[llvm] [AMDGPU][GlobalISel] Add RegBankLegalize support for G_FPEXT (PR #171483)
via llvm-commits
llvm-commits at lists.llvm.org
Tue Dec 9 09:59:46 PST 2025
https://github.com/vangthao95 created https://github.com/llvm/llvm-project/pull/171483
None
>From b3fd56606e39137b52dd363617c9f0e2c51242d5 Mon Sep 17 00:00:00 2001
From: Vang Thao <vthao at amd.com>
Date: Tue, 9 Dec 2025 09:25:34 -0800
Subject: [PATCH] [AMDGPU][GlobalISel] Add RegBankLegalize support for G_FPEXT
---
.../AMDGPU/AMDGPURegBankLegalizeRules.cpp | 7 +
llvm/test/CodeGen/AMDGPU/GlobalISel/fpext.ll | 263 ++++++++++++++++++
.../AMDGPU/GlobalISel/regbankselect-fpext.mir | 8 +-
3 files changed, 276 insertions(+), 2 deletions(-)
create mode 100644 llvm/test/CodeGen/AMDGPU/GlobalISel/fpext.ll
diff --git a/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp b/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp
index 9de309279a247..373b35e75f404 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPURegBankLegalizeRules.cpp
@@ -976,6 +976,13 @@ RegBankLegalizeRules::RegBankLegalizeRules(const GCNSubtarget &_ST,
.Any({{UniS32, S32}, {{Sgpr32}, {Sgpr32}}}, hasSALUFloat)
.Any({{UniS32, S32}, {{UniInVgprS32}, {Vgpr32}}}, !hasSALUFloat);
+ addRulesForGOpcs({G_FPEXT})
+ .Any({{UniS32, S16}, {{Sgpr32}, {Sgpr16}}}, hasSALUFloat)
+ .Any({{UniS32, S16}, {{UniInVgprS32}, {Vgpr16}}}, !hasSALUFloat)
+ .Any({{DivS32, S16}, {{Vgpr32}, {Vgpr16}}})
+ .Any({{UniS64, S32}, {{UniInVgprS64}, {Vgpr32}}})
+ .Any({{DivS64, S32}, {{Vgpr64}, {Vgpr32}}});
+
addRulesForGOpcs({G_IS_FPCLASS})
.Any({{DivS1, S16}, {{Vcc}, {Vgpr16}}})
.Any({{UniS1, S16}, {{UniInVcc}, {Vgpr16}}})
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/fpext.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/fpext.ll
new file mode 100644
index 0000000000000..990153ebce243
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/fpext.ll
@@ -0,0 +1,263 @@
+; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py
+; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=-real-true16 -mcpu=gfx1100 -o - %s | FileCheck -check-prefixes=GFX11,GFX11-FAKE16 %s
+; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=+real-true16 -mcpu=gfx1100 -o - %s | FileCheck -check-prefixes=GFX11,GFX11-TRUE16 %s
+; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=-real-true16 -mcpu=gfx1200 -o - %s | FileCheck -check-prefixes=GFX12,GFX12-FAKE16 %s
+; RUN: llc -global-isel -new-reg-bank-select -mtriple=amdgcn-amd-amdpal -mattr=+real-true16 -mcpu=gfx1200 -o - %s | FileCheck -check-prefixes=GFX12,GFX12-TRUE16 %s
+
+define amdgpu_ps float @fpext_f16_to_f32_uniform(half inreg %a) {
+; GFX11-LABEL: fpext_f16_to_f32_uniform:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: v_cvt_f32_f16_e32 v0, s0
+; GFX11-NEXT: ; return to shader part epilog
+;
+; GFX12-LABEL: fpext_f16_to_f32_uniform:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_cvt_f32_f16 s0, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_3)
+; GFX12-NEXT: v_mov_b32_e32 v0, s0
+; GFX12-NEXT: ; return to shader part epilog
+ %result = fpext half %a to float
+ ret float %result
+}
+
+define amdgpu_ps float @fpext_f16_to_f32_div(half %a) {
+; GFX11-FAKE16-LABEL: fpext_f16_to_f32_div:
+; GFX11-FAKE16: ; %bb.0:
+; GFX11-FAKE16-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX11-FAKE16-NEXT: ; return to shader part epilog
+;
+; GFX11-TRUE16-LABEL: fpext_f16_to_f32_div:
+; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: v_cvt_f32_f16_e32 v0, v0.l
+; GFX11-TRUE16-NEXT: ; return to shader part epilog
+;
+; GFX12-FAKE16-LABEL: fpext_f16_to_f32_div:
+; GFX12-FAKE16: ; %bb.0:
+; GFX12-FAKE16-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX12-FAKE16-NEXT: ; return to shader part epilog
+;
+; GFX12-TRUE16-LABEL: fpext_f16_to_f32_div:
+; GFX12-TRUE16: ; %bb.0:
+; GFX12-TRUE16-NEXT: v_cvt_f32_f16_e32 v0, v0.l
+; GFX12-TRUE16-NEXT: ; return to shader part epilog
+ %result = fpext half %a to float
+ ret float %result
+}
+
+define amdgpu_ps void @fpext_f32_to_f64_uniform(float inreg %a, ptr addrspace(1) %ptr) {
+; GFX11-LABEL: fpext_f32_to_f64_uniform:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: v_cvt_f64_f32_e32 v[2:3], s0
+; GFX11-NEXT: global_store_b64 v[0:1], v[2:3], off
+; GFX11-NEXT: s_endpgm
+;
+; GFX12-LABEL: fpext_f32_to_f64_uniform:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: v_cvt_f64_f32_e32 v[2:3], s0
+; GFX12-NEXT: global_store_b64 v[0:1], v[2:3], off
+; GFX12-NEXT: s_endpgm
+ %result = fpext float %a to double
+ store double %result, ptr addrspace(1) %ptr
+ ret void
+}
+
+define amdgpu_ps void @fpext_f32_to_f64_div(float %a, ptr addrspace(1) %ptr) {
+; GFX11-LABEL: fpext_f32_to_f64_div:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: v_cvt_f64_f32_e32 v[3:4], v0
+; GFX11-NEXT: global_store_b64 v[1:2], v[3:4], off
+; GFX11-NEXT: s_endpgm
+;
+; GFX12-LABEL: fpext_f32_to_f64_div:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: v_cvt_f64_f32_e32 v[3:4], v0
+; GFX12-NEXT: global_store_b64 v[1:2], v[3:4], off
+; GFX12-NEXT: s_endpgm
+ %result = fpext float %a to double
+ store double %result, ptr addrspace(1) %ptr
+ ret void
+}
+
+define amdgpu_ps void @fpext_f16_to_f64_uniform(half inreg %a, ptr addrspace(1) %ptr) {
+; GFX11-LABEL: fpext_f16_to_f64_uniform:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: v_cvt_f32_f16_e32 v2, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-NEXT: v_cvt_f64_f32_e32 v[2:3], v2
+; GFX11-NEXT: global_store_b64 v[0:1], v[2:3], off
+; GFX11-NEXT: s_endpgm
+;
+; GFX12-LABEL: fpext_f16_to_f64_uniform:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: s_cvt_f32_f16 s0, s0
+; GFX12-NEXT: s_delay_alu instid0(SALU_CYCLE_3)
+; GFX12-NEXT: v_cvt_f64_f32_e32 v[2:3], s0
+; GFX12-NEXT: global_store_b64 v[0:1], v[2:3], off
+; GFX12-NEXT: s_endpgm
+ %result = fpext half %a to double
+ store double %result, ptr addrspace(1) %ptr
+ ret void
+}
+
+define amdgpu_ps void @fpext_f16_to_f64_div(half %a, ptr addrspace(1) %ptr) {
+; GFX11-FAKE16-LABEL: fpext_f16_to_f64_div:
+; GFX11-FAKE16: ; %bb.0:
+; GFX11-FAKE16-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-FAKE16-NEXT: v_cvt_f64_f32_e32 v[3:4], v0
+; GFX11-FAKE16-NEXT: global_store_b64 v[1:2], v[3:4], off
+; GFX11-FAKE16-NEXT: s_endpgm
+;
+; GFX11-TRUE16-LABEL: fpext_f16_to_f64_div:
+; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: v_cvt_f32_f16_e32 v0, v0.l
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_cvt_f64_f32_e32 v[3:4], v0
+; GFX11-TRUE16-NEXT: global_store_b64 v[1:2], v[3:4], off
+; GFX11-TRUE16-NEXT: s_endpgm
+;
+; GFX12-FAKE16-LABEL: fpext_f16_to_f64_div:
+; GFX12-FAKE16: ; %bb.0:
+; GFX12-FAKE16-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-FAKE16-NEXT: v_cvt_f64_f32_e32 v[3:4], v0
+; GFX12-FAKE16-NEXT: global_store_b64 v[1:2], v[3:4], off
+; GFX12-FAKE16-NEXT: s_endpgm
+;
+; GFX12-TRUE16-LABEL: fpext_f16_to_f64_div:
+; GFX12-TRUE16: ; %bb.0:
+; GFX12-TRUE16-NEXT: v_cvt_f32_f16_e32 v0, v0.l
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX12-TRUE16-NEXT: v_cvt_f64_f32_e32 v[3:4], v0
+; GFX12-TRUE16-NEXT: global_store_b64 v[1:2], v[3:4], off
+; GFX12-TRUE16-NEXT: s_endpgm
+ %result = fpext half %a to double
+ store double %result, ptr addrspace(1) %ptr
+ ret void
+}
+
+define amdgpu_ps <2 x float> @fpext_v2f16_to_v2f32_uniform(<2 x half> inreg %a) {
+; GFX11-FAKE16-LABEL: fpext_v2f16_to_v2f32_uniform:
+; GFX11-FAKE16: ; %bb.0:
+; GFX11-FAKE16-NEXT: s_lshr_b32 s1, s0, 16
+; GFX11-FAKE16-NEXT: v_cvt_f32_f16_e32 v0, s0
+; GFX11-FAKE16-NEXT: v_cvt_f32_f16_e32 v1, s1
+; GFX11-FAKE16-NEXT: ; return to shader part epilog
+;
+; GFX11-TRUE16-LABEL: fpext_v2f16_to_v2f32_uniform:
+; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: v_cvt_f32_f16_e32 v0, s0
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_1)
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v1, v0
+; GFX11-TRUE16-NEXT: ; return to shader part epilog
+;
+; GFX12-FAKE16-LABEL: fpext_v2f16_to_v2f32_uniform:
+; GFX12-FAKE16: ; %bb.0:
+; GFX12-FAKE16-NEXT: s_cvt_f32_f16 s1, s0
+; GFX12-FAKE16-NEXT: s_cvt_hi_f32_f16 s0, s0
+; GFX12-FAKE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3)
+; GFX12-FAKE16-NEXT: v_dual_mov_b32 v0, s1 :: v_dual_mov_b32 v1, s0
+; GFX12-FAKE16-NEXT: ; return to shader part epilog
+;
+; GFX12-TRUE16-LABEL: fpext_v2f16_to_v2f32_uniform:
+; GFX12-TRUE16: ; %bb.0:
+; GFX12-TRUE16-NEXT: s_cvt_f32_f16 s0, s0
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(SALU_CYCLE_3)
+; GFX12-TRUE16-NEXT: v_dual_mov_b32 v0, s0 :: v_dual_mov_b32 v1, s0
+; GFX12-TRUE16-NEXT: ; return to shader part epilog
+ %result = fpext <2 x half> %a to <2 x float>
+ ret <2 x float> %result
+}
+
+define amdgpu_ps <2 x float> @fpext_v2f16_to_v2f32_div(<2 x half> %a) {
+; GFX11-FAKE16-LABEL: fpext_v2f16_to_v2f32_div:
+; GFX11-FAKE16: ; %bb.0:
+; GFX11-FAKE16-NEXT: v_lshrrev_b32_e32 v1, 16, v0
+; GFX11-FAKE16-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX11-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-FAKE16-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX11-FAKE16-NEXT: ; return to shader part epilog
+;
+; GFX11-TRUE16-LABEL: fpext_v2f16_to_v2f32_div:
+; GFX11-TRUE16: ; %bb.0:
+; GFX11-TRUE16-NEXT: v_cvt_f32_f16_e32 v2, v0.l
+; GFX11-TRUE16-NEXT: v_cvt_f32_f16_e32 v1, v0.h
+; GFX11-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-TRUE16-NEXT: v_mov_b32_e32 v0, v2
+; GFX11-TRUE16-NEXT: ; return to shader part epilog
+;
+; GFX12-FAKE16-LABEL: fpext_v2f16_to_v2f32_div:
+; GFX12-FAKE16: ; %bb.0:
+; GFX12-FAKE16-NEXT: v_lshrrev_b32_e32 v1, 16, v0
+; GFX12-FAKE16-NEXT: v_cvt_f32_f16_e32 v0, v0
+; GFX12-FAKE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-FAKE16-NEXT: v_cvt_f32_f16_e32 v1, v1
+; GFX12-FAKE16-NEXT: ; return to shader part epilog
+;
+; GFX12-TRUE16-LABEL: fpext_v2f16_to_v2f32_div:
+; GFX12-TRUE16: ; %bb.0:
+; GFX12-TRUE16-NEXT: v_cvt_f32_f16_e32 v2, v0.l
+; GFX12-TRUE16-NEXT: v_cvt_f32_f16_e32 v1, v0.h
+; GFX12-TRUE16-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-TRUE16-NEXT: v_mov_b32_e32 v0, v2
+; GFX12-TRUE16-NEXT: ; return to shader part epilog
+ %result = fpext <2 x half> %a to <2 x float>
+ ret <2 x float> %result
+}
+
+define amdgpu_ps void @fpext_v2f32_to_v2f64_uniform(<2 x float> inreg %a, ptr addrspace(1) %ptr) {
+; GFX11-LABEL: fpext_v2f32_to_v2f64_uniform:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: v_cvt_f64_f32_e32 v[4:5], s1
+; GFX11-NEXT: v_cvt_f64_f32_e32 v[2:3], s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX11-NEXT: v_readfirstlane_b32 s3, v5
+; GFX11-NEXT: v_readfirstlane_b32 s0, v2
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_1) | instid1(VALU_DEP_3)
+; GFX11-NEXT: v_readfirstlane_b32 s1, v3
+; GFX11-NEXT: v_readfirstlane_b32 s2, v4
+; GFX11-NEXT: v_dual_mov_b32 v5, s3 :: v_dual_mov_b32 v2, s0
+; GFX11-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX11-NEXT: v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v4, s2
+; GFX11-NEXT: global_store_b128 v[0:1], v[2:5], off
+; GFX11-NEXT: s_endpgm
+;
+; GFX12-LABEL: fpext_v2f32_to_v2f64_uniform:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: v_cvt_f64_f32_e32 v[4:5], s1
+; GFX12-NEXT: v_cvt_f64_f32_e32 v[2:3], s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2) | instskip(NEXT) | instid1(VALU_DEP_2)
+; GFX12-NEXT: v_readfirstlane_b32 s3, v5
+; GFX12-NEXT: v_readfirstlane_b32 s0, v2
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_3) | instskip(SKIP_2) | instid1(VALU_DEP_3)
+; GFX12-NEXT: v_readfirstlane_b32 s1, v3
+; GFX12-NEXT: v_readfirstlane_b32 s2, v4
+; GFX12-NEXT: s_wait_alu depctr_va_sdst(0)
+; GFX12-NEXT: v_dual_mov_b32 v5, s3 :: v_dual_mov_b32 v2, s0
+; GFX12-NEXT: s_delay_alu instid0(VALU_DEP_2)
+; GFX12-NEXT: v_dual_mov_b32 v3, s1 :: v_dual_mov_b32 v4, s2
+; GFX12-NEXT: global_store_b128 v[0:1], v[2:5], off
+; GFX12-NEXT: s_endpgm
+ %result = fpext <2 x float> %a to <2 x double>
+ store <2 x double> %result, ptr addrspace(1) %ptr
+ ret void
+}
+
+define amdgpu_ps void @fpext_v2f32_to_v2f64_div(<2 x float> %a, ptr addrspace(1) %ptr) {
+; GFX11-LABEL: fpext_v2f32_to_v2f64_div:
+; GFX11: ; %bb.0:
+; GFX11-NEXT: v_cvt_f64_f32_e32 v[4:5], v0
+; GFX11-NEXT: v_cvt_f64_f32_e32 v[6:7], v1
+; GFX11-NEXT: global_store_b128 v[2:3], v[4:7], off
+; GFX11-NEXT: s_endpgm
+;
+; GFX12-LABEL: fpext_v2f32_to_v2f64_div:
+; GFX12: ; %bb.0:
+; GFX12-NEXT: v_cvt_f64_f32_e32 v[4:5], v0
+; GFX12-NEXT: v_cvt_f64_f32_e32 v[6:7], v1
+; GFX12-NEXT: global_store_b128 v[2:3], v[4:7], off
+; GFX12-NEXT: s_endpgm
+ %result = fpext <2 x float> %a to <2 x double>
+ store <2 x double> %result, ptr addrspace(1) %ptr
+ ret void
+}
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-fpext.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-fpext.mir
index df88b99d2ab9b..328f24a236557 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-fpext.mir
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/regbankselect-fpext.mir
@@ -1,6 +1,6 @@
# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py
-# RUN: llc -mtriple=amdgcn -mcpu=fiji -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-fast | FileCheck %s
-# RUN: llc -mtriple=amdgcn -mcpu=fiji -run-pass=regbankselect %s -verify-machineinstrs -o - -regbankselect-greedy | FileCheck %s
+# RUN: llc -mtriple=amdgcn -mcpu=fiji -run-pass="amdgpu-regbankselect,amdgpu-regbanklegalize" %s -verify-machineinstrs -o - -regbankselect-fast | FileCheck %s
+# RUN: llc -mtriple=amdgcn -mcpu=fiji -run-pass="amdgpu-regbankselect,amdgpu-regbanklegalize" %s -verify-machineinstrs -o - -regbankselect-greedy | FileCheck %s
---
name: fpext_s
@@ -15,6 +15,10 @@ body: |
; CHECK-NEXT: [[COPY:%[0-9]+]]:sgpr(s32) = COPY $sgpr0
; CHECK-NEXT: [[COPY1:%[0-9]+]]:vgpr(s32) = COPY [[COPY]](s32)
; CHECK-NEXT: [[FPEXT:%[0-9]+]]:vgpr(s64) = G_FPEXT [[COPY1]](s32)
+ ; CHECK-NEXT: [[UV:%[0-9]+]]:vgpr(s32), [[UV1:%[0-9]+]]:vgpr(s32) = G_UNMERGE_VALUES [[FPEXT]](s64)
+ ; CHECK-NEXT: [[AMDGPU_READANYLANE:%[0-9]+]]:sgpr(s32) = G_AMDGPU_READANYLANE [[UV]]
+ ; CHECK-NEXT: [[AMDGPU_READANYLANE1:%[0-9]+]]:sgpr(s32) = G_AMDGPU_READANYLANE [[UV1]]
+ ; CHECK-NEXT: [[MV:%[0-9]+]]:sgpr(s64) = G_MERGE_VALUES [[AMDGPU_READANYLANE]](s32), [[AMDGPU_READANYLANE1]](s32)
%0:_(s32) = COPY $sgpr0
%1:_(s64) = G_FPEXT %0
...
More information about the llvm-commits
mailing list