Revert r203230, "CodeGenPrep: sink extends of illegal types into use block."

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

diff --git a/lib/Target/X86/README.txt b/lib/Target/X86/README.txt
index 52d3c01076de2f69d27112c8e1e8c6caf5897bcc..a30f8c7d9852fdaafc06906509412ad2fa7329b9 100644 (file)
--- a/lib/Target/X86/README.txt
+++ b/lib/Target/X86/README.txt
@@ -1444,6 +1444,54 @@ it would be nice to produce "into" someday.
 
 //===---------------------------------------------------------------------===//
 
+This code:
+
+void vec_mpys1(int y[], const int x[], int scaler) {
+int i;
+for (i = 0; i < 150; i++)
+ y[i] += (((long long)scaler * (long long)x[i]) >> 31);
+}
+
+Compiles to this loop with GCC 3.x:
+
+.L5:
+       movl    %ebx, %eax
+       imull   (%edi,%ecx,4)
+       shrdl   $31, %edx, %eax
+       addl    %eax, (%esi,%ecx,4)
+       incl    %ecx
+       cmpl    $149, %ecx
+       jle     .L5
+
+llvm-gcc compiles it to the much uglier:
+
+LBB1_1:        ## bb1
+       movl    24(%esp), %eax
+       movl    (%eax,%edi,4), %ebx
+       movl    %ebx, %ebp
+       imull   %esi, %ebp
+       movl    %ebx, %eax
+       mull    %ecx
+       addl    %ebp, %edx
+       sarl    $31, %ebx
+       imull   %ecx, %ebx
+       addl    %edx, %ebx
+       shldl   $1, %eax, %ebx
+       movl    20(%esp), %eax
+       addl    %ebx, (%eax,%edi,4)
+       incl    %edi
+       cmpl    $150, %edi
+       jne     LBB1_1  ## bb1
+
+The issue is that we hoist the cast of "scaler" to long long outside of the
+loop, the value comes into the loop as two values, and
+RegsForValue::getCopyFromRegs doesn't know how to put an AssertSext on the
+constructed BUILD_PAIR which represents the cast value.
+
+This can be handled by making CodeGenPrepare sink the cast.
+
+//===---------------------------------------------------------------------===//
+
 Test instructions can be eliminated by using EFLAGS values from arithmetic
 instructions. This is currently not done for mul, and, or, xor, neg, shl,
 sra, srl, shld, shrd, atomic ops, and others. It is also currently not done