This situation can occur:

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

diff --git a/lib/Target/X86/README.txt b/lib/Target/X86/README.txt
index 845da38f03bd94cf6abb39fa7b9374fa1724fd35..6f6abd478dbf35d128fe5796586bd88f89628df4 100644 (file)
--- a/lib/Target/X86/README.txt
+++ b/lib/Target/X86/README.txt
@@ -1612,3 +1612,100 @@ The coalescer could coalesce "edx" with "eax" to avoid the movl in LBB1_2
 if it commuted the addl in LBB1_1.
 
 //===---------------------------------------------------------------------===//
+
+See rdar://4653682.
+
+From flops:
+
+LBB1_15:        # bb310
+        cvtss2sd        LCPI1_0, %xmm1
+        addsd   %xmm1, %xmm0
+        movsd   176(%esp), %xmm2
+        mulsd   %xmm0, %xmm2
+        movapd  %xmm2, %xmm3
+        mulsd   %xmm3, %xmm3
+        movapd  %xmm3, %xmm4
+        mulsd   LCPI1_23, %xmm4
+        addsd   LCPI1_24, %xmm4
+        mulsd   %xmm3, %xmm4
+        addsd   LCPI1_25, %xmm4
+        mulsd   %xmm3, %xmm4
+        addsd   LCPI1_26, %xmm4
+        mulsd   %xmm3, %xmm4
+        addsd   LCPI1_27, %xmm4
+        mulsd   %xmm3, %xmm4
+        addsd   LCPI1_28, %xmm4
+        mulsd   %xmm3, %xmm4
+        addsd   %xmm1, %xmm4
+        mulsd   %xmm2, %xmm4
+        movsd   152(%esp), %xmm1
+        addsd   %xmm4, %xmm1
+        movsd   %xmm1, 152(%esp)
+        incl    %eax
+        cmpl    %eax, %esi
+        jge     LBB1_15 # bb310
+LBB1_16:        # bb358.loopexit
+        movsd   152(%esp), %xmm0
+        addsd   %xmm0, %xmm0
+        addsd   LCPI1_22, %xmm0
+        movsd   %xmm0, 152(%esp)
+
+Rather than spilling the result of the last addsd in the loop, we should have
+insert a copy to split the interval (one for the duration of the loop, one
+extending to the fall through). The register pressure in the loop isn't high
+enough to warrant the spill.
+
+Also check why xmm7 is not used at all in the function.
+
+//===---------------------------------------------------------------------===//
+
+Legalize loses track of the fact that bools are always zero extended when in
+memory.  This causes us to compile abort_gzip (from 164.gzip) from:
+
+target datalayout = "e-p:32:32:32-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:32:64-f32:32:32-f64:32:64-v64:64:64-v128:128:128-a0:0:64-f80:128:128"
+target triple = "i386-apple-darwin8"
+@in_exit.4870.b = internal global i1 false             ; <i1*> [#uses=2]
+define fastcc void @abort_gzip() noreturn nounwind  {
+entry:
+       %tmp.b.i = load i1* @in_exit.4870.b             ; <i1> [#uses=1]
+       br i1 %tmp.b.i, label %bb.i, label %bb4.i
+bb.i:          ; preds = %entry
+       tail call void @exit( i32 1 ) noreturn nounwind 
+       unreachable
+bb4.i:         ; preds = %entry
+       store i1 true, i1* @in_exit.4870.b
+       tail call void @exit( i32 1 ) noreturn nounwind 
+       unreachable
+}
+declare void @exit(i32) noreturn nounwind 
+
+into:
+
+_abort_gzip:
+       subl    $12, %esp
+       movb    _in_exit.4870.b, %al
+       notb    %al
+       testb   $1, %al
+       jne     LBB1_2  ## bb4.i
+LBB1_1:        ## bb.i
+  ...
+
+//===---------------------------------------------------------------------===//
+
+We compile:
+
+int test(int x, int y) {
+  return x-y-1;
+}
+
+into (-m64):
+
+_test:
+       decl    %edi
+       movl    %edi, %eax
+       subl    %esi, %eax
+       ret
+
+it would be better to codegen as: x+~y  (notl+addl)
+
+//===---------------------------------------------------------------------===//