[llvm] [llvm][InstCombine] Fold signum(x) into scmp(x, 0) (PR #143445)

[Yash Solanki via llvm-commits]

================
@@ -0,0 +1,254 @@
+; NOTE: Assertions have been autogenerated by utils/update_test_checks.py UTC_ARGS: --version 5
+; RUN: opt -passes=instcombine -S < %s | FileCheck %s
+
+; Fold (x < y) ? -1 : zext(x != y) to scmp(x, y)
+define i32 @scmp_x_y_from_lt(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @scmp_x_y_from_lt(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[SEL:%.*]] = call i32 @llvm.scmp.i32.i32(i32 [[X]], i32 [[Y]])
+; CHECK-NEXT:    ret i32 [[SEL]]
+;
+  %cmp1 = icmp ne i32 %x, %y
+  %zext = zext i1 %cmp1 to i32
+  %cmp2 = icmp slt i32 %x, %y
+  %sel = select i1 %cmp2, i32 -1, i32 %zext
+  ret i32 %sel
+}
+
+; Fold (x < y) ? -1 : zext(x > y) to scmp(x, y)
+define i32 @scmp_x_y_from_lt_gt(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @scmp_x_y_from_lt_gt(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[SEL:%.*]] = call i32 @llvm.scmp.i32.i32(i32 [[X]], i32 [[Y]])
+; CHECK-NEXT:    ret i32 [[SEL]]
+;
+  %cmp1 = icmp sgt i32 %x, %y
+  %zext = zext i1 %cmp1 to i32
+  %cmp2 = icmp slt i32 %x, %y
+  %sel = select i1 %cmp2, i32 -1, i32 %zext
+  ret i32 %sel
+}
+
+; Fold (x > y) ? 1 : sext(x != y) to scmp(x, y)
+define i32 @scmp_x_y_from_gt_sext_ne(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @scmp_x_y_from_gt_sext_ne(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[SEL:%.*]] = call i32 @llvm.scmp.i32.i32(i32 [[X]], i32 [[Y]])
+; CHECK-NEXT:    ret i32 [[SEL]]
+;
+  %cmp1 = icmp ne i32 %x, %y
+  %sext = sext i1 %cmp1 to i32
+  %cmp2 = icmp sgt i32 %x, %y
+  %sel = select i1 %cmp2, i32 1, i32 %sext
+  ret i32 %sel
+}
+
+; Fold (x > y) ? 1 : sext(x < y) to scmp(x, y)
+define i32 @scmp_x_y_from_gt_sext_lt(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @scmp_x_y_from_gt_sext_lt(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[SEL:%.*]] = call i32 @llvm.scmp.i32.i32(i32 [[X]], i32 [[Y]])
+; CHECK-NEXT:    ret i32 [[SEL]]
+;
+  %cmp1 = icmp slt i32 %x, %y
+  %sext = sext i1 %cmp1 to i32
+  %cmp2 = icmp sgt i32 %x, %y
+  %sel = select i1 %cmp2, i32 1, i32 %sext
+  ret i32 %sel
+}
+
+; Fold (x <= y - 1) ? -1 : zext(x != y) to ucmp(x, y)
+define i32 @ucmp_x_y_from_le_sub_zext_ne(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @ucmp_x_y_from_le_sub_zext_ne(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[SEL:%.*]] = call i32 @llvm.ucmp.i32.i32(i32 [[X]], i32 [[Y]])
+; CHECK-NEXT:    ret i32 [[SEL]]
+;
+  %ym1 = sub i32 %y, 1
+  %cmp1 = icmp ule i32 %x, %ym1
+  %neq = icmp ne i32 %x, %y
+  %zext = zext i1 %neq to i32
+  %sel = select i1 %cmp1, i32 -1, i32 %zext
+  ret i32 %sel
+}
+
+; Fold (x <= y - 1) ? -1 : zext(x > y) to ucmp(x, y)
+define i32 @ucmp_x_y_from_le_sub_zext_gt(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @ucmp_x_y_from_le_sub_zext_gt(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[SEL:%.*]] = call i32 @llvm.ucmp.i32.i32(i32 [[X]], i32 [[Y]])
+; CHECK-NEXT:    ret i32 [[SEL]]
+;
+  %ym1 = sub i32 %y, 1
+  %cmp1 = icmp ule i32 %x, %ym1
+  %gt = icmp ugt i32 %x, %y
+  %zext = zext i1 %gt to i32
+  %sel = select i1 %cmp1, i32 -1, i32 %zext
+  ret i32 %sel
+}
+
+; Fold (x >= y + 1) ? 1 : sext(x != y) to ucmp(x, y)
+define i32 @ucmp_x_y_from_ge_add_sext_ne(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @ucmp_x_y_from_ge_add_sext_ne(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[SEL:%.*]] = call i32 @llvm.ucmp.i32.i32(i32 [[X]], i32 [[Y]])
+; CHECK-NEXT:    ret i32 [[SEL]]
+;
+  %yp1 = add i32 %y, 1
+  %cmp1 = icmp uge i32 %x, %yp1
+  %neq = icmp ne i32 %x, %y
+  %sext = sext i1 %neq to i32
+  %sel = select i1 %cmp1, i32 1, i32 %sext
+  ret i32 %sel
+}
+
+; Fold (x >= y + 1) ? 1 : sext(x < y) to ucmp(x, y)
+define i32 @ucmp_x_y_from_ge_add_sext_lt(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @ucmp_x_y_from_ge_add_sext_lt(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[SEL:%.*]] = call i32 @llvm.ucmp.i32.i32(i32 [[X]], i32 [[Y]])
+; CHECK-NEXT:    ret i32 [[SEL]]
+;
+  %yp1 = add i32 %y, 1
+  %cmp1 = icmp uge i32 %x, %yp1
+  %lt = icmp ult i32 %x, %y
+  %sext = sext i1 %lt to i32
+  %sel = select i1 %cmp1, i32 1, i32 %sext
+  ret i32 %sel
+}
+
+; Fold (x > y - 1) ? zext(x != y) : -1 to scmp(x, y)
+define i32 @scmp_x_y_from_gt_sub_zext_ne(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @scmp_x_y_from_gt_sub_zext_ne(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[SEL:%.*]] = call i32 @llvm.scmp.i32.i32(i32 [[X]], i32 [[Y]])
+; CHECK-NEXT:    ret i32 [[SEL]]
+;
+  %ym1 = sub i32 %y, 1
+  %cmp1 = icmp sgt i32 %x, %ym1
+  %ne = icmp ne i32 %x, %y
+  %zext = zext i1 %ne to i32
+  %sel = select i1 %cmp1, i32 %zext, i32 -1
+  ret i32 %sel
+}
+
+; Fold (x > y - 1) ? zext(x > y) : -1 to ucmp(x, y)
+define i32 @ucmp_x_y_from_gt_sub_zext_gt(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @ucmp_x_y_from_gt_sub_zext_gt(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[SEL:%.*]] = call i32 @llvm.ucmp.i32.i32(i32 [[X]], i32 [[Y]])
+; CHECK-NEXT:    ret i32 [[SEL]]
+;
+  %ym1 = sub i32 %y, 1
+  %cmp1 = icmp ugt i32 %x, %ym1
+  %gt = icmp ugt i32 %x, %y
+  %zext = zext i1 %gt to i32
+  %sel = select i1 %cmp1, i32 %zext, i32 -1
+  ret i32 %sel
+}
+
+; Fold (x < y + 1) ? sext(x != y) : 1 to scmp(x, y)
+define i32 @scmp_x_y_from_lt_add_sext_ne(i32 %x, i32 %y) {
+; CHECK-LABEL: define i32 @scmp_x_y_from_lt_add_sext_ne(
+; CHECK-SAME: i32 [[X:%.*]], i32 [[Y:%.*]]) {
+; CHECK-NEXT:    [[SEL:%.*]] = call i32 @llvm.scmp.i32.i32(i32 [[X]], i32 [[Y]])
+; CHECK-NEXT:    ret i32 [[SEL]]
+;
+  %yp1 = add i32 %y, 1
+  %cmp1 = icmp slt i32 %x, %yp1
+  %ne = icmp ne i32 %x, %y
+  %sext = sext i1 %ne to i32
+  %sel = select i1 %cmp1, i32 %sext, i32 1
+  ret i32 %sel
+}
+
+; Fold (x < y + 1) ? sext(x < y) : 1 to ucmp(x, y)
+define i32 @ucmp_x_y_from_lt_add_sext_lt(i32 %x, i32 %y) {
----------------
yashnator wrote:

I have used `llvm::APInt` and checked overflow with its functions. Can you see if this seems fine now? If not, I'll put the original commit back with checks on the RHS.  Thanks for your patience!

https://github.com/llvm/llvm-project/pull/143445