[DAG] Fold (sext (sext_inreg x)) -> (sext (trunc x)) if the trunc is free

[RKSimon]

No description provided.

[llvmbot]

@llvm/pr-subscribers-backend-x86

@llvm/pr-subscribers-llvm-selectiondag

Author: Simon Pilgrim (RKSimon)

Changes

---

Full diff: https://github.com/llvm/llvm-project/pull/77616.diff

3 Files Affected:

• (modified) llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp (+2-2)
• (modified) llvm/test/CodeGen/X86/abds.ll (+4-4)
• (modified) llvm/test/CodeGen/X86/pr14088.ll (+1-2)

diff --git a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
index 54732237c91aa6..ecdf9ab9e989f0 100644
--- a/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
+++ b/llvm/lib/CodeGen/SelectionDAG/DAGCombiner.cpp
@@ -13381,9 +13381,9 @@ SDValue DAGCombiner::visitSIGN_EXTEND(SDNode *N) {
   if (N0.getOpcode() == ISD::SIGN_EXTEND_INREG) {
     SDValue N00 = N0.getOperand(0);
     EVT ExtVT = cast<VTSDNode>(N0->getOperand(1))->getVT();
-    if (N00.getOpcode() == ISD::TRUNCATE &&
+    if ((N00.getOpcode() == ISD::TRUNCATE || TLI.isTruncateFree(N00, ExtVT)) &&
         (!LegalTypes || TLI.isTypeLegal(ExtVT))) {
-      SDValue T = DAG.getNode(ISD::TRUNCATE, DL, ExtVT, N00.getOperand(0));
+      SDValue T = DAG.getNode(ISD::TRUNCATE, DL, ExtVT, N00);
       return DAG.getNode(ISD::SIGN_EXTEND, DL, VT, T);
     }
   }
diff --git a/llvm/test/CodeGen/X86/abds.ll b/llvm/test/CodeGen/X86/abds.ll
index 39ac47e99e6e98..a80339427984a8 100644
--- a/llvm/test/CodeGen/X86/abds.ll
+++ b/llvm/test/CodeGen/X86/abds.ll
@@ -141,9 +141,9 @@ define i16 @abd_ext_i16_i32(i16 %a, i32 %b) nounwind {
 ;
 ; X64-LABEL: abd_ext_i16_i32:
 ; X64:       # %bb.0:
+; X64-NEXT:    # kill: def $edi killed $edi def $rdi
+; X64-NEXT:    movswq %di, %rcx
 ; X64-NEXT:    movslq %esi, %rax
-; X64-NEXT:    movswl %di, %ecx
-; X64-NEXT:    movslq %ecx, %rcx
 ; X64-NEXT:    subq %rax, %rcx
 ; X64-NEXT:    movq %rcx, %rax
 ; X64-NEXT:    negq %rax
@@ -232,9 +232,9 @@ define i32 @abd_ext_i32_i16(i32 %a, i16 %b) nounwind {
 ;
 ; X64-LABEL: abd_ext_i32_i16:
 ; X64:       # %bb.0:
+; X64-NEXT:    # kill: def $esi killed $esi def $rsi
+; X64-NEXT:    movswq %si, %rax
 ; X64-NEXT:    movslq %edi, %rcx
-; X64-NEXT:    movswl %si, %eax
-; X64-NEXT:    cltq
 ; X64-NEXT:    subq %rax, %rcx
 ; X64-NEXT:    movq %rcx, %rax
 ; X64-NEXT:    negq %rax
diff --git a/llvm/test/CodeGen/X86/pr14088.ll b/llvm/test/CodeGen/X86/pr14088.ll
index 18c73f00571ca8..83bf13280f94ae 100644
--- a/llvm/test/CodeGen/X86/pr14088.ll
+++ b/llvm/test/CodeGen/X86/pr14088.ll
@@ -26,8 +26,7 @@ define i32 @f(i1 %foo, ptr %tm_year2, ptr %bar, i16 %zed, i32 %zed2) {
 ; CHECK-NEXT:    imull $100, %ecx, %ecx
 ; CHECK-NEXT:    subl %ecx, %eax
 ; CHECK-NEXT:    movw %ax, (%rsi)
-; CHECK-NEXT:    cwtl
-; CHECK-NEXT:    cltq
+; CHECK-NEXT:    movswq %ax, %rax
 ; CHECK-NEXT:    imulq $1717986919, %rax, %rax # imm = 0x66666667
 ; CHECK-NEXT:    movq %rax, %rcx
 ; CHECK-NEXT:    shrq $63, %rcx

[phoebewang]

Test changes look great.

[arsenm]

slightly surprised this didn't fire in any amdgpu tests

[RKSimon]

slightly surprised this didn't fire in any amdgpu tests

The legality checks is limiting this in a number of places - I did look at relaxing it but it started to cause hangs.