diff --git a/Makefile b/Makefile
index aa38ca8..d616d6f 100755
--- a/Makefile
+++ b/Makefile
@@ -31,6 +31,13 @@ endif
 
 # Detect the architecture
 ifeq ($(YQ2_OSTYPE), Windows)
+ifdef MINGW_CHOST
+ifeq ($(MINGW_CHOST), x86_64-w64-mingw32)
+YQ2_ARCH ?= x86_64
+else # i686-w64-mingw32
+YQ2_ARCH ?= i386
+endif
+else # windows, but MINGW_CHOST not defined
 ifdef PROCESSOR_ARCHITEW6432
 # 64 bit Windows
 YQ2_ARCH ?= $(PROCESSOR_ARCHITEW6432)
@@ -38,6 +45,7 @@ else
 # 32 bit Windows
 YQ2_ARCH ?= $(PROCESSOR_ARCHITECTURE)
 endif
+endif # windows but MINGW_CHOST not defined
 else
 # Normalize some abiguous YQ2_ARCH strings
 YQ2_ARCH ?= $(shell uname -m | sed -e 's/i.86/i386/' -e 's/amd64/x86_64/' -e 's/^arm.*/arm/')
@@ -120,13 +128,12 @@ endif
 
 # ----------
 
-# If we're building with gcc for i386 let's define -ffloat-store.
-# This helps the old and crappy x87 FPU to produce correct values.
-# Would be nice if Clang had something comparable.
+# Using the default x87 float math on 32bit x86 causes rounding trouble
+# -ffloat-store could work around that, but the better solution is to
+# just enforce SSE - every x86 CPU since Pentium3 supports that
+# and this should even improve the performance on old CPUs
 ifeq ($(YQ2_ARCH), i386)
-ifeq ($(COMPILER), gcc)
-override CFLAGS += -ffloat-store
-endif
+override CFLAGS += -msse -mfpmath=sse
 endif
 
 # Force SSE math on x86_64. All sane compilers should do this