merge the common darwin settings from the X86/PPC/ARM targets

[oota-llvm.git] / lib / Target / X86 / README-SSE.txt

diff --git a/lib/Target/X86/README-SSE.txt b/lib/Target/X86/README-SSE.txt
index 7c4bf37e34dbc32fd10e07cb2fc66245755fd5b4..71ad51c7984eb5ee07cd9563cb261ac35aff302d 100644 (file)
--- a/lib/Target/X86/README-SSE.txt
+++ b/lib/Target/X86/README-SSE.txt
@@ -17,7 +17,7 @@ other fast SSE modes.
 
 //===---------------------------------------------------------------------===//
 
-Think about doing i64 math in SSE regs.
+Think about doing i64 math in SSE regs on x86-32.
 
 //===---------------------------------------------------------------------===//
 
@@ -841,38 +841,78 @@ _t:
        movaps  %xmm0, 0
 
 //===---------------------------------------------------------------------===//
-rdar://6037315
+rdar://5907648
 
-llvm-gcc-4.2 does the following for uint32_t -> float conversions on i386:
+This function:
 
-       uint32_t x;
-       float y = (float)x;
+float foo(unsigned char x) {
+  return x;
+}
+
+compiles to (x86-32):
+
+define float @foo(i8 zeroext  %x) nounwind  {
+       %tmp12 = uitofp i8 %x to float          ; <float> [#uses=1]
+       ret float %tmp12
+}
+
+compiles to:
+
+_foo:
+       subl    $4, %esp
+       movzbl  8(%esp), %eax
+       cvtsi2ss        %eax, %xmm0
+       movss   %xmm0, (%esp)
+       flds    (%esp)
+       addl    $4, %esp
+       ret
+
+We should be able to use:
+  cvtsi2ss 8($esp), %xmm0
+since we know the stack slot is already zext'd.
+
+//===---------------------------------------------------------------------===//
 
-becomes:
+Consider using movlps instead of movsd to implement (scalar_to_vector (loadf64))
+when code size is critical. movlps is slower than movsd on core2 but it's one
+byte shorter.
 
-movl   %eax,           -8(%ebp)        // write x to the stack
-movl   $0x3ff00000,    -4(%ebp)        // 2^52 + x as a double at -4(%ebp)
-movsd  -8(%ebp),               %xmm0
-subsd  [2^52 double],  %xmm0   // subtract 2^52 -- this is exact
-cvtsd2ss %xmm0,                %xmm0   // convert to single -- rounding happens here
+//===---------------------------------------------------------------------===//
 
-On merom/yonah, this takes a substantial stall.  The following is a much 
-better option:
+We should use a dynamic programming based approach to tell when using FPStack
+operations is cheaper than SSE.  SciMark montecarlo contains code like this
+for example:
 
-movd   %eax,           %xmm0   // load x into low word of xmm0
-movsd  [2^52 double],  %xmm1   // load 2^52 into xmm1
-orpd   %xmm1,          %xmm0   // 2^52 + x in double precision
-subsd  %xmm1,          %xmm0   // x in double precision
-cvtsd2ss %xmm0,                %xmm0   // x rounded to single precision
+double MonteCarlo_num_flops(int Num_samples) {
+    return ((double) Num_samples)* 4.0;
+}
 
-IF we don't already need PIC, then the following is even faster still, at a 
-small cost to code size:
+In fpstack mode, this compiles into:
 
-movl           $0x3ff00000,    %ecx            // conjure high word of 2^52
-movd           %ecx,           %xmm1
-movss  %eax,           %xmm0   // load x into low word of xmm0
-psllq          $32,                    %xmm1   // 2^52
-orpd           %xmm1,          %xmm0   // 2^52 + x in double precision
-subsd          %xmm1,          %xmm0   // x in double precision
-cvtsd2ss       %xmm0,          %xmm0   // x in single precision
+LCPI1_0:                                       
+       .long   1082130432      ## float 4.000000e+00
+_MonteCarlo_num_flops:
+       subl    $4, %esp
+       movl    8(%esp), %eax
+       movl    %eax, (%esp)
+       fildl   (%esp)
+       fmuls   LCPI1_0
+       addl    $4, %esp
+       ret
+        
+in SSE mode, it compiles into significantly slower code:
 
+_MonteCarlo_num_flops:
+       subl    $12, %esp
+       cvtsi2sd        16(%esp), %xmm0
+       mulsd   LCPI1_0, %xmm0
+       movsd   %xmm0, (%esp)
+       fldl    (%esp)
+       addl    $12, %esp
+       ret
+
+There are also other cases in scimark where using fpstack is better, it is
+cheaper to do fld1 than load from a constant pool for example, so
+"load, add 1.0, store" is better done in the fp stack, etc.
+
+//===---------------------------------------------------------------------===//