Use ADX instructions where applicable in SIKE R2 assembly (aws#1712)

* Use ADX instructions where applicable in SIKE R2 assembly

* Fix sike_r2 Makefile

CMakeLists.txt

@@ -107,6 +107,8 @@ else()
 endif()
 
 # The PQ ASM try_compile has to come after we turn on pthread
+set(PQ_ASM_COMPILES_ADX false)
+
 if(S2N_NO_PQ_ASM)
     message(STATUS "S2N_NO_PQ_ASM flag was detected - forcing usage of generic C code for PQ crypto")
 else()
@@ -121,6 +123,19 @@ else()
         message(STATUS "PQ ASM try_compile succeeded - using optimized x86_64 assembly for PQ crypto")
         file(GLOB PQ_X86_64_ASM "pq-crypto/sike_r2/fp_x64_asm.S")
         list(APPEND PQ_SRC ${PQ_X86_64_ASM})
+
+        message(STATUS "Attempting to try_compile PQ ASM with ADX support")
+        try_compile(PQ_ASM_COMPILES_ADX ${CMAKE_BINARY_DIR}
+                SOURCES
+                "${CMAKE_CURRENT_LIST_DIR}/tests/unit/s2n_pq_asm_noop_test.c"
+                "${CMAKE_CURRENT_LIST_DIR}/pq-crypto/sike_r2/fp_x64_asm.S"
+                COMPILE_DEFINITIONS "-D_ADX_")
+        if(PQ_ASM_COMPILES_ADX)
+            message(STATUS "PQ ASM try_compile with ADX support succeeded - using ASM code with ADX instructions")
+            # The -D_ADX_ compile flag is added to the project below
+        else()
+            message(STATUS "PQ ASM try_compile with ADX support failed - using ASM code without ADX instructions")
+        endif()
     else()
         message(STATUS "PQ ASM try_compile failed - using generic C code for PQ crypto")
         set(S2N_NO_PQ_ASM ON)
@@ -173,6 +188,10 @@ if(S2N_NO_PQ)
     target_compile_options(${PROJECT_NAME} PUBLIC -DS2N_NO_PQ)
 endif()
 
+if(PQ_ASM_COMPILES_ADX)
+    target_compile_options(${PROJECT_NAME} PUBLIC -D_ADX_)
+endif()
+
 target_compile_options(${PROJECT_NAME} PUBLIC -fPIC)
 
 target_compile_definitions(${PROJECT_NAME} PRIVATE -D_POSIX_C_SOURCE=200809L)

pq-crypto/Makefile

@@ -53,9 +53,9 @@ run-lcov: lcov
 
 .PHONY : clean
 clean: decruft
-	${MAKE} -C bike_r1 decruft
+	$(MAKE) -C bike_r1 decruft
 	$(MAKE) -C bike_r2 decruft
-	${MAKE} -C sike_r1 decruft
-	${MAKE} -C sike_r2 decruft
+	$(MAKE) -C sike_r1 decruft
+	$(MAKE) -C sike_r2 decruft
 
 include ../s2n.mk

pq-crypto/sike_r2/Makefile

@@ -40,6 +40,11 @@ ifndef S2N_NO_PQ_ASM
 	ifeq ($(TRY_COMPILE_PQ_ASM), 0)
 		ASRC=fp_x64_asm.S
 		OBJS+=$(ASRC:.S=.o)
+
+		TRY_COMPILE_PQ_ASM_ADX := $(shell $(CC) -D_ADX_ -c -o ./fp_x64_asm.o ./fp_x64_asm.S > /dev/null 2>&1; echo $$?)
+		ifeq ($(TRY_COMPILE_PQ_ASM_ADX), 0)
+			CFLAGS += -D_ADX_
+		endif
 	else
 		CFLAGS += -DS2N_NO_PQ_ASM
 	endif

pq-crypto/sike_r2/fp_x64_asm.S

@@ -6,31 +6,29 @@
 
 .intel_syntax noprefix
 
-#define _MULX_
+/* Requires bmi2 instruction set for mulx. adx instructions are optional, but preferred. */
 
 // Registers that are used for parameter passing:
 #define reg_p1  rdi
 #define reg_p2  rsi
 #define reg_p3  rdx
 
 // Define addition instructions
-#ifdef _MULX_
 #ifdef _ADX_
 
 #define ADD1    adox
 #define ADC1    adox
 #define ADD2    adcx
 #define ADC2    adcx
 
-#else
+#else // _ADX_
 
 #define ADD1    add
 #define ADC1    adc
 #define ADD2    add
 #define ADC2    adc
 
-#endif    
-#endif
+#endif // _ADX_
 
 // The constants below (asm_p434, asm_p434p1, and asm_p434x2) are duplicated from
 // P434.c, and correspond to the arrays p434, p434p1, and p434x2. The values are
@@ -147,7 +145,6 @@ fpadd434_asm:
   pop    r12
   ret
 
-
 //***********************************************************************
 //  Field subtraction
 //  Operation: c [reg_p3] = a [reg_p1] - b [reg_p2]
@@ -209,9 +206,6 @@ fpsub434_asm:
   pop    r13
   pop    r12
   ret
-
-
-#ifdef _MULX_
 
 ///////////////////////////////////////////////////////////////// MACRO
 // Schoolbook integer multiplication, a full row at a time
@@ -335,7 +329,7 @@ fpsub434_asm:
     mov    56\C, \T8         // C7_final
 .endm 
 
-#else
+#else // _ADX_
 
 .macro MUL192_SCHOOL M0, M1, C, T0, T1, T2, T3, T4, T5, T6
     mov    rdx, \M0
@@ -445,8 +439,7 @@ fpsub434_asm:
     adc    \T8, 0
     mov    56\C, \T8         // C7_final
 .endm
-#endif
-
+#endif // _ADX_
 
 //*****************************************************************************
 //  434-bit multiplication using Karatsuba (one level), schoolbook (one level)
@@ -595,26 +588,6 @@ mul434_asm:
     pop    r12
     ret
 
-#else
-
-//***********************************************************************                       
-//  Integer multiplication
-//  Based on Karatsuba method
-//  Operation: c [reg_p3] = a [reg_p1] * b [reg_p2]
-//  NOTE: a=c or b=c are not allowed
-//***********************************************************************
-.global mul434_asm
-mul434_asm:
-
-  ret
-
-# error "CONFIGURATION NOT SUPPORTED. TRY USE_MULX=TRUE"
-
-#endif
-
-
-#ifdef _MULX_
-
 ///////////////////////////////////////////////////////////////// MACRO
 // Schoolbook integer multiplication
 // Inputs:  memory pointers M0 and M1
@@ -664,7 +637,7 @@ mul434_asm:
     ADC1   \T5, rax 
 .endm
 
-#else
+#else // _ADX_
 
 .macro MUL128x256_SCHOOL M0, M1, T0, T1, T2, T3, T4, T5, T6 
     mov    rdx, \M0
@@ -692,9 +665,8 @@ mul434_asm:
     adc    \T4, rdx  
     adc    \T5, 0 
 .endm
-#endif
+#endif // _ADX_
 
-
 //**************************************************************************************
 //  Montgomery reduction
 //  Based on method described in Faz-Hernandez et al. https://eprint.iacr.org/2017/1015
@@ -800,24 +772,6 @@ rdc434_asm:
     pop    r12
     ret
 
-#else
-
-//***********************************************************************
-//  Montgomery reduction
-//  Based on comba method
-//  Operation: c [reg_p2] = a [reg_p1]                                                      
-//  NOTE: a=c is not allowed
-//*********************************************************************** 
-.global rdc434_asm
-rdc434_asm:
-
-  ret
-
-# error "CONFIGURATION NOT SUPPORTED. TRY USE_MULX=TRUE"
-
-#endif
-
-
 //***********************************************************************
 //  434-bit multiprecision addition
 //  Operation: c [reg_p3] = a [reg_p1] + b [reg_p2]
@@ -848,7 +802,6 @@ mp_add434_asm:
   mov    [reg_p3+48], r10
   ret
 
-
 //***********************************************************************
 //  2x434-bit multiprecision subtraction/addition
 //  Operation: c [x2] = a [x0] - b [x1]. If c < 0, add p434*2^448
@@ -933,7 +886,6 @@ mp_subadd434x2_asm:
   pop    r12
   ret
 
-
 //***********************************************************************
 //  Double 2x434-bit multiprecision subtraction
 //  Operation: c [reg_p3] = c [reg_p3] - a [reg_p1] - b [reg_p2]