[Transforms][LICM] A test case for the upcoming fix D152281 for the issue with reassociation profitability

[pawosm01]

Summary

This commit introduces a test for the upcoming change addressing
the following issue: https://github.com/llvm/llvm-project/issues/62736

[fhahn]

llvm/test/Transforms/LICM/expr-reassociate.ll

3

Usually the tests for a specific transform shouldn't depend on other passes; You could add a test to PhaseOrdering that tests the end-to-end flow (e.g. with -O2) and then here have just the ones that run licm on the input IR that requires reassoication to move it out of the loop

[pawosm01]

llvm/test/Transforms/LICM/expr-reassociate.ll

3

I was about to move a vital part of this test into the Transforms/PhaseOrdering but I was encouraged not to. The test as it is now captures the interplay between those two passes quite well and is self-contained. Those end-to-end tests are pretty rare.

[pawosm01]

Extending it with an additional test case exposing shortcomings that have been fixed.

[pawosm01]

Followed a reviewer's suggestion to limit the applicability of the attempted transformation.

[pawosm01]

One more test case added.

llvm/test/Transforms/LICM/expr-reassociate.ll

+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 2
+; RUN: opt -passes='reassociate' -S < %s | FileCheck %s --check-prefix=REASSOCIATE_ONLY
+; RUN: opt -passes='licm' -S < %s | FileCheck %s --check-prefix=LICM_ONLY
+; RUN: opt -passes='reassociate,loop-mssa(licm)' -S < %s | FileCheck %s --check-prefix=LICM_AFTER_REASSOCIATE
+
+;
+; A simple loop, should not get modified:
+;
+;  int j;
+;  const double d1d = d1 * delta;
+;
+;  for (j = 0; j <= i; j++)
+;    cells[j] = d1d * cells[j + 1];
+;
+
+define void @innermost_loop_1d_fast(i32 %i, double %d1, double %delta, ptr %cells) {
+; REASSOCIATE_ONLY-LABEL: define void @innermost_loop_1d_fast
+; REASSOCIATE_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; REASSOCIATE_ONLY-NEXT:  entry:
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_D1:%.*]] = fmul fast double [[DELTA]], [[D1]]
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; REASSOCIATE_ONLY:       for.cond:
+; REASSOCIATE_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; REASSOCIATE_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; REASSOCIATE_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; REASSOCIATE_ONLY:       for.body:
+; REASSOCIATE_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[FMUL_D1]], [[CELL_1]]
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; REASSOCIATE_ONLY-NEXT:    store double [[FMUL_1]], ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND]]
+; REASSOCIATE_ONLY:       for.end:
+; REASSOCIATE_ONLY-NEXT:    ret void
+;
+; LICM_ONLY-LABEL: define void @innermost_loop_1d_fast
+; LICM_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_ONLY-NEXT:  entry:
+; LICM_ONLY-NEXT:    [[FMUL_D1:%.*]] = fmul fast double [[D1]], [[DELTA]]
+; LICM_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_ONLY:       for.cond:
+; LICM_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_ONLY:       for.body:
+; LICM_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[FMUL_D1]], [[CELL_1]]
+; LICM_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_ONLY-NEXT:    store double [[FMUL_1]], ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    br label [[FOR_COND]]
+; LICM_ONLY:       for.end:
+; LICM_ONLY-NEXT:    ret void
+;
+; LICM_AFTER_REASSOCIATE-LABEL: define void @innermost_loop_1d_fast
+; LICM_AFTER_REASSOCIATE-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_AFTER_REASSOCIATE-NEXT:  entry:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_D1:%.*]] = fmul fast double [[DELTA]], [[D1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_AFTER_REASSOCIATE:       for.cond:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_AFTER_REASSOCIATE-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_AFTER_REASSOCIATE:       for.body:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[FMUL_D1]], [[CELL_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_AFTER_REASSOCIATE-NEXT:    store double [[FMUL_1]], ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND]]
+; LICM_AFTER_REASSOCIATE:       for.end:
+; LICM_AFTER_REASSOCIATE-NEXT:    ret void
+;
+entry:
+  %fmul.d1 = fmul fast double %d1, %delta
+  br label %for.cond
+
+for.cond:
+  %j = phi i32 [ 0, %entry ], [ %add.j.1, %for.body ]
+  %cmp.not = icmp sgt i32 %j, %i
+  br i1 %cmp.not, label %for.end, label %for.body
+
+for.body:
+  %add.j.1 = add nuw nsw i32 %j, 1
+  %idxprom.j.1 = zext i32 %add.j.1 to i64
+  %arrayidx.j.1 = getelementptr inbounds double, ptr %cells, i64 %idxprom.j.1
+  %cell.1 = load double, ptr %arrayidx.j.1, align 8
+  %fmul.1 = fmul fast double %fmul.d1, %cell.1
+  %idxprom.j = zext i32 %j to i64
+  %arrayidx.j = getelementptr inbounds double, ptr %cells, i64 %idxprom.j
+  store double %fmul.1, ptr %arrayidx.j, align 8
+  br label %for.cond
+
+for.end:
+  ret void
+}
+
+;
+; A simple loop:
+;
+;  int j;
+;
+;  for (j = 0; j <= i; j++)
+;    cells[j] = d1 * cells[j + 1] * delta;
+;
+; ...should be transformed by the LICM pass into this:
+;
+;  int j;
+;  const double d1d = d1 * delta;
+;
+;  for (j = 0; j <= i; j++)
+;    cells[j] = d1d * cells[j + 1];
+;
+
+define void @innermost_loop_1d_shouldhoist_fast(i32 %i, double %d1, double %delta, ptr %cells) {
+; REASSOCIATE_ONLY-LABEL: define void @innermost_loop_1d_shouldhoist_fast
+; REASSOCIATE_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; REASSOCIATE_ONLY-NEXT:  entry:
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; REASSOCIATE_ONLY:       for.cond:
+; REASSOCIATE_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; REASSOCIATE_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; REASSOCIATE_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; REASSOCIATE_ONLY:       for.body:
+; REASSOCIATE_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[DELTA]], [[D1]]
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[FMUL_1]], [[CELL_1]]
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; REASSOCIATE_ONLY-NEXT:    store double [[FMUL_2]], ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND]]
+; REASSOCIATE_ONLY:       for.end:
+; REASSOCIATE_ONLY-NEXT:    ret void
+;
+; LICM_ONLY-LABEL: define void @innermost_loop_1d_shouldhoist_fast
+; LICM_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_ONLY-NEXT:  entry:
+; LICM_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_ONLY:       for.cond:
+; LICM_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_ONLY:       for.body:
+; LICM_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[D1]], [[CELL_1]]
+; LICM_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[FMUL_1]], [[DELTA]]
+; LICM_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_ONLY-NEXT:    store double [[FMUL_2]], ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    br label [[FOR_COND]]
+; LICM_ONLY:       for.end:
+; LICM_ONLY-NEXT:    ret void
+;
+; LICM_AFTER_REASSOCIATE-LABEL: define void @innermost_loop_1d_shouldhoist_fast
+; LICM_AFTER_REASSOCIATE-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_AFTER_REASSOCIATE-NEXT:  entry:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[DELTA]], [[D1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_AFTER_REASSOCIATE:       for.cond:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_AFTER_REASSOCIATE-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_AFTER_REASSOCIATE:       for.body:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[FMUL_1]], [[CELL_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_AFTER_REASSOCIATE-NEXT:    store double [[FMUL_2]], ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND]]
+; LICM_AFTER_REASSOCIATE:       for.end:
+; LICM_AFTER_REASSOCIATE-NEXT:    ret void
+;
+entry:
+  br label %for.cond
+
+for.cond:
+  %j = phi i32 [ 0, %entry ], [ %add.j.1, %for.body ]
+  %cmp.not = icmp sgt i32 %j, %i
+  br i1 %cmp.not, label %for.end, label %for.body
+
+for.body:
+  %add.j.1 = add nuw nsw i32 %j, 1
+  %idxprom.j.1 = zext i32 %add.j.1 to i64
+  %arrayidx.j.1 = getelementptr inbounds double, ptr %cells, i64 %idxprom.j.1
+  %cell.1 = load double, ptr %arrayidx.j.1, align 8
+  %fmul.1 = fmul fast double %d1, %cell.1
+  %fmul.2 = fmul fast double %fmul.1, %delta
+  %idxprom.j = zext i32 %j to i64
+  %arrayidx.j = getelementptr inbounds double, ptr %cells, i64 %idxprom.j
+  store double %fmul.2, ptr %arrayidx.j, align 8
+  br label %for.cond
+
+for.end:
+  ret void
+}
+
+;
+; The following loop will be modified by the 'Reassociate expressions' pass,
+;
+;  int j;
+;  const double d1d = d1 * delta;
+;  const double d2d = d2 * delta;
+;
+;  for (j = 0; j <= i; j++)
+;    cells[j] = d1d * cells[j + 1] + d2d * cells[j];
+;
+; ...into this:
+;
+;  int j;
+;
+;  for (j = 0; j <= i; j++)
+;    cells[j] = (d1 * cells[j + 1] + d2 * cells[j]) * delta;
+;
+; We expect the LICM pass to undo this transformation.
+;
+
+define void @innermost_loop_2d_fast(i32 %i, double %d1, double %d2, double %delta, ptr %cells) {
+; REASSOCIATE_ONLY-LABEL: define void @innermost_loop_2d_fast
+; REASSOCIATE_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; REASSOCIATE_ONLY-NEXT:  entry:
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; REASSOCIATE_ONLY:       for.cond:
+; REASSOCIATE_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; REASSOCIATE_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; REASSOCIATE_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; REASSOCIATE_ONLY:       for.body:
+; REASSOCIATE_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[CELL_1]], [[D1]]
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[CELL_2]], [[D2]]
+; REASSOCIATE_ONLY-NEXT:    [[REASS_ADD:%.*]] = fadd fast double [[FMUL_2]], [[FMUL_1]]
+; REASSOCIATE_ONLY-NEXT:    [[REASS_MUL:%.*]] = fmul fast double [[REASS_ADD]], [[DELTA]]
+; REASSOCIATE_ONLY-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND]]
+; REASSOCIATE_ONLY:       for.end:
+; REASSOCIATE_ONLY-NEXT:    ret void
+;
+; LICM_ONLY-LABEL: define void @innermost_loop_2d_fast
+; LICM_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_ONLY-NEXT:  entry:
+; LICM_ONLY-NEXT:    [[FMUL_D1:%.*]] = fmul fast double [[D1]], [[DELTA]]
+; LICM_ONLY-NEXT:    [[FMUL_D2:%.*]] = fmul fast double [[D2]], [[DELTA]]
+; LICM_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_ONLY:       for.cond:
+; LICM_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_ONLY:       for.body:
+; LICM_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[FMUL_D1]], [[CELL_1]]
+; LICM_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[FMUL_D2]], [[CELL_2]]
+; LICM_ONLY-NEXT:    [[FADD_1:%.*]] = fadd fast double [[FMUL_1]], [[FMUL_2]]
+; LICM_ONLY-NEXT:    store double [[FADD_1]], ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    br label [[FOR_COND]]
+; LICM_ONLY:       for.end:
+; LICM_ONLY-NEXT:    ret void
+;
+; LICM_AFTER_REASSOCIATE-LABEL: define void @innermost_loop_2d_fast
+; LICM_AFTER_REASSOCIATE-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_AFTER_REASSOCIATE-NEXT:  entry:
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_AFTER_REASSOCIATE:       for.cond:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_AFTER_REASSOCIATE-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_AFTER_REASSOCIATE:       for.body:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[CELL_1]], [[D1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[CELL_2]], [[D2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[REASS_ADD:%.*]] = fadd fast double [[FMUL_2]], [[FMUL_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[REASS_MUL:%.*]] = fmul fast double [[REASS_ADD]], [[DELTA]]
+; LICM_AFTER_REASSOCIATE-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND]]
+; LICM_AFTER_REASSOCIATE:       for.end:
+; LICM_AFTER_REASSOCIATE-NEXT:    ret void
+;
+entry:
+  %fmul.d1 = fmul fast double %d1, %delta
+  %fmul.d2 = fmul fast double %d2, %delta
+  br label %for.cond
+
+for.cond:
+  %j = phi i32 [ 0, %entry ], [ %add.j.1, %for.body ]
+  %cmp.not = icmp sgt i32 %j, %i
+  br i1 %cmp.not, label %for.end, label %for.body
+
+for.body:
+  %add.j.1 = add nuw nsw i32 %j, 1
+  %idxprom.j.1 = zext i32 %add.j.1 to i64
+  %arrayidx.j.1 = getelementptr inbounds double, ptr %cells, i64 %idxprom.j.1
+  %cell.1 = load double, ptr %arrayidx.j.1, align 8
+  %fmul.1 = fmul fast double %fmul.d1, %cell.1
+  %idxprom.j = zext i32 %j to i64
+  %arrayidx.j = getelementptr inbounds double, ptr %cells, i64 %idxprom.j
+  %cell.2 = load double, ptr %arrayidx.j, align 8
+  %fmul.2 = fmul fast double %fmul.d2, %cell.2
+  %fadd.1 = fadd fast double %fmul.1, %fmul.2
+  store double %fadd.1, ptr %arrayidx.j, align 8
+  br label %for.cond
+
+for.end:
+  ret void
+}
+
+;
+; The following loop will be modified by the 'Reassociate expressions' pass,
+;
+;  int j;
+;  const double d1d = d1 * delta;
+;  const double d2d = d2 * delta;
+;  const double d3d = d3 * delta;
+;
+;  for (j = 0; j <= i; j++)
+;    cells[j] = d1d * cells[j + 1] + d2d * cells[j] + d3d * cells[j + 2];
+;
+; ...into this:
+;
+;  int j;
+;
+;  for (j = 0; j <= i; j++)
+;    cells[j] = (d1 * cells[j + 1] + d2 * cells[j] + d3 * cells[j + 2]) * delta;
+;
+; We expect the LICM pass to undo this transformation.
+;
+
+
+define void @innermost_loop_3d_fast(i32 %i, double %d1, double %d2, double %d3, double %delta, ptr %cells) {
+; REASSOCIATE_ONLY-LABEL: define void @innermost_loop_3d_fast
+; REASSOCIATE_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[D3:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; REASSOCIATE_ONLY-NEXT:  entry:
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; REASSOCIATE_ONLY:       for.cond:
+; REASSOCIATE_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; REASSOCIATE_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; REASSOCIATE_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; REASSOCIATE_ONLY:       for.body:
+; REASSOCIATE_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[CELL_1]], [[D1]]
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[CELL_2]], [[D2]]
+; REASSOCIATE_ONLY-NEXT:    [[ADD_J_2:%.*]] = add nuw nsw i32 [[J]], 2
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J_2:%.*]] = zext i32 [[ADD_J_2]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J_2:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_2]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_3:%.*]] = load double, ptr [[ARRAYIDX_J_2]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_3:%.*]] = fmul fast double [[CELL_3]], [[D3]]
+; REASSOCIATE_ONLY-NEXT:    [[REASS_ADD:%.*]] = fadd fast double [[FMUL_2]], [[FMUL_1]]
+; REASSOCIATE_ONLY-NEXT:    [[REASS_ADD1:%.*]] = fadd fast double [[REASS_ADD]], [[FMUL_3]]
+; REASSOCIATE_ONLY-NEXT:    [[REASS_MUL:%.*]] = fmul fast double [[REASS_ADD1]], [[DELTA]]
+; REASSOCIATE_ONLY-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J_2]], align 8
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND]]
+; REASSOCIATE_ONLY:       for.end:
+; REASSOCIATE_ONLY-NEXT:    ret void
+;
+; LICM_ONLY-LABEL: define void @innermost_loop_3d_fast
+; LICM_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[D3:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_ONLY-NEXT:  entry:
+; LICM_ONLY-NEXT:    [[FMUL_D1:%.*]] = fmul fast double [[D1]], [[DELTA]]
+; LICM_ONLY-NEXT:    [[FMUL_D2:%.*]] = fmul fast double [[D2]], [[DELTA]]
+; LICM_ONLY-NEXT:    [[FMUL_D3:%.*]] = fmul fast double [[D3]], [[DELTA]]
+; LICM_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_ONLY:       for.cond:
+; LICM_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_ONLY:       for.body:
+; LICM_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[FMUL_D1]], [[CELL_1]]
+; LICM_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[FMUL_D2]], [[CELL_2]]
+; LICM_ONLY-NEXT:    [[FADD_1:%.*]] = fadd fast double [[FMUL_1]], [[FMUL_2]]
+; LICM_ONLY-NEXT:    [[ADD_J_2:%.*]] = add nuw nsw i32 [[J]], 2
+; LICM_ONLY-NEXT:    [[IDXPROM_J_2:%.*]] = zext i32 [[ADD_J_2]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J_2:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_2]]
+; LICM_ONLY-NEXT:    [[CELL_3:%.*]] = load double, ptr [[ARRAYIDX_J_2]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_3:%.*]] = fmul fast double [[FMUL_D3]], [[CELL_3]]
+; LICM_ONLY-NEXT:    [[FADD_2:%.*]] = fadd fast double [[FADD_1]], [[FMUL_3]]
+; LICM_ONLY-NEXT:    store double [[FADD_2]], ptr [[ARRAYIDX_J_2]], align 8
+; LICM_ONLY-NEXT:    br label [[FOR_COND]]
+; LICM_ONLY:       for.end:
+; LICM_ONLY-NEXT:    ret void
+;
+; LICM_AFTER_REASSOCIATE-LABEL: define void @innermost_loop_3d_fast
+; LICM_AFTER_REASSOCIATE-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[D3:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_AFTER_REASSOCIATE-NEXT:  entry:
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_AFTER_REASSOCIATE:       for.cond:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_AFTER_REASSOCIATE-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_AFTER_REASSOCIATE:       for.body:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[CELL_1]], [[D1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[CELL_2]], [[D2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ADD_J_2:%.*]] = add nuw nsw i32 [[J]], 2
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J_2:%.*]] = zext i32 [[ADD_J_2]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J_2:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_3:%.*]] = load double, ptr [[ARRAYIDX_J_2]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_3:%.*]] = fmul fast double [[CELL_3]], [[D3]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[REASS_ADD:%.*]] = fadd fast double [[FMUL_2]], [[FMUL_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[REASS_ADD1:%.*]] = fadd fast double [[REASS_ADD]], [[FMUL_3]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[REASS_MUL:%.*]] = fmul fast double [[REASS_ADD1]], [[DELTA]]
+; LICM_AFTER_REASSOCIATE-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J_2]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND]]
+; LICM_AFTER_REASSOCIATE:       for.end:
+; LICM_AFTER_REASSOCIATE-NEXT:    ret void
+;
+entry:
+  %fmul.d1 = fmul fast double %d1, %delta
+  %fmul.d2 = fmul fast double %d2, %delta
+  %fmul.d3 = fmul fast double %d3, %delta
+  br label %for.cond
+
+for.cond:
+  %j = phi i32 [ 0, %entry ], [ %add.j.1, %for.body ]
+  %cmp.not = icmp sgt i32 %j, %i
+  br i1 %cmp.not, label %for.end, label %for.body
+
+for.body:
+  %add.j.1 = add nuw nsw i32 %j, 1
+  %idxprom.j.1 = zext i32 %add.j.1 to i64
+  %arrayidx.j.1 = getelementptr inbounds double, ptr %cells, i64 %idxprom.j.1
+  %cell.1 = load double, ptr %arrayidx.j.1, align 8
+  %fmul.1 = fmul fast double %fmul.d1, %cell.1
+  %idxprom.j = zext i32 %j to i64
+  %arrayidx.j = getelementptr inbounds double, ptr %cells, i64 %idxprom.j
+  %cell.2 = load double, ptr %arrayidx.j, align 8
+  %fmul.2 = fmul fast double %fmul.d2, %cell.2
+  %fadd.1 = fadd fast double %fmul.1, %fmul.2
+  %add.j.2 = add nuw nsw i32 %j, 2
+  %idxprom.j.2 = zext i32 %add.j.2 to i64
+  %arrayidx.j.2 = getelementptr inbounds double, ptr %cells, i64 %idxprom.j.2
+  %cell.3 = load double, ptr %arrayidx.j.2, align 8
+  %fmul.3 = fmul fast double %fmul.d3, %cell.3
+  %fadd.2 = fadd fast double %fadd.1, %fmul.3
+  store double %fadd.2, ptr %arrayidx.j.2, align 8
+  br label %for.cond
+
+for.end:
+  ret void
+}
+
+;
+; When there is no 'fast' attribute, the transformation should not happen.
+;
+
+define void @innermost_loop_2d_nofast(i32 %i, double %d1, double %d2, double %delta, ptr %cells) {
+; REASSOCIATE_ONLY-LABEL: define void @innermost_loop_2d_nofast
+; REASSOCIATE_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; REASSOCIATE_ONLY-NEXT:  entry:
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_D1:%.*]] = fmul double [[D1]], [[DELTA]]
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_D2:%.*]] = fmul double [[D2]], [[DELTA]]
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; REASSOCIATE_ONLY:       for.cond:
+; REASSOCIATE_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; REASSOCIATE_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; REASSOCIATE_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; REASSOCIATE_ONLY:       for.body:
+; REASSOCIATE_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul double [[FMUL_D1]], [[CELL_1]]
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul double [[FMUL_D2]], [[CELL_2]]
+; REASSOCIATE_ONLY-NEXT:    [[FADD_1:%.*]] = fadd double [[FMUL_1]], [[FMUL_2]]
+; REASSOCIATE_ONLY-NEXT:    store double [[FADD_1]], ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND]]
+; REASSOCIATE_ONLY:       for.end:
+; REASSOCIATE_ONLY-NEXT:    ret void
+;
+; LICM_ONLY-LABEL: define void @innermost_loop_2d_nofast
+; LICM_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_ONLY-NEXT:  entry:
+; LICM_ONLY-NEXT:    [[FMUL_D1:%.*]] = fmul double [[D1]], [[DELTA]]
+; LICM_ONLY-NEXT:    [[FMUL_D2:%.*]] = fmul double [[D2]], [[DELTA]]
+; LICM_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_ONLY:       for.cond:
+; LICM_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_ONLY:       for.body:
+; LICM_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul double [[FMUL_D1]], [[CELL_1]]
+; LICM_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul double [[FMUL_D2]], [[CELL_2]]
+; LICM_ONLY-NEXT:    [[FADD_1:%.*]] = fadd double [[FMUL_1]], [[FMUL_2]]
+; LICM_ONLY-NEXT:    store double [[FADD_1]], ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    br label [[FOR_COND]]
+; LICM_ONLY:       for.end:
+; LICM_ONLY-NEXT:    ret void
+;
+; LICM_AFTER_REASSOCIATE-LABEL: define void @innermost_loop_2d_nofast
+; LICM_AFTER_REASSOCIATE-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_AFTER_REASSOCIATE-NEXT:  entry:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_D1:%.*]] = fmul double [[D1]], [[DELTA]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_D2:%.*]] = fmul double [[D2]], [[DELTA]]
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_AFTER_REASSOCIATE:       for.cond:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_AFTER_REASSOCIATE-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_AFTER_REASSOCIATE:       for.body:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_1:%.*]] = fmul double [[FMUL_D1]], [[CELL_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_2:%.*]] = fmul double [[FMUL_D2]], [[CELL_2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FADD_1:%.*]] = fadd double [[FMUL_1]], [[FMUL_2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    store double [[FADD_1]], ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND]]
+; LICM_AFTER_REASSOCIATE:       for.end:
+; LICM_AFTER_REASSOCIATE-NEXT:    ret void
+;
+entry:
+  %fmul.d1 = fmul double %d1, %delta
+  %fmul.d2 = fmul double %d2, %delta
+  br label %for.cond
+
+for.cond:
+  %j = phi i32 [ 0, %entry ], [ %add.j.1, %for.body ]
+  %cmp.not = icmp sgt i32 %j, %i
+  br i1 %cmp.not, label %for.end, label %for.body
+
+for.body:
+  %add.j.1 = add nuw nsw i32 %j, 1
+  %idxprom.j.1 = zext i32 %add.j.1 to i64
+  %arrayidx.j.1 = getelementptr inbounds double, ptr %cells, i64 %idxprom.j.1
+  %cell.1 = load double, ptr %arrayidx.j.1, align 8
+  %fmul.1 = fmul double %fmul.d1, %cell.1
+  %idxprom.j = zext i32 %j to i64
+  %arrayidx.j = getelementptr inbounds double, ptr %cells, i64 %idxprom.j
+  %cell.2 = load double, ptr %arrayidx.j, align 8
+  %fmul.2 = fmul double %fmul.d2, %cell.2
+  %fadd.1 = fadd double %fmul.1, %fmul.2
+  store double %fadd.1, ptr %arrayidx.j, align 8
+  br label %for.cond
+
+for.end:
+  ret void
+}
+
+;
+; The following loop should be modified by the LICM pass,
+;
+;  int j;
+;
+;  for (j = 0; j <= i; j++)
+;    cells[j] = (d1 * cells[j + 1] + d2 * cells[j]) * delta;
+;
+; ...into this:
+;
+;  int j;
+;  const double d1d = d1 * delta;
+;  const double d2d = d2 * delta;
+;
+;  for (j = 0; j <= i; j++)
+;    cells[j] = d1d * cells[j + 1] + d2d * cells[j];
+;
+
+define void @innermost_loop_2d_fast_reassociated(i32 %i, double %d1, double %d2, double %delta, ptr %cells) {
+; REASSOCIATE_ONLY-LABEL: define void @innermost_loop_2d_fast_reassociated
+; REASSOCIATE_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; REASSOCIATE_ONLY-NEXT:  entry:
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; REASSOCIATE_ONLY:       for.cond:
+; REASSOCIATE_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; REASSOCIATE_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; REASSOCIATE_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; REASSOCIATE_ONLY:       for.body:
+; REASSOCIATE_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[CELL_1]], [[D1]]
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[CELL_2]], [[D2]]
+; REASSOCIATE_ONLY-NEXT:    [[REASS_ADD:%.*]] = fadd fast double [[FMUL_2]], [[FMUL_1]]
+; REASSOCIATE_ONLY-NEXT:    [[REASS_MUL:%.*]] = fmul fast double [[REASS_ADD]], [[DELTA]]
+; REASSOCIATE_ONLY-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND]]
+; REASSOCIATE_ONLY:       for.end:
+; REASSOCIATE_ONLY-NEXT:    ret void
+;
+; LICM_ONLY-LABEL: define void @innermost_loop_2d_fast_reassociated
+; LICM_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_ONLY-NEXT:  entry:
+; LICM_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_ONLY:       for.cond:
+; LICM_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_ONLY:       for.body:
+; LICM_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[CELL_1]], [[D1]]
+; LICM_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[CELL_2]], [[D2]]
+; LICM_ONLY-NEXT:    [[REASS_ADD:%.*]] = fadd fast double [[FMUL_2]], [[FMUL_1]]
+; LICM_ONLY-NEXT:    [[REASS_MUL:%.*]] = fmul fast double [[REASS_ADD]], [[DELTA]]
+; LICM_ONLY-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    br label [[FOR_COND]]
+; LICM_ONLY:       for.end:
+; LICM_ONLY-NEXT:    ret void
+;
+; LICM_AFTER_REASSOCIATE-LABEL: define void @innermost_loop_2d_fast_reassociated
+; LICM_AFTER_REASSOCIATE-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_AFTER_REASSOCIATE-NEXT:  entry:
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_AFTER_REASSOCIATE:       for.cond:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_AFTER_REASSOCIATE-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_AFTER_REASSOCIATE:       for.body:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[CELL_1]], [[D1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[CELL_2]], [[D2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[REASS_ADD:%.*]] = fadd fast double [[FMUL_2]], [[FMUL_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[REASS_MUL:%.*]] = fmul fast double [[REASS_ADD]], [[DELTA]]
+; LICM_AFTER_REASSOCIATE-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND]]
+; LICM_AFTER_REASSOCIATE:       for.end:
+; LICM_AFTER_REASSOCIATE-NEXT:    ret void
+;
+entry:
+  br label %for.cond
+
+for.cond:
+  %j = phi i32 [ 0, %entry ], [ %add.j.1, %for.body ]
+  %cmp.not = icmp sgt i32 %j, %i
+  br i1 %cmp.not, label %for.end, label %for.body
+
+for.body:
+  %add.j.1 = add nuw nsw i32 %j, 1
+  %idxprom.j.1 = zext i32 %add.j.1 to i64
+  %arrayidx.j.1 = getelementptr inbounds double, ptr %cells, i64 %idxprom.j.1
+  %cell.1 = load double, ptr %arrayidx.j.1, align 8
+  %fmul.1 = fmul fast double %cell.1, %d1
+  %idxprom.j = zext i32 %j to i64
+  %arrayidx.j = getelementptr inbounds double, ptr %cells, i64 %idxprom.j
+  %cell.2 = load double, ptr %arrayidx.j, align 8
+  %fmul.2 = fmul fast double %cell.2, %d2
+  %reass.add = fadd fast double %fmul.2, %fmul.1
+  %reass.mul = fmul fast double %reass.add, %delta
+  store double %reass.mul, ptr %arrayidx.j, align 8
+  br label %for.cond
+
+for.end:
+  ret void
+}
+
+;
+; When there is no 'fast' attribute, the transformation should not happen.
+;
+
+define void @innermost_loop_2d_nofast_reassociated(i32 %i, double %d1, double %d2, double %delta, ptr %cells) {
+; REASSOCIATE_ONLY-LABEL: define void @innermost_loop_2d_nofast_reassociated
+; REASSOCIATE_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; REASSOCIATE_ONLY-NEXT:  entry:
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; REASSOCIATE_ONLY:       for.cond:
+; REASSOCIATE_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; REASSOCIATE_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; REASSOCIATE_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; REASSOCIATE_ONLY:       for.body:
+; REASSOCIATE_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul double [[D1]], [[CELL_1]]
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul double [[D2]], [[CELL_2]]
+; REASSOCIATE_ONLY-NEXT:    [[REASS_ADD:%.*]] = fadd double [[FMUL_1]], [[FMUL_2]]
+; REASSOCIATE_ONLY-NEXT:    [[REASS_MUL:%.*]] = fmul double [[DELTA]], [[REASS_ADD]]
+; REASSOCIATE_ONLY-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND]]
+; REASSOCIATE_ONLY:       for.end:
+; REASSOCIATE_ONLY-NEXT:    ret void
+;
+; LICM_ONLY-LABEL: define void @innermost_loop_2d_nofast_reassociated
+; LICM_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_ONLY-NEXT:  entry:
+; LICM_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_ONLY:       for.cond:
+; LICM_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_ONLY:       for.body:
+; LICM_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul double [[CELL_1]], [[D1]]
+; LICM_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul double [[CELL_2]], [[D2]]
+; LICM_ONLY-NEXT:    [[REASS_ADD:%.*]] = fadd double [[FMUL_2]], [[FMUL_1]]
+; LICM_ONLY-NEXT:    [[REASS_MUL:%.*]] = fmul double [[REASS_ADD]], [[DELTA]]
+; LICM_ONLY-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    br label [[FOR_COND]]
+; LICM_ONLY:       for.end:
+; LICM_ONLY-NEXT:    ret void
+;
+; LICM_AFTER_REASSOCIATE-LABEL: define void @innermost_loop_2d_nofast_reassociated
+; LICM_AFTER_REASSOCIATE-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_AFTER_REASSOCIATE-NEXT:  entry:
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_AFTER_REASSOCIATE:       for.cond:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_AFTER_REASSOCIATE-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_AFTER_REASSOCIATE:       for.body:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_1:%.*]] = fmul double [[D1]], [[CELL_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_2:%.*]] = fmul double [[D2]], [[CELL_2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[REASS_ADD:%.*]] = fadd double [[FMUL_1]], [[FMUL_2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[REASS_MUL:%.*]] = fmul double [[DELTA]], [[REASS_ADD]]
+; LICM_AFTER_REASSOCIATE-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND]]
+; LICM_AFTER_REASSOCIATE:       for.end:
+; LICM_AFTER_REASSOCIATE-NEXT:    ret void
+;
+entry:
+  br label %for.cond
+
+for.cond:
+  %j = phi i32 [ 0, %entry ], [ %add.j.1, %for.body ]
+  %cmp.not = icmp sgt i32 %j, %i
+  br i1 %cmp.not, label %for.end, label %for.body
+
+for.body:
+  %add.j.1 = add nuw nsw i32 %j, 1
+  %idxprom.j.1 = zext i32 %add.j.1 to i64
+  %arrayidx.j.1 = getelementptr inbounds double, ptr %cells, i64 %idxprom.j.1
+  %cell.1 = load double, ptr %arrayidx.j.1, align 8
+  %fmul.1 = fmul double %cell.1, %d1
+  %idxprom.j = zext i32 %j to i64
+  %arrayidx.j = getelementptr inbounds double, ptr %cells, i64 %idxprom.j
+  %cell.2 = load double, ptr %arrayidx.j, align 8
+  %fmul.2 = fmul double %cell.2, %d2
+  %reass.add = fadd double %fmul.2, %fmul.1
+  %reass.mul = fmul double %reass.add, %delta
+  store double %reass.mul, ptr %arrayidx.j, align 8
+  br label %for.cond
+
+for.end:
+  ret void
+}
+
+;
+; The following loop will not be modified by the LICM pass:
+;
+;  int j;
+;
+;  for (j = 0; j <= i; j++)
+;    cells[j] = (d1 * cells[j + 1] + d2 * cells[j] +
+;                cells[j] * cells[j + 1]) * delta;
+;
+; This case differs as one of the multiplications involves no invariants.
+;
+
+define void @innermost_loop_3d_fast_reassociated_different(i32 %i, double %d1, double %d2, double %delta, ptr %cells) {
+; REASSOCIATE_ONLY-LABEL: define void @innermost_loop_3d_fast_reassociated_different
+; REASSOCIATE_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; REASSOCIATE_ONLY-NEXT:  entry:
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; REASSOCIATE_ONLY:       for.cond:
+; REASSOCIATE_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; REASSOCIATE_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; REASSOCIATE_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; REASSOCIATE_ONLY:       for.body:
+; REASSOCIATE_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J_2:%.*]] = zext i32 [[ADD_J_1]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J_2:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_2]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J_2]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J_3:%.*]] = zext i32 [[ADD_J_1]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[CELL_3:%.*]] = load double, ptr [[ARRAYIDX_J_2]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; REASSOCIATE_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; REASSOCIATE_ONLY-NEXT:    [[CELL_4:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[CELL_1]], [[D1]]
+; REASSOCIATE_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[CELL_4]], [[D2]]
+; REASSOCIATE_ONLY-NEXT:    [[EXTRA_MUL:%.*]] = fmul fast double [[CELL_3]], [[CELL_2]]
+; REASSOCIATE_ONLY-NEXT:    [[REASS_ADD:%.*]] = fadd fast double [[EXTRA_MUL]], [[FMUL_1]]
+; REASSOCIATE_ONLY-NEXT:    [[EXTRA_ADD:%.*]] = fadd fast double [[REASS_ADD]], [[FMUL_2]]
+; REASSOCIATE_ONLY-NEXT:    [[REASS_MUL:%.*]] = fmul fast double [[EXTRA_ADD]], [[DELTA]]
+; REASSOCIATE_ONLY-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
+; REASSOCIATE_ONLY-NEXT:    br label [[FOR_COND]]
+; REASSOCIATE_ONLY:       for.end:
+; REASSOCIATE_ONLY-NEXT:    ret void
+;
+; LICM_ONLY-LABEL: define void @innermost_loop_3d_fast_reassociated_different
+; LICM_ONLY-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_ONLY-NEXT:  entry:
+; LICM_ONLY-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_ONLY:       for.cond:
+; LICM_ONLY-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_ONLY-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_ONLY-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_ONLY:       for.body:
+; LICM_ONLY-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_ONLY-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_ONLY-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_ONLY-NEXT:    [[IDXPROM_J_2:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J_2:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_2]]
+; LICM_ONLY-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J_2]], align 8
+; LICM_ONLY-NEXT:    [[CELL_3:%.*]] = load double, ptr [[ARRAYIDX_J_2]], align 8
+; LICM_ONLY-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_ONLY-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_ONLY-NEXT:    [[CELL_4:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[CELL_1]], [[D1]]
+; LICM_ONLY-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[CELL_4]], [[D2]]
+; LICM_ONLY-NEXT:    [[REASS_ADD:%.*]] = fadd fast double [[FMUL_2]], [[FMUL_1]]
+; LICM_ONLY-NEXT:    [[EXTRA_MUL:%.*]] = fmul fast double [[CELL_3]], [[CELL_2]]
+; LICM_ONLY-NEXT:    [[EXTRA_ADD:%.*]] = fadd fast double [[EXTRA_MUL]], [[REASS_ADD]]
+; LICM_ONLY-NEXT:    [[REASS_MUL:%.*]] = fmul fast double [[EXTRA_ADD]], [[DELTA]]
+; LICM_ONLY-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
+; LICM_ONLY-NEXT:    br label [[FOR_COND]]
+; LICM_ONLY:       for.end:
+; LICM_ONLY-NEXT:    ret void
+;
+; LICM_AFTER_REASSOCIATE-LABEL: define void @innermost_loop_3d_fast_reassociated_different
+; LICM_AFTER_REASSOCIATE-SAME: (i32 [[I:%.*]], double [[D1:%.*]], double [[D2:%.*]], double [[DELTA:%.*]], ptr [[CELLS:%.*]]) {
+; LICM_AFTER_REASSOCIATE-NEXT:  entry:
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND:%.*]]
+; LICM_AFTER_REASSOCIATE:       for.cond:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[J:%.*]] = phi i32 [ 0, [[ENTRY:%.*]] ], [ [[ADD_J_1:%.*]], [[FOR_BODY:%.*]] ]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CMP_NOT:%.*]] = icmp sgt i32 [[J]], [[I]]
+; LICM_AFTER_REASSOCIATE-NEXT:    br i1 [[CMP_NOT]], label [[FOR_END:%.*]], label [[FOR_BODY]]
+; LICM_AFTER_REASSOCIATE:       for.body:
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ADD_J_1]] = add nuw nsw i32 [[J]], 1
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J_1:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J_1:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_1:%.*]] = load double, ptr [[ARRAYIDX_J_1]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J_2:%.*]] = zext i32 [[ADD_J_1]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J_2:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J_2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_2:%.*]] = load double, ptr [[ARRAYIDX_J_2]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_3:%.*]] = load double, ptr [[ARRAYIDX_J_2]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[IDXPROM_J:%.*]] = zext i32 [[J]] to i64
+; LICM_AFTER_REASSOCIATE-NEXT:    [[ARRAYIDX_J:%.*]] = getelementptr inbounds double, ptr [[CELLS]], i64 [[IDXPROM_J]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[CELL_4:%.*]] = load double, ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_1:%.*]] = fmul fast double [[CELL_1]], [[D1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[FMUL_2:%.*]] = fmul fast double [[CELL_4]], [[D2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[EXTRA_MUL:%.*]] = fmul fast double [[CELL_3]], [[CELL_2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[REASS_ADD:%.*]] = fadd fast double [[EXTRA_MUL]], [[FMUL_1]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[EXTRA_ADD:%.*]] = fadd fast double [[REASS_ADD]], [[FMUL_2]]
+; LICM_AFTER_REASSOCIATE-NEXT:    [[REASS_MUL:%.*]] = fmul fast double [[EXTRA_ADD]], [[DELTA]]
+; LICM_AFTER_REASSOCIATE-NEXT:    store double [[REASS_MUL]], ptr [[ARRAYIDX_J]], align 8
+; LICM_AFTER_REASSOCIATE-NEXT:    br label [[FOR_COND]]
+; LICM_AFTER_REASSOCIATE:       for.end:
+; LICM_AFTER_REASSOCIATE-NEXT:    ret void
+;
+entry:
+  br label %for.cond
+
+for.cond:
+  %j = phi i32 [ 0, %entry ], [ %add.j.1, %for.body ]
+  %cmp.not = icmp sgt i32 %j, %i
+  br i1 %cmp.not, label %for.end, label %for.body
+
+for.body:
+  %add.j.1 = add nuw nsw i32 %j, 1
+  %idxprom.j.1 = zext i32 %add.j.1 to i64
+  %arrayidx.j.1 = getelementptr inbounds double, ptr %cells, i64 %idxprom.j.1
+  %cell.1 = load double, ptr %arrayidx.j.1, align 8
+  %add.j.2 = add nuw nsw i32 %j, 2
+  %idxprom.j.2 = zext i32 %add.j.1 to i64
+  %arrayidx.j.2 = getelementptr inbounds double, ptr %cells, i64 %idxprom.j.2
+  %cell.2 = load double, ptr %arrayidx.j.2, align 8
+  %add.j.3 = add nuw nsw i32 %j, 3
+  %idxprom.j.3 = zext i32 %add.j.1 to i64
+  %arrayidx.j.3 = getelementptr inbounds double, ptr %cells, i64 %idxprom.j.3
+  %cell.3 = load double, ptr %arrayidx.j.2, align 8
+  %idxprom.j = zext i32 %j to i64
+  %arrayidx.j = getelementptr inbounds double, ptr %cells, i64 %idxprom.j
+  %cell.4 = load double, ptr %arrayidx.j, align 8
+  %fmul.1 = fmul fast double %cell.1, %d1
+  %fmul.2 = fmul fast double %cell.4, %d2
+  %reass.add = fadd fast double %fmul.2, %fmul.1
+  %extra.mul = fmul fast double %cell.3, %cell.2
+  %extra.add = fadd fast double %extra.mul, %reass.add
+  %reass.mul = fmul fast double %extra.add, %delta
+  store double %reass.mul, ptr %arrayidx.j, align 8
+  br label %for.cond
+
+for.end:
+  ret void
+}