[llvm-commits] [llvm] r47431 - /llvm/trunk/lib/Target/X86/README.txt
Chris Lattner
sabre at nondot.org
Wed Feb 20 22:51:30 PST 2008
Author: lattner
Date: Thu Feb 21 00:51:29 2008
New Revision: 47431
URL: http://llvm.org/viewvc/llvm-project?rev=47431&view=rev
Log:
Dan implemented one multiply issue. Replace it with another. :)
Modified:
llvm/trunk/lib/Target/X86/README.txt
Modified: llvm/trunk/lib/Target/X86/README.txt
URL: http://llvm.org/viewvc/llvm-project/llvm/trunk/lib/Target/X86/README.txt?rev=47431&r1=47430&r2=47431&view=diff
==============================================================================
--- llvm/trunk/lib/Target/X86/README.txt (original)
+++ llvm/trunk/lib/Target/X86/README.txt Thu Feb 21 00:51:29 2008
@@ -1392,23 +1392,44 @@
//===---------------------------------------------------------------------===//
-We're missing an obvious fold of a load into imul:
+We're codegen'ing multiply of long longs inefficiently:
-int test(long a, long b) { return a * b; }
+unsigned long long LLM(unsigned long long arg1, unsigned long long arg2) {
+ return arg1 * arg2;
+}
-LLVM produces:
-_test:
- movl 4(%esp), %ecx
- movl 8(%esp), %eax
- imull %ecx, %eax
- ret
+We compile to (fomit-frame-pointer):
-vs:
-_test:
- movl 8(%esp), %eax
- imull 4(%esp), %eax
+_LLM:
+ pushl %esi
+ movl 8(%esp), %ecx
+ movl 16(%esp), %esi
+ movl %esi, %eax
+ mull %ecx
+ imull 12(%esp), %esi
+ addl %edx, %esi
+ imull 20(%esp), %ecx
+ movl %esi, %edx
+ addl %ecx, %edx
+ popl %esi
+ ret
+
+This looks like a scheduling deficiency and lack of remat of the load from
+the argument area. ICC apparently produces:
+
+ movl 8(%esp), %ecx
+ imull 12(%esp), %ecx
+ movl 16(%esp), %eax
+ imull 4(%esp), %eax
+ addl %eax, %ecx
+ movl 4(%esp), %eax
+ mull 12(%esp)
+ addl %ecx, %edx
ret
+Note that it remat'd loads from 4(esp) and 12(esp). See this GCC PR:
+http://gcc.gnu.org/bugzilla/show_bug.cgi?id=17236
+
//===---------------------------------------------------------------------===//
We can fold a store into "zeroing a reg". Instead of:
More information about the llvm-commits
mailing list