benchmark: rework output formatting and per-algorithm reporting

linuxkm/Kbuild

@@ -230,7 +230,25 @@ endif
 $(obj)/$(LIBWOLFSSL_NAME).mod.o: ccflags-y := $(PIE_SUPPORT_FLAGS)
 $(obj)/wolfcrypt/test/test.o: ccflags-y += -DNO_MAIN_DRIVER -DWOLFSSL_NO_OPTIONS_H
 $(obj)/wolfcrypt/src/aes.o: ccflags-y := $(WOLFSSL_CFLAGS) $(WOLFSSL_CFLAGS_YES_VECTOR_INSNS) $(PIE_FLAGS) $(PIE_SUPPORT_FLAGS)
-$(obj)/wolfcrypt/benchmark/benchmark.o: ccflags-y := $(WOLFSSL_CFLAGS) $(CFLAGS_FPU_ENABLE) $(CFLAGS_SIMD_ENABLE) $(CFLAGS_AUTO_VECTORIZE_DISABLE) $(PIE_SUPPORT_FLAGS) -DNO_MAIN_FUNCTION -DWOLFSSL_NO_OPTIONS_H
+# benchmark.c used to compute its statistics in double, so this object alone
+# was built with the FPU and SSE enabled while every other object in the module
+# gets $(WOLFSSL_CFLAGS_NO_VECTOR_INSNS).  That permission outlived its reason:
+# under WC_BENCH_INT_ONLY, which is automatic for WOLFSSL_LINUXKM, the
+# statistics are 64-bit integer work and no floating point remains.
+#
+# Leaving it enabled was not free.  gcc used the permission for things that
+# have nothing to do with arithmetic -- zeroing a stack buffer with
+# "pxor %xmm0,%xmm0" and a run of movaps -- and emitted them in the function
+# PROLOGUE, before the SAVE_VECTOR_REGISTERS() that bench_stats_start() opens.
+# 21 reachable bench_* functions wrote %xmm0 with no kernel_fpu_begin() in
+# effect, and nothing outside brackets them: init_module() calls
+# wolfcrypt_benchmark_main() directly, and neither it nor benchmark_test()
+# opens a bracket.
+#
+# Disabling vector instructions here fixes that at the source rather than by
+# widening a bracket.  The bracket in bench_stats_start() still has to exist:
+# it is there for the benchmarked CRYPTO, which does use SIMD.
+$(obj)/wolfcrypt/benchmark/benchmark.o: ccflags-y := $(WOLFSSL_CFLAGS) $(WOLFSSL_CFLAGS_NO_VECTOR_INSNS) $(CFLAGS_AUTO_VECTORIZE_DISABLE) $(PIE_SUPPORT_FLAGS) -DNO_MAIN_FUNCTION -DWOLFSSL_NO_OPTIONS_H
 
 asflags-y := $(WOLFSSL_ASFLAGS) $(ASFLAGS_FPUSIMD_DISABLE)