[SimplifyIndVars] Ignore dead users

IndVarSimplify sometimes makes transforms basing on users that are trivially dead. In particular,
if DCE wasn't run before it, there may be a dead `sext/zext` in loop that will trigger widening
transforms, however it makes no sense to do it.

This patch teaches IndVarsSimplify ignore the mist trivial cases of that.

Differential Revision: https://reviews.llvm.org/D47974
Reviewed By: sanjoy

llvm-svn: 334567

llvm/lib/Transforms/Utils/SimplifyIndVar.cpp

@@ -26,6 +26,7 @@
 #include "llvm/IR/PatternMatch.h"
 #include "llvm/Support/Debug.h"
 #include "llvm/Support/raw_ostream.h"
+#include "llvm/Transforms/Utils/Local.h"
 
 using namespace llvm;
 
@@ -772,6 +773,15 @@ void SimplifyIndvar::simplifyUsers(PHINode *CurrIV, IVVisitor *V) {
       SimpleIVUsers.pop_back_val();
     Instruction *UseInst = UseOper.first;
 
+    // If a user of the IndVar is trivially dead, we prefer just to mark it dead
+    // rather than try to do some complex analysis or transformation (such as
+    // widening) basing on it.
+    // TODO: Propagate TLI and pass it here to handle more cases.
+    if (isInstructionTriviallyDead(UseInst, /* TLI */ nullptr)) {
+      DeadInsts.emplace_back(UseInst);
+      continue;
+    }
+
     // Bypass back edges to avoid extra work.
     if (UseInst == CurrIV) continue;
 

llvm/test/Analysis/ScalarEvolution/2012-05-29-MulAddRec.ll

@@ -8,6 +8,8 @@
 ;
 ; PR12929: cast<Ty>() argument of incompatible type
 
+declare void @use(i8 %x)
+
 ; CHECK: @func
 ; CHECK: for.cond:
 ; CHECK: %inc1 = phi i8 [ 0, %entry ], [ %0, %for.body ]
@@ -33,6 +35,7 @@ for.cond:                                         ; preds = %for.cond.loopexit,
   %indvars.iv = phi i8 [ %indvars.iv.next, %for.cond.loopexit ], [ 10, %entry ]
   %mul3 = phi i8 [ undef, %entry ], [ %mul.lcssa, %for.cond.loopexit ]
   %inc1 = phi i8 [ 0, %entry ], [ %0, %for.cond.loopexit ]
+  call void @use(i8 %inc1)
   br label %for.body
 
 for.body:                                         ; preds = %for.body, %for.cond
@@ -40,6 +43,7 @@ for.body:                                         ; preds = %for.body, %for.cond
   %mul45 = phi i8 [ %mul3, %for.cond ], [ %mul, %for.body ]
   %inc = add i8 %inc26, 1
   %mul = mul i8 %inc26, %mul45
+  call void @use(i8 %inc)
   %exitcond = icmp ne i8 %inc, %indvars.iv
   br i1 %exitcond, label %for.body, label %for.cond.loopexit
 }

llvm/test/Analysis/ScalarEvolution/guards.ll

@@ -10,6 +10,8 @@ target datalayout = "n8:16:32:64"
 
 declare void @llvm.experimental.guard(i1, ...)
 
+declare void @use(i64 %x)
+
 define void @test_1(i1* %cond_buf, i32* %len_buf) {
 ; CHECK-LABEL: @test_1(
 entry:
@@ -60,6 +62,7 @@ loop:
   %iv.inc = add i32 %iv, 1
 
   %iv.sext = sext i32 %iv to i64
+  call void @use(i64 %iv.sext)
 
   %iv.inc.cmp = icmp slt i32 %iv.inc, %len
   call void(i1, ...) @llvm.experimental.guard(i1 %iv.inc.cmp) [ "deopt"() ]

llvm/test/Transforms/IndVarSimplify/2011-09-10-widen-nsw.ll

@@ -5,6 +5,8 @@
 
 target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f32:32:32-f64:64:64-v64:64:64-v128:128:128-a0:0:64-s0:64:64-f80:128:128-n8:16:32:64"
 
+declare void @use(i64 %x)
+
 define void @test() nounwind {
 entry:
   br i1 undef, label %for.body11, label %for.end285
@@ -26,15 +28,19 @@ for.body170:                                      ; preds = %for.body170, %for.b
 
   %add = add nsw i32 %i2.19, 1
   %add.idxprom = sext i32 %add to i64
+  call void @use(i64 %add.idxprom)
 
   %sub = sub nsw i32 %i2.19, 2
   %sub.idxprom = sext i32 %sub to i64
+  call void @use(i64 %sub.idxprom)
 
   %sub.neg = sub nsw i32 4, %i2.19
   %sub.neg.idxprom = sext i32 %sub.neg to i64
+  call void @use(i64 %sub.neg.idxprom)
 
   %mul = mul nsw i32 %i2.19, 8
   %mul.idxprom = sext i32 %mul to i64
+  call void @use(i64 %mul.idxprom)
 
   %add249 = add nsw i32 %i2.19, %shl132
   br label %for.body170

llvm/test/Transforms/IndVarSimplify/iv-widen.ll

@@ -7,10 +7,12 @@ target datalayout = "n8:16:32:64"
 
 target triple = "x86_64-apple-darwin"
 
+declare void @use(i64 %x)
+
 ; CHECK-LABEL: @loop_0
 ; CHECK-LABEL: B18:
 ; Only one phi now.
-; CHECK: phi
+; CHECK: phi i64
 ; CHECK-NOT: phi
 ; One trunc for the gep.
 ; CHECK: trunc i64 %indvars.iv to i32
@@ -21,6 +23,37 @@ define void @loop_0(i32* %a) {
 Prologue:
   br i1 undef, label %B18, label %B6
 
+B18:                                        ; preds = %B24, %Prologue
+  %.02 = phi i32 [ 0, %Prologue ], [ %tmp33, %B24 ]
+  %tmp23 = zext i32 %.02 to i64
+  call void @use(i64 %tmp23)
+  %tmp33 = add i32 %.02, 1
+  %o = getelementptr i32, i32* %a, i32 %.02
+  %v = load i32, i32* %o
+  %t = icmp eq i32 %v, 0
+  br i1 %t, label %exit24, label %B24
+
+B24:                                        ; preds = %B18
+  %t2 = icmp eq i32 %tmp33, 20
+  br i1 %t2, label %B6, label %B18
+
+B6:                                       ; preds = %Prologue
+  ret void
+
+exit24:                      ; preds = %B18
+  call void @dummy(i32 %.02)
+  unreachable
+}
+
+; Make sure that dead zext is removed and no widening happens.
+; CHECK-LABEL: @loop_0.dead
+; CHECK: phi i32
+; CHECK-NOT: zext
+; CHECK-NOT: trunc
+define void @loop_0.dead(i32* %a) {
+Prologue:
+  br i1 undef, label %B18, label %B6
+
 B18:                                        ; preds = %B24, %Prologue
   %.02 = phi i32 [ 0, %Prologue ], [ %tmp33, %B24 ]
   %tmp23 = zext i32 %.02 to i64

llvm/test/Transforms/IndVarSimplify/lftr-reuse.ll

@@ -6,6 +6,8 @@ target datalayout = "e-p:64:64:64-i1:8:8-i8:8:8-i16:16:16-i32:32:32-i64:64:64-f3
 
 ; Perform LFTR using the original pointer-type IV.
 
+declare void @use(double %x)
+
 ;  for(char* p = base; p < base + n; ++p) {
 ;    *p = p-base;
 ;  }
@@ -102,9 +104,11 @@ loop:
   %diagidxw = sext i32 %diagidx to i64
   %matrixp = getelementptr inbounds [0 x double], [0 x double]* %matrix, i32 0, i64 %diagidxw
   %v1 = load double, double* %matrixp
+  call void @use(double %v1)
   %iw = sext i32 %i to i64
   %vectorp = getelementptr inbounds [0 x double], [0 x double]* %vector, i32 0, i64 %iw
   %v2 = load double, double* %vectorp
+  call void @use(double %v2)
   %row.inc = add nsw i32 %rowidx, %ilead
   %i.inc = add nsw i32 %i, 1
   %cmp196 = icmp slt i32 %i.inc, %irow
@@ -125,7 +129,6 @@ entry:
 ; CHECK-NOT: add
 ; CHECK: loop:
 ; CHECK: phi i64
-; CHECK: phi i64
 ; CHECK-NOT: phi
 ; CHECK: icmp slt
 ; CHECK: br i1

llvm/test/Transforms/IndVarSimplify/loop-invariant-conditions.ll

@@ -293,6 +293,8 @@ for.end:                                          ; preds = %if.end, %entry
   ret void
 }
 
+declare void @use(i1 %x)
+
 ; check that we handle conditions with loop invariant operands which
 ; *aren't* in the header - this is a very rare and fragile case where
 ; we have a "loop" which is known to run exactly one iteration but
@@ -320,6 +322,7 @@ latch:
   br i1 true, label %exit, label %loop
 
 exit:
+  call void @use(i1 %cmp)
   ret void
 }
 

llvm/test/Transforms/IndVarSimplify/no-iv-rewrite.ll

@@ -357,6 +357,9 @@ exit:
   ret void
 }
 
+declare void @use32(i32 %x)
+declare void @use64(i64 %x)
+
 ; Test a widened IV that is used by a phi on different paths within the loop.
 ;
 ; CHECK: for.body:
@@ -385,10 +388,12 @@ if.else:
 
 if.then97:
   %idxprom100 = sext i32 %iv to i64
+  call void @use64(i64 %idxprom100)
   br label %for.inc
 
 for.inc:
   %kmin.1 = phi i32 [ %iv, %if.then33 ], [ 0, %if.then ], [ %iv, %if.then97 ], [ 0, %if.else ]
+  call void @use32(i32 %kmin.1)
   %inc = add nsw i32 %iv, 1
   br i1 undef, label %for.body, label %for.end
 

llvm/test/Transforms/IndVarSimplify/pr25578.ll

@@ -1,6 +1,8 @@
 ; RUN: opt < %s -indvars -S | FileCheck %s
 target datalayout = "e-m:o-i64:64-f80:128-n8:16:32:64-S128"
 
+declare void @use(i64 %x)
+
 ; CHECK-LABEL: @foo
 define void @foo() {
 entry:
@@ -15,6 +17,7 @@ L1_header:
 L2_header:
   %i = phi i32 [ 0, %L1_header ], [ %i_next, %L2_latch ]
   %i_prom = sext i32 %i to i64
+  call void @use(i64 %i_prom)
   br label %L3_header
 
 L3_header: