[X86] Fast-ISel was incorrectly always killing the source of a truncate.

A trunc from i32 to i1 on x86_64 generates an instruction such as

%vreg19<def> = COPY %vreg9:sub_8bit<kill>; GR8:%vreg19 GR32:%vreg9

However, the copy here should only have the kill flag on the 32-bit path, not the 64-bit one.
Otherwise, we are killing the source of the truncate which could be used later in the program.

git-svn-id: https://llvm.org/svn/llvm-project/llvm/trunk@236890 91177308-0d34-0410-b5e6-96231b3b80d8

diff --git a/lib/Target/X86/X86FastISel.cpp b/lib/Target/X86/X86FastISel.cpp
index 9d817164b05367d569c3b5b95889a29794a39037..3bf4fb824fd523729639b6ab7381e5e71bb22371 100644 (file)
--- a/lib/Target/X86/X86FastISel.cpp
+++ b/lib/Target/X86/X86FastISel.cpp
@@ -2185,6 +2185,7 @@ bool X86FastISel::X86SelectTrunc(const Instruction *I) {
     return true;
   }
 
+  bool KillInputReg = false;
   if (!Subtarget->is64Bit()) {
     // If we're on x86-32; we can't extract an i8 from a general register.
     // First issue a copy to GR16_ABCD or GR32_ABCD.
@@ -2194,11 +2195,12 @@ bool X86FastISel::X86SelectTrunc(const Instruction *I) {
     BuildMI(*FuncInfo.MBB, FuncInfo.InsertPt, DbgLoc,
             TII.get(TargetOpcode::COPY), CopyReg).addReg(InputReg);
     InputReg = CopyReg;
+    KillInputReg = true;
   }
 
   // Issue an extract_subreg.
   unsigned ResultReg = fastEmitInst_extractsubreg(MVT::i8,
-                                                  InputReg, /*Kill=*/true,
+                                                  InputReg, KillInputReg,
                                                   X86::sub_8bit);
   if (!ResultReg)
     return false;

diff --git a/test/CodeGen/X86/fast-isel-trunc-kill-subreg.ll b/test/CodeGen/X86/fast-isel-trunc-kill-subreg.ll
new file mode 100644 (file)
index 0000000..96a1a2d
--- /dev/null
+++ b/test/CodeGen/X86/fast-isel-trunc-kill-subreg.ll
@@ -0,0 +1,40 @@
+; RUN: llc %s -o - -fast-isel=true -O1 -verify-machineinstrs | FileCheck %s
+
+target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
+target triple = "x86_64-apple-unknown"
+
+; This test failed the machine verifier because the trunc at the start of the
+; method was extracing a subreg and killing the source register.  The kill flag was
+; invalid here as the source of the trunc could still be used elsewhere.
+
+; CHECK-LABEL: @test
+
+define i32 @test(i32 %block8x8) {
+bb:
+  %tmp9 = trunc i32 %block8x8 to i1
+  %tmp10 = zext i1 %tmp9 to i32
+  %tmp11 = mul i32 %tmp10, 8
+  %tmp12 = zext i32 %tmp11 to i64
+  br label %bb241
+
+bb241:                                            ; preds = %bb241, %bb
+  %lsr.iv3 = phi i64 [ %lsr.iv.next4, %bb241 ], [ %tmp12, %bb ]
+  %lsr.iv1 = phi i32 [ %lsr.iv.next2, %bb241 ], [ 0, %bb ]
+  %lsr.iv.next2 = add nuw nsw i32 %lsr.iv1, 1
+  %lsr.iv.next4 = add i64 %lsr.iv3, 32
+  %exitcond = icmp eq i32 %lsr.iv.next2, 8
+  br i1 %exitcond, label %.preheader.preheader, label %bb241
+
+.preheader.preheader:                             ; preds = %bb241
+  %tmp18 = lshr i32 %block8x8, 1
+  br label %bb270
+
+bb270:                                            ; preds = %bb270, %.preheader.preheader
+  %lsr.iv = phi i32 [ %lsr.iv.next, %bb270 ], [ %tmp18, %.preheader.preheader ]
+  %lsr.iv.next = add i32 %lsr.iv, 4
+  %tmp272 = icmp slt i32 %lsr.iv.next, 100
+  br i1 %tmp272, label %bb270, label %.loopexit
+
+.loopexit:                                        ; preds = %bb270
+  ret i32 0
+}