GFX13: Fix incorrect shrinking of V_ADD_CO_U32_e64 (#3361)

This instruction is VOP3-only so it should not be shrunk to
V_ADD_CO_U32_e32.

Fixes: SC1-6350

Author: jayfoad

llvm/lib/Target/AMDGPU/SIFoldOperands.cpp

@@ -877,6 +877,9 @@ bool SIFoldOperandsImpl::tryAddToFoldList(
       // Make sure to get the 32-bit version of the commuted opcode.
       unsigned MaybeCommutedOpc = MI->getOpcode();
       Op32 = AMDGPU::getVOPe32(MaybeCommutedOpc);
+
+      if (TII->pseudoToMCOpcode(Op32) == -1)
+        Op32 = -1;
     }
 
     appendFoldCandidate(FoldList, MI, CommuteOpNo, OpToFold, /*Commuted=*/true,

llvm/test/CodeGen/AMDGPU/literal64-symbol.ll

@@ -0,0 +1,66 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 3
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx1200 -mattr=wavefrontsize32 < %s | FileCheck %s --check-prefixes=GFX12-W32
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx1200 -mattr=wavefrontsize64 < %s | FileCheck %s --check-prefixes=GFX12-W64
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx1250 -mattr=wavefrontsize32 < %s | FileCheck %s --check-prefixes=GFX1250
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx1300 -mattr=wavefrontsize32 < %s | FileCheck %s --check-prefixes=GFX13-W32
+; RUN: llc -mtriple=amdgcn--amdpal -mcpu=gfx1300 -mattr=wavefrontsize64 < %s | FileCheck %s --check-prefixes=GFX13-W64
+
+@sym = external addrspace(4) constant [2 x <3 x i32>]
+
+define amdgpu_gs ptr addrspace(4) @v_add_symbol(i64 %arg) {
+; GFX12-W32-LABEL: v_add_symbol:
+; GFX12-W32:       ; %bb.0:
+; GFX12-W32-NEXT:    v_lshlrev_b64_e32 v[0:1], 4, v[0:1]
+; GFX12-W32-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-W32-NEXT:    v_add_co_u32 v0, vcc_lo, sym@abs32@lo, v0
+; GFX12-W32-NEXT:    v_add_co_ci_u32_e64 v1, null, sym@abs32@hi, v1, vcc_lo
+; GFX12-W32-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-W32-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX12-W32-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX12-W32-NEXT:    ; return to shader part epilog
+;
+; GFX12-W64-LABEL: v_add_symbol:
+; GFX12-W64:       ; %bb.0:
+; GFX12-W64-NEXT:    v_lshlrev_b64_e32 v[0:1], 4, v[0:1]
+; GFX12-W64-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX12-W64-NEXT:    v_add_co_u32 v0, vcc, sym@abs32@lo, v0
+; GFX12-W64-NEXT:    v_add_co_ci_u32_e64 v1, null, sym@abs32@hi, v1, vcc
+; GFX12-W64-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-W64-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX12-W64-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX12-W64-NEXT:    ; return to shader part epilog
+;
+; GFX1250-LABEL: v_add_symbol:
+; GFX1250:       ; %bb.0:
+; GFX1250-NEXT:    v_lshl_add_u64 v[0:1], v[0:1], 4, sym@abs64
+; GFX1250-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1250-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX1250-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX1250-NEXT:    ; return to shader part epilog
+;
+; GFX13-W32-LABEL: v_add_symbol:
+; GFX13-W32:       ; %bb.0:
+; GFX13-W32-NEXT:    v_lshlrev_b64_e32 v[0:1], 4, v[0:1]
+; GFX13-W32-NEXT:    s_mov_b64 s[0:1], sym@abs64
+; GFX13-W32-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX13-W32-NEXT:    v_add_co_u32 v0, vcc_lo, v0, sym@abs64
+; GFX13-W32-NEXT:    v_add_co_ci_u32_e64 v1, null, s1, v1, vcc_lo
+; GFX13-W32-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX13-W32-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX13-W32-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX13-W32-NEXT:    ; return to shader part epilog
+;
+; GFX13-W64-LABEL: v_add_symbol:
+; GFX13-W64:       ; %bb.0:
+; GFX13-W64-NEXT:    v_lshlrev_b64_e32 v[0:1], 4, v[0:1]
+; GFX13-W64-NEXT:    s_mov_b64 s[0:1], sym@abs64
+; GFX13-W64-NEXT:    s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX13-W64-NEXT:    v_add_co_u32 v0, vcc, v0, sym@abs64
+; GFX13-W64-NEXT:    v_add_co_ci_u32_e64 v1, null, s1, v1, vcc
+; GFX13-W64-NEXT:    s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX13-W64-NEXT:    v_readfirstlane_b32 s0, v0
+; GFX13-W64-NEXT:    v_readfirstlane_b32 s1, v1
+; GFX13-W64-NEXT:    ; return to shader part epilog
+  %i = getelementptr [2 x <3 x i32>], ptr addrspace(4) @sym, i64 0, i64 %arg
+  ret ptr addrspace(4) %i
+}