[llvm-commits] [llvm] r123135 - /llvm/trunk/lib/Target/README.txt

[Chandler Carruth]

Author: chandlerc
Date: Sun Jan  9 16:36:18 2011
New Revision: 123135

URL: http://llvm.org/viewvc/llvm-project?rev=123135&view=rev
Log:
Add a note about the inability to model FP -> int conversions which
perform rounding other than truncation in the IR. Common C code for this
turns into really an LLVM intrinsic call that blocks a lot of further
optimizations.

Modified:
    llvm/trunk/lib/Target/README.txt

Modified: llvm/trunk/lib/Target/README.txt
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/README.txt?rev=123135&r1=123134&r2=123135&view=diff
==============================================================================
--- llvm/trunk/lib/Target/README.txt (original)
+++ llvm/trunk/lib/Target/README.txt Sun Jan  9 16:36:18 2011
@@ -2273,3 +2273,58 @@
 the largest integer value as < inf.
 
 //===---------------------------------------------------------------------===//
+
+clang -O3 currently compiles this code:
+
+#include <emmintrin.h>
+int f(double x) { return _mm_cvtsd_si32(_mm_set_sd(x)); }
+int g(double x) { return _mm_cvttsd_si32(_mm_set_sd(x)); }
+
+into
+
+define i32 @_Z1fd(double %x) nounwind readnone {
+entry:
+  %vecinit.i = insertelement <2 x double> undef, double %x, i32 0
+  %vecinit1.i = insertelement <2 x double> %vecinit.i, double 0.000000e+00, i32 1
+  %0 = tail call i32 @llvm.x86.sse2.cvtsd2si(<2 x double> %vecinit1.i) nounwind
+  ret i32 %0
+}
+
+define i32 @_Z1gd(double %x) nounwind readnone {
+entry:
+  %conv.i = fptosi double %x to i32
+  ret i32 %conv.i
+}
+
+This difference carries over to the assmebly produced, resulting in:
+
+_Z1fd:                                  # @_Z1fd
+# BB#0:                                 # %entry
+        pushq   %rbp
+        movq    %rsp, %rbp
+        xorps   %xmm1, %xmm1
+        movsd   %xmm0, %xmm1
+        cvtsd2sil       %xmm1, %eax
+        popq    %rbp
+        ret
+
+_Z1gd:                                  # @_Z1gd
+# BB#0:                                 # %entry
+        pushq   %rbp
+        movq    %rsp, %rbp
+        cvttsd2si       %xmm0, %eax
+        popq    %rbp
+        ret
+
+The problem is that we can't see through the intrinsic call used for cvtsd2si,
+and fold away the unnecessary manipulation of the function parameter. When
+these functions are inlined, it forms a barrier preventing many further
+optimizations. LLVM IR doesn't have a good way to model the logic of
+'cvtsd2si', its only FP -> int conversion path forces truncation. We should add
+a rounding flag onto fptosi so that it can represent this type of rounding
+naturally in the IR rather than using intrinsics. We might need to use a
+'system_rounding_mode' flag to encode that the semantics of the rounding mode
+can be changed by the program, but ideally we could just say that isn't
+supported, and hard code the rounding.
+
+//===---------------------------------------------------------------------===//