X-Git-Url: http://demsky.eecs.uci.edu/git/?a=blobdiff_plain;f=lib%2FTarget%2FX86%2FREADME-MMX.txt;h=a6c8616b6d2c6b77a897de92552ef3111b20d25b;hb=d9e3385ced2dc887e2fe8e1c071bd2611e4d3ede;hp=1a6f129c569c24859de5498f4a57ef79dee8e43e;hpb=aadcea33d18f3fc1cfe81d5e9d9d701e8c287ec6;p=oota-llvm.git

diff --git a/lib/Target/X86/README-MMX.txt b/lib/Target/X86/README-MMX.txt
index 1a6f129c569..a6c8616b6d2 100644
--- a/lib/Target/X86/README-MMX.txt
+++ b/lib/Target/X86/README-MMX.txt
@@ -3,3 +3,69 @@
 //===---------------------------------------------------------------------===//
 
 //===---------------------------------------------------------------------===//
+
+This:
+
+#include <mmintrin.h>
+
+__v2si qux(int A) {
+  return (__v2si){ 0, A };
+}
+
+is compiled into:
+
+_qux:
+        subl $28, %esp
+        movl 32(%esp), %eax
+        movd %eax, %mm0
+        movq %mm0, (%esp)
+        movl (%esp), %eax
+        movl %eax, 20(%esp)
+        movq %mm0, 8(%esp)
+        movl 12(%esp), %eax
+        movl %eax, 16(%esp)
+        movq 16(%esp), %mm0
+        addl $28, %esp
+        ret
+
+Yuck!
+
+GCC gives us:
+
+_qux:
+        subl    $12, %esp
+        movl    16(%esp), %eax
+        movl    20(%esp), %edx
+        movl    $0, (%eax)
+        movl    %edx, 4(%eax)
+        addl    $12, %esp
+        ret     $4
+
+//===---------------------------------------------------------------------===//
+
+We generate crappy code for this:
+
+__m64 t() {
+  return _mm_cvtsi32_si64(1);
+}
+
+_t:
+	subl	$12, %esp
+	movl	$1, %eax
+	movd	%eax, %mm0
+	movq	%mm0, (%esp)
+	movl	(%esp), %eax
+	movl	4(%esp), %edx
+	addl	$12, %esp
+	ret
+
+The extra stack traffic is covered in the previous entry. But the other reason
+is we are not smart about materializing constants in MMX registers. With -m64
+
+	movl	$1, %eax
+	movd	%eax, %mm0
+	movd	%mm0, %rax
+	ret
+
+We should be using a constantpool load instead:
+	movq	LC0(%rip), %rax