte_cpu_get_flag_enabled(enum rte_cpu_flag_t feature)
 {
-	int value;
+	const struct feature_entry *feat;
+	cpuid_registers_t regs;
+
 
 	if (feature >= RTE_CPUFLAG_NUMFLAGS)
 		/* Flag does not match anything in the feature tables */
 		return -ENOENT;
 
-	/* get value of the register containing the desired feature */
-	value = rte_cpu_get_features(cpu_feature_table[feature].params);
+	feat = &cpu_feature_table[feature];
+
+	if (!feat->leaf)
+		/* This entry in the table wasn't filled out! */
+		return -EFAULT;
+
+	rte_cpu_get_features(feat->leaf & 0xffff0000, 0, regs);
+	if (((regs[REG_EAX] ^ feat->leaf) & 0xffff0000) ||
+	      regs[REG_EAX] < feat->leaf)
+		return 0;
+
+	/* get the cpuid leaf containing the desired feature */
+	rte_cpu_get_features(feat->leaf, feat->subleaf, regs);
 
 	/* check if the feature is enabled */
-	return (cpu_feature_table[feature].feature_mask & value) > 0;
+	return (regs[feat->reg] >> feat->bit) & 1;
 }
 
 /**
@@ -271,9 +253,18 @@ rte_cpu_check_supported(void)
 			RTE_COMPILE_TIME_CPUFLAGS
 	};
 	unsigned i;
+	int ret;
 
 	for (i = 0; i < sizeof(compile_time_flags)/sizeof(compile_time_flags[0]); i++)
-		if (rte_cpu_get_flag_enabled(compile_time_flags[i]) < 1) {
+		ret = rte_cpu_get_flag_enabled(compile_time_flags[i]);
+
+		if (ret < 0) {
+			fprintf(stderr,
+				"ERROR: CPU feature flag lookup failed with error %d\n",
+				ret);
+			exit(1);
+		}
+		if (!ret) {
 			fprintf(stderr,
 			        "ERROR: This system does not support \"%s\".\n"
 			        "Please check that RTE_MACHINE is set correctly.\n",
-- 
1.8.3.1