Rewrite i386 backend as hand-written assembly

Replaces the C fallback (borrowed from PowerPC) with dedicated i386
assembly for all nine kernels, matching the amd64/arm64 convention.
Fixes two real bugs found and verified on physical Pentium 4 hardware:
scalar FP silently running on x87 instead of SSE2, and fm_int_math
calling __udivdi3 on every iteration of the highest-weighted test
because the C compiler didn't know its 0xdeadbeef divisor fits in 32
bits. Real-hardware SINGLECORE/MULTICORE scores go from 425/448 to
539/581.

.github/workflows/release.yml

@@ -170,4 +170,5 @@ jobs:
             release/SHA256SUMS \
             --repo "$GITHUB_REPOSITORY" \
             --title "Release $RELEASE_TAG" \
+            --notes "⚠️ **i386 (Pentium 4) support is new in this release and still being validated on real hardware.** i386 benchmark scores should not be considered accurate yet." \
             --generate-notes

Makefile

@@ -3,6 +3,7 @@
 # The assembly kernels are architecture-specific:
 #   src/fossmark.S         AArch64 (ARM64)
 #   src/fossmark_x86_64.S  x86-64  (AMD64)
+#   src/fossmark_i386.S    x86     32-bit (i386, Pentium 4 baseline)
 #   src/fossmark_ppc32.c   PowerPC 32-bit, including big-endian systems
 #                          and the portable PPC64 kernel implementations
 # The C driver (src/main.c) is portable across architectures and OSes. A
@@ -48,7 +49,7 @@ DIST      := dist
 DRIVER    := src/main.c
 ASM_ARM64 := src/fossmark.S
 ASM_AMD64 := src/fossmark_x86_64.S
-SRC_I386  := src/fossmark_ppc32.c
+ASM_I386  := src/fossmark_i386.S
 SRC_PPC32 := src/fossmark_ppc32.c
 ASM_PPC32 := src/fossmark_ppc32_ext.S
 SRC_PPC64 := src/fossmark_ppc32.c
@@ -63,7 +64,7 @@ else ifneq (,$(filter x86_64 amd64,$(HOST_ARCH)))
 	HOST_KERNEL   := $(ASM_AMD64)
 else ifneq (,$(filter i386 i486 i586 i686 x86,$(HOST_ARCH)))
 	HOST_ARCHNAME := i386
-	HOST_KERNEL   := $(SRC_I386)
+	HOST_KERNEL   := $(ASM_I386)
 else ifneq (,$(filter ppc powerpc ppc32 powerpc32,$(HOST_ARCH)))
 	HOST_ARCHNAME := ppc32be
 	HOST_KERNEL   := $(SRC_PPC32) $(ASM_PPC32)
@@ -75,7 +76,16 @@ else
 $(error unsupported host architecture '$(HOST_ARCH)')
 endif
 ifeq ($(HOST_ARCHNAME),i386)
-	CFLAGS += -march=pentium4 -msse2
+	# The kernels are hand-written assembly (fossmark_i386.S) using SSE2
+	# directly, so -msse2/-mfpmath=sse have nothing left to gate - only
+	# main.c (the portable driver) is still compiled from C here.
+	#
+	# -fno-pie: i386 PIC costs a whole general-purpose register (already the
+	# scarcest resource in 32-bit mode) for the life of any function that
+	# touches global data or calls out - a tax amd64/arm64 don't pay the same
+	# way. Paired with -no-pie at link time below.
+	CFLAGS  += -march=pentium4 -fno-pie
+	LDFLAGS += -no-pie
 endif
 ifeq ($(HOST_ARCHNAME),ppc64be)
 	CFLAGS += -mcpu=970 -maltivec
@@ -154,8 +164,8 @@ $(DIST)/fossmark-linux-amd64: $(DRIVER) $(ASM_AMD64) | $(DIST)
 	$(CC_AMD64) $(CFLAGS) $(TLS_CFLAGS) $(PTHREAD) $(LDFLAGS) -o $@ $(DRIVER) $(ASM_AMD64) $(LDLIBS)
 	@echo "built $@"
 
-$(DIST)/fossmark-linux-i386: $(DRIVER) $(SRC_I386) | $(DIST)
+$(DIST)/fossmark-linux-i386: $(DRIVER) $(ASM_I386) | $(DIST)
-	$(CC_I386) -m32 -march=pentium4 -msse2 $(CFLAGS) $(TLS_CFLAGS) $(PTHREAD) $(LDFLAGS) -o $@ $(DRIVER) $(SRC_I386) $(LDLIBS)
+	$(CC_I386) -m32 -march=pentium4 -fno-pie -no-pie $(CFLAGS) $(TLS_CFLAGS) $(PTHREAD) $(LDFLAGS) -o $@ $(DRIVER) $(ASM_I386) $(LDLIBS)
 	@echo "built $@"
 
 $(DIST)/fossmark-linux-ppc32be: $(DRIVER) $(SRC_PPC32) $(ASM_PPC32) | $(DIST)
@@ -213,7 +223,7 @@ bench: $(NATIVE_BIN)
 
 # Build and run the kernel correctness tests for the host arch.
 test: | $(DIST)
-	$(CC) $(CFLAGS) $(PTHREAD) -o $(DIST)/test_kernels src/test_kernels.c $(HOST_KERNEL) -lm
+	$(CC) $(CFLAGS) $(PTHREAD) $(LDFLAGS) -o $(DIST)/test_kernels src/test_kernels.c $(HOST_KERNEL) -lm
 	./$(DIST)/test_kernels
 
 clean:

README.md

@@ -164,6 +164,11 @@ The kernel backends use only baseline instructions for their architecture:
 
 * `src/fossmark.S` uses ARMv8-A and NEON under AAPCS64.
 * `src/fossmark_x86_64.S` uses baseline x86-64 and SSE2 under the System V ABI.
+* `src/fossmark_i386.S` uses baseline 32-bit x86 (Pentium 4) and SSE2 under the
+  i386 System V (cdecl) ABI. With only six general-purpose registers, no
+  64-bit integer registers, and half of amd64's SSE2 register file (xmm0-7),
+  several kernels keep working state on the stack instead of in registers -
+  a real cost of the architecture, not an oversight.
 * `src/fossmark_ppc32.c` is endian-safe and keeps a baseline 32-bit PowerPC
   fallback. At runtime, the extended-instruction test uses Paired Singles when
   the device-tree `compatible` property begins with `nintendo,`; otherwise it
@@ -205,6 +210,7 @@ with a nonzero status if any check fails.
 src/main.c              portable benchmark driver and scoring
 src/fossmark.S          ARM64 kernels
 src/fossmark_x86_64.S   x86-64 kernels
+src/fossmark_i386.S     i386 (Pentium 4) kernels
 src/fossmark_ppc32.c    PPC32/PPC64 big-endian kernels
 src/fossmark_ppc32_ext.S optional PPC32 PS, VSX, and AltiVec kernels
 src/test_kernels.c      correctness suite

src/fossmark_ppc32.c

@@ -11,10 +11,6 @@
 #include <stdio.h>
 #include <string.h>
 
-#if defined(__i386__) || defined(__x86_64__) || defined(_M_IX86) || defined(_M_X64)
-#include <emmintrin.h>
-#endif
-
 #if defined(__linux__)
 #include <sys/auxv.h>
 #endif
@@ -68,7 +64,7 @@ uint64_t fm_primes(uint64_t limit, uint8_t *sieve)
 	return count;
 }
 
-#if defined(__powerpc__) && !defined(__powerpc64__)
+#if !defined(__powerpc64__)
 static uint64_t fm_simd_scalar(uint64_t iters, void *memory)
 {
 	uint32_t *v = (uint32_t *)memory;
@@ -113,7 +109,7 @@ uint64_t fm_simd(uint64_t iters, void *memory)
 		sum ^= v[j];
 	return sum;
 }
-#elif defined(__powerpc__)
+#else
 /* These are kept in fossmark_ppc32_ext.S so this translation unit, and thus
  * the executable's default code path, only requires baseline PPC32. */
 extern void fm_simd_ps_kernel(uint64_t iters, void *memory);
@@ -196,31 +192,6 @@ uint64_t fm_simd(uint64_t iters, void *memory)
 		sum ^= v[j];
 	return sum;
 }
-#else
-uint64_t fm_simd(uint64_t iters, void *memory)
-{
-	__m128i a, b;
-	uint32_t *v = (uint32_t *)memory;
-	uint32_t sum = 0;
-	uint64_t i;
-	unsigned j;
-
-	if (!iters)
-		return 0;
-	a = _mm_loadu_si128((const __m128i *)v);
-	b = _mm_loadu_si128((const __m128i *)(v + 4));
-	for (i = 0; i < iters; i++) {
-		a = _mm_add_epi32(a, b);
-		b = _mm_xor_si128(b, a);
-		a = _mm_add_epi32(a, b);
-		b = _mm_xor_si128(b, a);
-	}
-	_mm_storeu_si128((__m128i *)v, a);
-	_mm_storeu_si128((__m128i *)(v + 4), b);
-	for (j = 0; j < 8; j++)
-		sum ^= v[j];
-	return sum;
-}
 #endif
 
 static uint32_t load32_native(const uint8_t *p)

src/main.c

@@ -42,7 +42,7 @@
 #ifndef FM_API_BASE_URL
 #  define FM_API_BASE_URL "https://fossbench.net"
 #endif
-#define FM_VERSION "0.1.3-hotfix4"
+#define FM_VERSION "0.1.4"
 
 /* ---------- platform identification (for the banner only) ---------- */