From fb00e2715b109e06de04b276e69015a69d1dfe16 Mon Sep 17 00:00:00 2001
From: Chandler Carruth <chandlerc@gmail.com>
Date: Sun, 9 Jan 2011 22:36:18 +0000
Subject: [PATCH] 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.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@123135 91177308-0d34-0410-b5e6-96231b3b80d8
---
 lib/Target/README.txt | 55 +++++++++++++++++++++++++++++++++++++++++++++++++++
 1 file changed, 55 insertions(+)

diff --git a/lib/Target/README.txt b/lib/Target/README.txt
index a9afffd95ae..bea240c0921 100644
--- a/lib/Target/README.txt
+++ b/lib/Target/README.txt
@@ -2273,3 +2273,58 @@ this whenever the floating point type has enough exponent bits to represent
 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.
+
+//===---------------------------------------------------------------------===//
-- 
2.11.0