Revert rG25a97c3a43d7 - "[InstCombine] visitCallInst - retain undefs in vector funnel...

This reverts commit 25a97c3a43d7bc469ec67dd4e901a507b9b11116.

We have other constant folds that fold undef funnel shift amounts to 0 - so we need to be consistent.

If we end up with regressions where we lose a splat shift amount pattern we'll have to investigate other canonicalizations, but matchFunnelShift currently protects us from that.

diff --git a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
index 7416516..c069657 100644 (file)
--- a/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
+++ b/llvm/lib/Transforms/InstCombine/InstCombineCalls.cpp
@@ -885,15 +885,15 @@ Instruction *InstCombinerImpl::visitCallInst(CallInst &CI) {
     Constant *ShAmtC;
     if (match(II->getArgOperand(2), m_Constant(ShAmtC)) &&
         !isa<ConstantExpr>(ShAmtC) && !ShAmtC->containsConstantExpression()) {
-      Constant *WidthC = ConstantInt::get(Ty, BitWidth);
-
       // Canonicalize a shift amount constant operand to modulo the bit-width.
-      if (!match(ShAmtC, m_SpecificInt_ICMP(ICmpInst::ICMP_ULT,
-                                            APInt(BitWidth, BitWidth)))) {
-        Constant *ModuloC = ConstantExpr::getURem(ShAmtC, WidthC);
-        ModuloC = Constant::mergeUndefsWith(ModuloC, ShAmtC);
+      Constant *WidthC = ConstantInt::get(Ty, BitWidth);
+      Constant *ModuloC = ConstantExpr::getURem(ShAmtC, WidthC);
+      if (ModuloC != ShAmtC)
         return replaceOperand(*II, 2, ModuloC);
-      }
+
+      assert(ConstantExpr::getICmp(ICmpInst::ICMP_UGT, WidthC, ShAmtC) ==
+                 ConstantInt::getTrue(CmpInst::makeCmpResultType(Ty)) &&
+             "Shift amount expected to be modulo bitwidth");
 
       // Canonicalize funnel shift right by constant to funnel shift left. This
       // is not entirely arbitrary. For historical reasons, the backend may

diff --git a/llvm/test/Transforms/InstCombine/funnel.ll b/llvm/test/Transforms/InstCombine/funnel.ll
index 7c60f15..5c94279 100644 (file)
--- a/llvm/test/Transforms/InstCombine/funnel.ll
+++ b/llvm/test/Transforms/InstCombine/funnel.ll
@@ -116,7 +116,7 @@ define <2 x i32> @fshr_v2i32_constant_nonsplat(<2 x i32> %x, <2 x i32> %y) {
 
 define <2 x i32> @fshr_v2i32_constant_nonsplat_undef0(<2 x i32> %x, <2 x i32> %y) {
 ; CHECK-LABEL: @fshr_v2i32_constant_nonsplat_undef0(
-; CHECK-NEXT:    [[R:%.*]] = call <2 x i32> @llvm.fshl.v2i32(<2 x i32> [[Y:%.*]], <2 x i32> [[X:%.*]], <2 x i32> <i32 undef, i32 13>)
+; CHECK-NEXT:    [[R:%.*]] = call <2 x i32> @llvm.fshl.v2i32(<2 x i32> [[Y:%.*]], <2 x i32> [[X:%.*]], <2 x i32> <i32 0, i32 13>)
 ; CHECK-NEXT:    ret <2 x i32> [[R]]
 ;
   %shr = lshr <2 x i32> %x, <i32 undef, i32 19>
@@ -127,7 +127,7 @@ define <2 x i32> @fshr_v2i32_constant_nonsplat_undef0(<2 x i32> %x, <2 x i32> %y
 
 define <2 x i32> @fshr_v2i32_constant_nonsplat_undef1(<2 x i32> %x, <2 x i32> %y) {
 ; CHECK-LABEL: @fshr_v2i32_constant_nonsplat_undef1(
-; CHECK-NEXT:    [[R:%.*]] = call <2 x i32> @llvm.fshl.v2i32(<2 x i32> [[Y:%.*]], <2 x i32> [[X:%.*]], <2 x i32> <i32 15, i32 undef>)
+; CHECK-NEXT:    [[R:%.*]] = call <2 x i32> @llvm.fshl.v2i32(<2 x i32> [[Y:%.*]], <2 x i32> [[X:%.*]], <2 x i32> <i32 15, i32 0>)
 ; CHECK-NEXT:    ret <2 x i32> [[R]]
 ;
   %shr = lshr <2 x i32> %x, <i32 17, i32 19>
@@ -149,7 +149,7 @@ define <2 x i36> @fshl_v2i36_constant_nonsplat(<2 x i36> %x, <2 x i36> %y) {
 
 define <3 x i36> @fshl_v3i36_constant_nonsplat_undef0(<3 x i36> %x, <3 x i36> %y) {
 ; CHECK-LABEL: @fshl_v3i36_constant_nonsplat_undef0(
-; CHECK-NEXT:    [[R:%.*]] = call <3 x i36> @llvm.fshl.v3i36(<3 x i36> [[X:%.*]], <3 x i36> [[Y:%.*]], <3 x i36> <i36 21, i36 11, i36 undef>)
+; CHECK-NEXT:    [[R:%.*]] = call <3 x i36> @llvm.fshl.v3i36(<3 x i36> [[X:%.*]], <3 x i36> [[Y:%.*]], <3 x i36> <i36 21, i36 11, i36 0>)
 ; CHECK-NEXT:    ret <3 x i36> [[R]]
 ;
   %shl = shl <3 x i36> %x, <i36 21, i36 11, i36 undef>

diff --git a/llvm/test/Transforms/InstCombine/rotate.ll b/llvm/test/Transforms/InstCombine/rotate.ll
index c911f29..270d778 100644 (file)
--- a/llvm/test/Transforms/InstCombine/rotate.ll
+++ b/llvm/test/Transforms/InstCombine/rotate.ll
@@ -138,7 +138,7 @@ define <2 x i32> @rotr_v2i32_constant_nonsplat(<2 x i32> %x) {
 
 define <2 x i32> @rotr_v2i32_constant_nonsplat_undef0(<2 x i32> %x) {
 ; CHECK-LABEL: @rotr_v2i32_constant_nonsplat_undef0(
-; CHECK-NEXT:    [[R:%.*]] = call <2 x i32> @llvm.fshl.v2i32(<2 x i32> [[X:%.*]], <2 x i32> [[X]], <2 x i32> <i32 undef, i32 19>)
+; CHECK-NEXT:    [[R:%.*]] = call <2 x i32> @llvm.fshl.v2i32(<2 x i32> [[X:%.*]], <2 x i32> [[X]], <2 x i32> <i32 0, i32 19>)
 ; CHECK-NEXT:    ret <2 x i32> [[R]]
 ;
   %shl = shl <2 x i32> %x, <i32 undef, i32 19>
@@ -149,7 +149,7 @@ define <2 x i32> @rotr_v2i32_constant_nonsplat_undef0(<2 x i32> %x) {
 
 define <2 x i32> @rotr_v2i32_constant_nonsplat_undef1(<2 x i32> %x) {
 ; CHECK-LABEL: @rotr_v2i32_constant_nonsplat_undef1(
-; CHECK-NEXT:    [[R:%.*]] = call <2 x i32> @llvm.fshl.v2i32(<2 x i32> [[X:%.*]], <2 x i32> [[X]], <2 x i32> <i32 17, i32 undef>)
+; CHECK-NEXT:    [[R:%.*]] = call <2 x i32> @llvm.fshl.v2i32(<2 x i32> [[X:%.*]], <2 x i32> [[X]], <2 x i32> <i32 17, i32 0>)
 ; CHECK-NEXT:    ret <2 x i32> [[R]]
 ;
   %shl = shl <2 x i32> %x, <i32 17, i32 19>
@@ -171,7 +171,7 @@ define <2 x i36> @rotl_v2i36_constant_nonsplat(<2 x i36> %x) {
 
 define <3 x i36> @rotl_v3i36_constant_nonsplat_undef0(<3 x i36> %x) {
 ; CHECK-LABEL: @rotl_v3i36_constant_nonsplat_undef0(
-; CHECK-NEXT:    [[R:%.*]] = call <3 x i36> @llvm.fshl.v3i36(<3 x i36> [[X:%.*]], <3 x i36> [[X]], <3 x i36> <i36 21, i36 11, i36 undef>)
+; CHECK-NEXT:    [[R:%.*]] = call <3 x i36> @llvm.fshl.v3i36(<3 x i36> [[X:%.*]], <3 x i36> [[X]], <3 x i36> <i36 21, i36 11, i36 0>)
 ; CHECK-NEXT:    ret <3 x i36> [[R]]
 ;
   %shl = shl <3 x i36> %x, <i36 21, i36 11, i36 undef>