[llvm] [NFC][SLP] Precommit tests for the phantom load saving and the cost context (PR #228901)
via llvm-commits
llvm-commits at lists.llvm.org
Sun Oct 4 08:13:13 PDT 2026
llvmorg-github-actions[bot] wrote:
<!--LLVM PR SUMMARY COMMENT-->
@llvm/pr-subscribers-backend-amdgpu
Author: Dmitry Sidorov (MrSidims)
<details>
<summary>Changes</summary>
<sub>Stack created with <a href="https://github.com/github/gh-stack">GitHub Stacks CLI</a> • <a href="https://gh.io/stacks-feedback">Give Feedback 💬</a></sub>
---
Patch is 187.70 KiB, truncated to 20.00 KiB below, full version: https://github.com/llvm/llvm-project/pull/228901.diff
7 Files Affected:
- (added) llvm/test/CodeGen/AMDGPU/slp-coalesced-loads.ll (+1536)
- (added) llvm/test/Transforms/SLPVectorizer/AMDGPU/elementwise-coalesced-loads.ll (+454)
- (added) llvm/test/Transforms/SLPVectorizer/AMDGPU/fdiv-afn-reduction.ll (+70)
- (added) llvm/test/Transforms/SLPVectorizer/AMDGPU/int-reduction-coalesced-loads.ll (+527)
- (added) llvm/test/Transforms/SLPVectorizer/AMDGPU/ordered-reduction-coalesced-loads.ll (+684)
- (added) llvm/test/Transforms/SLPVectorizer/AMDGPU/reassoc-reduction-coalesced-loads.ll (+159)
- (added) llvm/test/Transforms/SLPVectorizer/X86/udiv-strictfp.ll (+112)
``````````diff
diff --git a/llvm/test/CodeGen/AMDGPU/slp-coalesced-loads.ll b/llvm/test/CodeGen/AMDGPU/slp-coalesced-loads.ll
new file mode 100644
index 0000000000000..4ac9d6adf090e
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/slp-coalesced-loads.ll
@@ -0,0 +1,1536 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6
+; RUN: opt -mtriple=amdgpu9.0a-amd-amdhsa -passes=slp-vectorizer < %s | llc -mtriple=amdgpu9.0a-amd-amdhsa | FileCheck -check-prefix=GFX90A %s
+; RUN: opt -mtriple=amdgpu10.30-amd-amdhsa -passes=slp-vectorizer < %s | llc -mtriple=amdgpu10.30-amd-amdhsa | FileCheck -check-prefix=GFX1030 %s
+; RUN: opt -mtriple=amdgpu12.50-amd-amdhsa -passes=slp-vectorizer < %s | llc -mtriple=amdgpu12.50-amd-amdhsa | FileCheck -check-prefix=GFX1250 %s
+
+define float @dot8_ordered(ptr addrspace(1) noalias readonly align 4 %a, ptr addrspace(1) noalias readonly align 4 %b, float %acc) {
+; GFX90A-LABEL: dot8_ordered:
+; GFX90A: ; %bb.0: ; %entry
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx4 v[6:9], v[0:1], off
+; GFX90A-NEXT: global_load_dwordx4 v[10:13], v[0:1], off offset:16
+; GFX90A-NEXT: global_load_dwordx4 v[14:17], v[2:3], off offset:16
+; GFX90A-NEXT: global_load_dwordx4 v[18:21], v[2:3], off
+; GFX90A-NEXT: s_waitcnt vmcnt(1)
+; GFX90A-NEXT: v_pk_mul_f32 v[0:1], v[12:13], v[16:17]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mul_f32 v[6:7], v[6:7], v[18:19]
+; GFX90A-NEXT: v_add_f32_e32 v4, v4, v6
+; GFX90A-NEXT: v_pk_mul_f32 v[2:3], v[8:9], v[20:21]
+; GFX90A-NEXT: v_add_f32_e32 v4, v4, v7
+; GFX90A-NEXT: v_add_f32_e32 v2, v4, v2
+; GFX90A-NEXT: v_pk_mul_f32 v[8:9], v[10:11], v[14:15]
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v8
+; GFX90A-NEXT: v_add_f32_e32 v2, v2, v9
+; GFX90A-NEXT: v_add_f32_e32 v0, v2, v0
+; GFX90A-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1030-LABEL: dot8_ordered:
+; GFX1030: ; %bb.0: ; %entry
+; GFX1030-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX1030-NEXT: global_load_dwordx4 v[5:8], v[2:3], off
+; GFX1030-NEXT: s_clause 0x1
+; GFX1030-NEXT: global_load_dwordx4 v[9:12], v[0:1], off
+; GFX1030-NEXT: global_load_dwordx4 v[13:16], v[0:1], off offset:16
+; GFX1030-NEXT: global_load_dwordx4 v[17:20], v[2:3], off offset:16
+; GFX1030-NEXT: s_waitcnt vmcnt(2)
+; GFX1030-NEXT: v_fma_f32 v0, v9, v5, v4
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v10, v6
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v11, v7
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v12, v8
+; GFX1030-NEXT: s_waitcnt vmcnt(0)
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v13, v17
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v14, v18
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v15, v19
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v16, v20
+; GFX1030-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: dot8_ordered:
+; GFX1250: ; %bb.0: ; %entry
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_load_b128 v[6:9], v[2:3], off
+; GFX1250-NEXT: s_clause 0x1
+; GFX1250-NEXT: global_load_b128 v[10:13], v[0:1], off
+; GFX1250-NEXT: global_load_b128 v[14:17], v[0:1], off offset:16
+; GFX1250-NEXT: global_load_b128 v[18:21], v[2:3], off offset:16
+; GFX1250-NEXT: s_wait_loadcnt 0x2
+; GFX1250-NEXT: s_wait_xcnt 0x1
+; GFX1250-NEXT: v_pk_mul_f32 v[0:1], v[10:11], v[6:7]
+; GFX1250-NEXT: s_wait_xcnt 0x0
+; GFX1250-NEXT: v_pk_mul_f32 v[2:3], v[12:13], v[8:9]
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1250-NEXT: v_add_f32_e32 v0, v4, v0
+; GFX1250-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_2) | instid1(VALU_DEP_2)
+; GFX1250-NEXT: v_add_f32_e32 v2, v0, v2
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: v_pk_mul_f32 v[0:1], v[14:15], v[18:19]
+; GFX1250-NEXT: v_add_f32_e32 v2, v2, v3
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(SKIP_1) | instid1(VALU_DEP_2)
+; GFX1250-NEXT: v_add_f32_e32 v0, v2, v0
+; GFX1250-NEXT: v_pk_mul_f32 v[2:3], v[16:17], v[20:21]
+; GFX1250-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1250-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX1250-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+entry:
+ %pa0 = getelementptr inbounds float, ptr addrspace(1) %a, i64 0
+ %la0 = load float, ptr addrspace(1) %pa0, align 4
+ %pb0 = getelementptr inbounds float, ptr addrspace(1) %b, i64 0
+ %lb0 = load float, ptr addrspace(1) %pb0, align 4
+ %m0 = fmul contract float %la0, %lb0
+ %pa1 = getelementptr inbounds float, ptr addrspace(1) %a, i64 1
+ %la1 = load float, ptr addrspace(1) %pa1, align 4
+ %pb1 = getelementptr inbounds float, ptr addrspace(1) %b, i64 1
+ %lb1 = load float, ptr addrspace(1) %pb1, align 4
+ %m1 = fmul contract float %la1, %lb1
+ %pa2 = getelementptr inbounds float, ptr addrspace(1) %a, i64 2
+ %la2 = load float, ptr addrspace(1) %pa2, align 4
+ %pb2 = getelementptr inbounds float, ptr addrspace(1) %b, i64 2
+ %lb2 = load float, ptr addrspace(1) %pb2, align 4
+ %m2 = fmul contract float %la2, %lb2
+ %pa3 = getelementptr inbounds float, ptr addrspace(1) %a, i64 3
+ %la3 = load float, ptr addrspace(1) %pa3, align 4
+ %pb3 = getelementptr inbounds float, ptr addrspace(1) %b, i64 3
+ %lb3 = load float, ptr addrspace(1) %pb3, align 4
+ %m3 = fmul contract float %la3, %lb3
+ %pa4 = getelementptr inbounds float, ptr addrspace(1) %a, i64 4
+ %la4 = load float, ptr addrspace(1) %pa4, align 4
+ %pb4 = getelementptr inbounds float, ptr addrspace(1) %b, i64 4
+ %lb4 = load float, ptr addrspace(1) %pb4, align 4
+ %m4 = fmul contract float %la4, %lb4
+ %pa5 = getelementptr inbounds float, ptr addrspace(1) %a, i64 5
+ %la5 = load float, ptr addrspace(1) %pa5, align 4
+ %pb5 = getelementptr inbounds float, ptr addrspace(1) %b, i64 5
+ %lb5 = load float, ptr addrspace(1) %pb5, align 4
+ %m5 = fmul contract float %la5, %lb5
+ %pa6 = getelementptr inbounds float, ptr addrspace(1) %a, i64 6
+ %la6 = load float, ptr addrspace(1) %pa6, align 4
+ %pb6 = getelementptr inbounds float, ptr addrspace(1) %b, i64 6
+ %lb6 = load float, ptr addrspace(1) %pb6, align 4
+ %m6 = fmul contract float %la6, %lb6
+ %pa7 = getelementptr inbounds float, ptr addrspace(1) %a, i64 7
+ %la7 = load float, ptr addrspace(1) %pa7, align 4
+ %pb7 = getelementptr inbounds float, ptr addrspace(1) %b, i64 7
+ %lb7 = load float, ptr addrspace(1) %pb7, align 4
+ %m7 = fmul contract float %la7, %lb7
+ %s0 = fadd contract float %acc, %m0
+ %s1 = fadd contract float %s0, %m1
+ %s2 = fadd contract float %s1, %m2
+ %s3 = fadd contract float %s2, %m3
+ %s4 = fadd contract float %s3, %m4
+ %s5 = fadd contract float %s4, %m5
+ %s6 = fadd contract float %s5, %m6
+ %s7 = fadd contract float %s6, %m7
+ ret float %s7
+}
+
+define float @dot8_reassoc(ptr addrspace(1) noalias readonly align 4 %a, ptr addrspace(1) noalias readonly align 4 %b) {
+; GFX90A-LABEL: dot8_reassoc:
+; GFX90A: ; %bb.0: ; %entry
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx4 v[4:7], v[0:1], off
+; GFX90A-NEXT: global_load_dwordx4 v[8:11], v[0:1], off offset:16
+; GFX90A-NEXT: global_load_dwordx4 v[12:15], v[2:3], off offset:16
+; GFX90A-NEXT: global_load_dwordx4 v[16:19], v[2:3], off
+; GFX90A-NEXT: s_waitcnt vmcnt(1)
+; GFX90A-NEXT: v_pk_mul_f32 v[0:1], v[10:11], v[14:15]
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_fma_f32 v[0:1], v[6:7], v[18:19], v[0:1]
+; GFX90A-NEXT: v_pk_fma_f32 v[0:1], v[8:9], v[12:13], v[0:1]
+; GFX90A-NEXT: v_pk_fma_f32 v[0:1], v[4:5], v[16:17], v[0:1]
+; GFX90A-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1030-LABEL: dot8_reassoc:
+; GFX1030: ; %bb.0: ; %entry
+; GFX1030-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX1030-NEXT: global_load_dwordx4 v[4:7], v[2:3], off offset:16
+; GFX1030-NEXT: s_clause 0x1
+; GFX1030-NEXT: global_load_dwordx4 v[8:11], v[0:1], off offset:16
+; GFX1030-NEXT: global_load_dwordx4 v[12:15], v[0:1], off
+; GFX1030-NEXT: global_load_dwordx4 v[16:19], v[2:3], off
+; GFX1030-NEXT: s_waitcnt vmcnt(2)
+; GFX1030-NEXT: v_mul_f32_e32 v0, v11, v7
+; GFX1030-NEXT: s_waitcnt vmcnt(0)
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v15, v19
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v9, v5
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v13, v17
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v10, v6
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v14, v18
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v8, v4
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v12, v16
+; GFX1030-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: dot8_reassoc:
+; GFX1250: ; %bb.0: ; %entry
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_load_b128 v[4:7], v[2:3], off offset:16
+; GFX1250-NEXT: s_clause 0x1
+; GFX1250-NEXT: global_load_b128 v[8:11], v[0:1], off offset:16
+; GFX1250-NEXT: global_load_b128 v[12:15], v[0:1], off
+; GFX1250-NEXT: global_load_b128 v[16:19], v[2:3], off
+; GFX1250-NEXT: s_wait_loadcnt 0x2
+; GFX1250-NEXT: s_wait_xcnt 0x1
+; GFX1250-NEXT: v_pk_mul_f32 v[0:1], v[10:11], v[6:7]
+; GFX1250-NEXT: s_wait_xcnt 0x0
+; GFX1250-NEXT: v_pk_mul_f32 v[2:3], v[8:9], v[4:5]
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1250-NEXT: v_pk_fma_f32 v[0:1], v[14:15], v[18:19], v[0:1]
+; GFX1250-NEXT: v_pk_fma_f32 v[2:3], v[12:13], v[16:17], v[2:3]
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1250-NEXT: v_pk_add_f32 v[0:1], v[2:3], v[0:1]
+; GFX1250-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+entry:
+ %pa0 = getelementptr inbounds float, ptr addrspace(1) %a, i64 0
+ %la0 = load float, ptr addrspace(1) %pa0, align 4
+ %pb0 = getelementptr inbounds float, ptr addrspace(1) %b, i64 0
+ %lb0 = load float, ptr addrspace(1) %pb0, align 4
+ %m0 = fmul contract float %la0, %lb0
+ %pa1 = getelementptr inbounds float, ptr addrspace(1) %a, i64 1
+ %la1 = load float, ptr addrspace(1) %pa1, align 4
+ %pb1 = getelementptr inbounds float, ptr addrspace(1) %b, i64 1
+ %lb1 = load float, ptr addrspace(1) %pb1, align 4
+ %m1 = fmul contract float %la1, %lb1
+ %pa2 = getelementptr inbounds float, ptr addrspace(1) %a, i64 2
+ %la2 = load float, ptr addrspace(1) %pa2, align 4
+ %pb2 = getelementptr inbounds float, ptr addrspace(1) %b, i64 2
+ %lb2 = load float, ptr addrspace(1) %pb2, align 4
+ %m2 = fmul contract float %la2, %lb2
+ %pa3 = getelementptr inbounds float, ptr addrspace(1) %a, i64 3
+ %la3 = load float, ptr addrspace(1) %pa3, align 4
+ %pb3 = getelementptr inbounds float, ptr addrspace(1) %b, i64 3
+ %lb3 = load float, ptr addrspace(1) %pb3, align 4
+ %m3 = fmul contract float %la3, %lb3
+ %pa4 = getelementptr inbounds float, ptr addrspace(1) %a, i64 4
+ %la4 = load float, ptr addrspace(1) %pa4, align 4
+ %pb4 = getelementptr inbounds float, ptr addrspace(1) %b, i64 4
+ %lb4 = load float, ptr addrspace(1) %pb4, align 4
+ %m4 = fmul contract float %la4, %lb4
+ %pa5 = getelementptr inbounds float, ptr addrspace(1) %a, i64 5
+ %la5 = load float, ptr addrspace(1) %pa5, align 4
+ %pb5 = getelementptr inbounds float, ptr addrspace(1) %b, i64 5
+ %lb5 = load float, ptr addrspace(1) %pb5, align 4
+ %m5 = fmul contract float %la5, %lb5
+ %pa6 = getelementptr inbounds float, ptr addrspace(1) %a, i64 6
+ %la6 = load float, ptr addrspace(1) %pa6, align 4
+ %pb6 = getelementptr inbounds float, ptr addrspace(1) %b, i64 6
+ %lb6 = load float, ptr addrspace(1) %pb6, align 4
+ %m6 = fmul contract float %la6, %lb6
+ %pa7 = getelementptr inbounds float, ptr addrspace(1) %a, i64 7
+ %la7 = load float, ptr addrspace(1) %pa7, align 4
+ %pb7 = getelementptr inbounds float, ptr addrspace(1) %b, i64 7
+ %lb7 = load float, ptr addrspace(1) %pb7, align 4
+ %m7 = fmul contract float %la7, %lb7
+ %s1 = fadd fast float %m0, %m1
+ %s2 = fadd fast float %s1, %m2
+ %s3 = fadd fast float %s2, %m3
+ %s4 = fadd fast float %s3, %m4
+ %s5 = fadd fast float %s4, %m5
+ %s6 = fadd fast float %s5, %m6
+ %s7 = fadd fast float %s6, %m7
+ ret float %s7
+}
+
+define float @dot_i16_window(ptr addrspace(1) %a, ptr addrspace(1) %b, float %acc) {
+; GFX90A-LABEL: dot_i16_window:
+; GFX90A: ; %bb.0: ; %entry
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dwordx2 v[10:11], v[0:1], off
+; GFX90A-NEXT: global_load_dwordx4 v[6:9], v[2:3], off
+; GFX90A-NEXT: s_waitcnt vmcnt(1)
+; GFX90A-NEXT: v_cvt_f32_i32_sdwa v1, sext(v10) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
+; GFX90A-NEXT: v_cvt_f32_i32_sdwa v0, sext(v10) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
+; GFX90A-NEXT: v_cvt_f32_i32_sdwa v3, sext(v11) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
+; GFX90A-NEXT: v_cvt_f32_i32_sdwa v2, sext(v11) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_pk_mul_f32 v[0:1], v[0:1], v[6:7]
+; GFX90A-NEXT: v_add_f32_e32 v0, v4, v0
+; GFX90A-NEXT: v_pk_mul_f32 v[2:3], v[2:3], v[8:9]
+; GFX90A-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX90A-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX90A-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1030-LABEL: dot_i16_window:
+; GFX1030: ; %bb.0: ; %entry
+; GFX1030-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX1030-NEXT: global_load_dwordx2 v[5:6], v[0:1], off
+; GFX1030-NEXT: global_load_dwordx4 v[0:3], v[2:3], off
+; GFX1030-NEXT: s_waitcnt vmcnt(1)
+; GFX1030-NEXT: v_cvt_f32_i32_sdwa v7, sext(v5) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
+; GFX1030-NEXT: v_cvt_f32_i32_sdwa v5, sext(v5) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
+; GFX1030-NEXT: s_waitcnt vmcnt(0)
+; GFX1030-NEXT: v_fma_f32 v0, v7, v0, v4
+; GFX1030-NEXT: v_cvt_f32_i32_sdwa v4, sext(v6) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v5, v1
+; GFX1030-NEXT: v_cvt_f32_i32_sdwa v1, sext(v6) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_1
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v4, v2
+; GFX1030-NEXT: v_fmac_f32_e32 v0, v1, v3
+; GFX1030-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: dot_i16_window:
+; GFX1250: ; %bb.0: ; %entry
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_load_b64 v[10:11], v[0:1], off
+; GFX1250-NEXT: global_load_b128 v[6:9], v[2:3], off
+; GFX1250-NEXT: s_wait_loadcnt 0x1
+; GFX1250-NEXT: s_wait_xcnt 0x1
+; GFX1250-NEXT: v_ashrrev_i32_e32 v0, 16, v10
+; GFX1250-NEXT: s_wait_xcnt 0x0
+; GFX1250-NEXT: v_bfe_i32 v2, v10, 0, 16
+; GFX1250-NEXT: v_bfe_i32 v5, v11, 0, 16
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_3)
+; GFX1250-NEXT: v_cvt_f32_i32_e32 v1, v0
+; GFX1250-NEXT: v_cvt_f32_i32_e32 v0, v2
+; GFX1250-NEXT: v_ashrrev_i32_e32 v2, 16, v11
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1250-NEXT: v_pk_mul_f32 v[0:1], v[0:1], v[6:7]
+; GFX1250-NEXT: v_cvt_f32_i32_e32 v3, v2
+; GFX1250-NEXT: v_cvt_f32_i32_e32 v2, v5
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX1250-NEXT: v_add_f32_e32 v0, v4, v0
+; GFX1250-NEXT: v_pk_mul_f32 v[2:3], v[2:3], v[8:9]
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_1)
+; GFX1250-NEXT: v_add_f32_e32 v0, v0, v1
+; GFX1250-NEXT: v_add_f32_e32 v0, v0, v2
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX1250-NEXT: v_add_f32_e32 v0, v0, v3
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+entry:
+ %a1 = getelementptr inbounds nuw i8, ptr addrspace(1) %a, i64 2
+ %a2 = getelementptr inbounds nuw i8, ptr addrspace(1) %a, i64 4
+ %a3 = getelementptr inbounds nuw i8, ptr addrspace(1) %a, i64 6
+ %b1 = getelementptr inbounds nuw i8, ptr addrspace(1) %b, i64 4
+ %b2 = getelementptr inbounds nuw i8, ptr addrspace(1) %b, i64 8
+ %b3 = getelementptr inbounds nuw i8, ptr addrspace(1) %b, i64 12
+ %i0 = load i16, ptr addrspace(1) %a, align 2
+ %i1 = load i16, ptr addrspace(1) %a1, align 2
+ %i2 = load i16, ptr addrspace(1) %a2, align 2
+ %i3 = load i16, ptr addrspace(1) %a3, align 2
+ %f0 = sitofp i16 %i0 to float
+ %f1 = sitofp i16 %i1 to float
+ %f2 = sitofp i16 %i2 to float
+ %f3 = sitofp i16 %i3 to float
+ %w0 = load float, ptr addrspace(1) %b, align 4
+ %w1 = load float, ptr addrspace(1) %b1, align 4
+ %w2 = load float, ptr addrspace(1) %b2, align 4
+ %w3 = load float, ptr addrspace(1) %b3, align 4
+ %m0 = fmul contract float %f0, %w0
+ %m1 = fmul contract float %f1, %w1
+ %m2 = fmul contract float %f2, %w2
+ %m3 = fmul contract float %f3, %w3
+ %s0 = fadd contract float %acc, %m0
+ %s1 = fadd contract float %s0, %m1
+ %s2 = fadd contract float %s1, %m2
+ %s3 = fadd contract float %s2, %m3
+ ret float %s3
+}
+
+define void @store4_sitofp(ptr addrspace(1) %in, ptr addrspace(1) %out) {
+; GFX90A-LABEL: store4_sitofp:
+; GFX90A: ; %bb.0:
+; GFX90A-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX90A-NEXT: global_load_dword v0, v[0:1], off
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: v_ashrrev_i16_e32 v1, 8, v0
+; GFX90A-NEXT: v_cvt_f32_i32_sdwa v7, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3
+; GFX90A-NEXT: v_cvt_f32_i32_sdwa v6, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2
+; GFX90A-NEXT: v_cvt_f32_i32_sdwa v4, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0
+; GFX90A-NEXT: v_cvt_f32_i32_sdwa v5, sext(v1) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:WORD_0
+; GFX90A-NEXT: global_store_dwordx4 v[2:3], v[4:7], off
+; GFX90A-NEXT: s_waitcnt vmcnt(0)
+; GFX90A-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1030-LABEL: store4_sitofp:
+; GFX1030: ; %bb.0:
+; GFX1030-NEXT: s_waitcnt vmcnt(0) expcnt(0) lgkmcnt(0)
+; GFX1030-NEXT: global_load_dword v0, v[0:1], off
+; GFX1030-NEXT: s_waitcnt vmcnt(0)
+; GFX1030-NEXT: v_cvt_f32_i32_sdwa v4, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_0
+; GFX1030-NEXT: v_cvt_f32_i32_sdwa v5, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_1
+; GFX1030-NEXT: v_cvt_f32_i32_sdwa v6, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_2
+; GFX1030-NEXT: v_cvt_f32_i32_sdwa v7, sext(v0) dst_sel:DWORD dst_unused:UNUSED_PAD src0_sel:BYTE_3
+; GFX1030-NEXT: global_store_dwordx4 v[2:3], v[4:7], off
+; GFX1030-NEXT: s_setpc_b64 s[30:31]
+;
+; GFX1250-LABEL: store4_sitofp:
+; GFX1250: ; %bb.0:
+; GFX1250-NEXT: s_wait_loadcnt_dscnt 0x0
+; GFX1250-NEXT: s_wait_kmcnt 0x0
+; GFX1250-NEXT: global_load_b32 v0, v[0:1], off
+; GFX1250-NEXT: s_wait_loadcnt 0x0
+; GFX1250-NEXT: s_wait_xcnt 0x0
+; GFX1250-NEXT: v_bfe_i32 v1, v0, 0, 8
+; GFX1250-NEXT: v_bfe_i32 v5, v0, 8, 8
+; GFX1250-NEXT: v_bfe_i32 v6, v0, 16, 8
+; GFX1250-NEXT: v_ashrrev_i32_e32 v0, 24, v0
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX1250-NEXT: v_cvt_f32_i32_e32 v4, v1
+; GFX1250-NEXT: v_cvt_f32_i32_e32 v5, v5
+; GFX1250-NEXT: s_delay_alu instid0(VALU_DEP_4) | instskip(NEXT) | instid1(VALU_DEP_4)
+; GFX1250-NEXT: v_cvt_f32_i32_e32 v6, v6
+; GFX1250-NEXT: v_cvt_f32_i32_e32 v7, v0
+; GFX1250-NEXT: global_store_b128 v[2:3], v[4:7], off
+; GFX1250-NEXT: s_set_pc_i64 s[30:31]
+ %p1 = getelementptr inbounds i8, ptr addrspace(1) %in, i64 1
+ %p2 = getelementptr inbounds i8, ptr addrspace(1) %in, i64 2
+ %p3 = getelementptr inbounds i8, ptr addrspace(1) %in, i64 3
+ %q0 = load i8, ptr addrspace(1) %in, align 4
+ %q1 = load i8, ptr addrspace(1) %p1, align 1
+ %q2 = load i8, ptr addrspace(1) %p2, align 2
+ %q3 = load i8, ptr addrspace(1) %p3, align 1
+ %f0 ...
[truncated]
``````````
</details>
https://github.com/llvm/llvm-project/pull/228901
More information about the llvm-commits
mailing list