[dpdk-dev] [PATCH v3] eal: Fix up assembly for x86_64 in rte_cpu_get_features

[Neil Horman]

x86_64 doesn't need to save off and restore ebx when issuing cpuid, since x86_64
uses RIP relative addressing.  Doing the save actually clobbers the lower half
of rbx, which could be used and not saved off independently, leading to
undefined behavior.  Fix up the defines so that for x86_64 we just issue the
cpuid instruction, which is safe.  Also, while we're at it, lets clean up the
input and output constraints on the inline asm, so that we don't load registers
that the cpuid instruction isn't sensitive to.

Note that this patch does alter the API, in that specifcations to ebx and edx
are ignored.  I chose to go ahead and do that because there is only a single
caller of this function and neither register is ever written currently.

Signed-off-by: Neil Horman <nhorman at tuxdriver.com>
CC: "H. Peter Anvin" <hpa at zytor.com>
---
 lib/librte_eal/common/eal_common_cpuflags.c | 18 +++++++-----------
 1 file changed, 7 insertions(+), 11 deletions(-)

diff --git a/lib/librte_eal/common/eal_common_cpuflags.c b/lib/librte_eal/common/eal_common_cpuflags.c
index 1ebf78c..0a18d53 100644
--- a/lib/librte_eal/common/eal_common_cpuflags.c
+++ b/lib/librte_eal/common/eal_common_cpuflags.c
@@ -192,7 +192,7 @@ rte_cpu_get_features(struct cpuid_parameters_t params)
 {
 	int eax, ebx, ecx, edx;            /* registers */
 
-#ifndef __PIC__
+#if !defined(__PIC__) || !defined(__i386__)
    asm volatile ("cpuid"
                  /* output */
                  : "=a" (eax),
@@ -201,23 +201,19 @@ rte_cpu_get_features(struct cpuid_parameters_t params)
                    "=d" (edx)
                  /* input */
                  : "a" (params.eax),
-                   "b" (params.ebx),
-                   "c" (params.ecx),
-                   "d" (params.edx));
+                   "c" (params.ecx));
 #else
 	asm volatile ( 
-            "mov %%ebx, %%edi\n"
+            "mov %%ebx, %0\n"
             "cpuid\n"
-            "xchgl %%ebx, %%edi;\n"
-            : "=a" (eax),
-              "=D" (ebx),
+            "xchgl %%ebx, %0\n"
+            : "=r" (ebx),
+	      "=a" (eax),
               "=c" (ecx),
               "=d" (edx)
             /* input */
             : "a" (params.eax),
-              "D" (params.ebx),
-              "c" (params.ecx),
-              "d" (params.edx));
+              "c" (params.ecx));
 #endif
 
 	switch (params.return_register) {
-- 
1.8.3.1