diff --git a/source/lib/sysdep/ia32.asm b/source/lib/sysdep/ia32.asm
index 47109e5a0d..1a14a5ed44 100644
--- a/source/lib/sysdep/ia32.asm
+++ b/source/lib/sysdep/ia32.asm
@@ -232,14 +232,17 @@ __SECT__
 ; extern "C" bool __cdecl ia32_cpuid(u32 func, u32* regs)
 global _ia32_cpuid
 _ia32_cpuid:
+	; note: must preserve before .one_time_init because it does cpuid
+	push		ebx
+	push		edi
+
+.retry:
 	; if unknown, detect; if not available, fail.
 	xor		eax, eax				; return val on failure
 	cmp		[cpuid_available], eax
 	jl		.one_time_init
 	je		.ret
 
-	push		ebx
-	push		edi
 	mov		ecx, [esp+8+4+0]			; func
 	mov		edi, [esp+8+4+4]			; -> regs
 
@@ -291,7 +294,7 @@ _ia32_cpuid:
 	cpuid
 	mov		[max_ext_func], eax
 
-	jmp		_ia32_cpuid					; now try again
+	jmp		.retry
 
 
 
diff --git a/source/lib/sysdep/ia32.cpp b/source/lib/sysdep/ia32.cpp
index 6f5ef783cf..d4d3a18b28 100755
--- a/source/lib/sysdep/ia32.cpp
+++ b/source/lib/sysdep/ia32.cpp
@@ -312,11 +312,13 @@ static void get_cpu_type()
 	const uint family = bits(regs[EAX], 8, 11);
 
 	// get brand string (if available)
+	// note: ia32_cpuid writes 4 u32s directly to cpu_type -
+	// be very careful with pointer arithmetic!
 	u32* cpu_type_u32 = (u32*)cpu_type;
 	bool have_brand_string = false;
 	if(ia32_cpuid(0x80000002, cpu_type_u32+0 ) &&
-	   ia32_cpuid(0x80000003, cpu_type_u32+16) &&
-	   ia32_cpuid(0x80000004, cpu_type_u32+32))
+	   ia32_cpuid(0x80000003, cpu_type_u32+4) &&
+	   ia32_cpuid(0x80000004, cpu_type_u32+8))
 		have_brand_string = true;