[llvm] [AMDGPU][GlobalISel] Align `selectVOP3PMadMixModsImpl` with the `SelectionDAG` counterpart (PR #110168)

Shilei Tian via llvm-commits llvm-commits at lists.llvm.org
Mon Sep 30 11:26:27 PDT 2024


https://github.com/shiltian updated https://github.com/llvm/llvm-project/pull/110168

>From 9f1f15ea6961ecb90cf868e0425035ae281e0d63 Mon Sep 17 00:00:00 2001
From: Shilei Tian <shilei.tian at amd.com>
Date: Thu, 26 Sep 2024 16:45:11 -0400
Subject: [PATCH] [AMDGPU][GlobalISel] Make `selectVOP3PMadMixModsImpl` same as
 the SelectionDAG counterpart

The current `selectVOP3PMadMixModsImpl` can produce `V_MAD_FIX_F32` instruction
that violates constant bus restriction, while its `SelectionDAG` counterpart
doesn't. The culprit is in the copy stripping while the SelectionDAG version
only has a bitcast stripping. This PR simply aligns the two version.
---
 .../AMDGPU/AMDGPUInstructionSelector.cpp      | 17 +++-----
 .../GlobalISel/combine-fma-add-ext-mul.ll     | 20 ++++++---
 .../CodeGen/AMDGPU/GlobalISel/fdiv.f16.ll     | 10 ++---
 .../fmamix-constant-bus-violation.mir         | 42 +++++++++++++++++++
 .../madmix-constant-bus-violation.mir         | 42 +++++++++++++++++++
 5 files changed, 109 insertions(+), 22 deletions(-)
 create mode 100644 llvm/test/CodeGen/AMDGPU/GlobalISel/fmamix-constant-bus-violation.mir
 create mode 100644 llvm/test/CodeGen/AMDGPU/GlobalISel/madmix-constant-bus-violation.mir

diff --git a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
index febf0711c7d574..c3df962d448f94 100644
--- a/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
+++ b/llvm/lib/Target/AMDGPU/AMDGPUInstructionSelector.cpp
@@ -5312,26 +5312,20 @@ AMDGPUInstructionSelector::selectVOP3PMadMixModsImpl(MachineOperand &Root,
     // Only change Src if src modifier could be gained. In such cases new Src
     // could be sgpr but this does not violate constant bus restriction for
     // instruction that is being selected.
-    // Note: Src is not changed when there is only a simple sgpr to vgpr copy
-    // since this could violate constant bus restriction.
-    Register PeekSrc = stripCopy(Src, *MRI);
+    Src = stripBitCast(Src, *MRI);
 
     const auto CheckAbsNeg = [&]() {
       // Be careful about folding modifiers if we already have an abs. fneg is
       // applied last, so we don't want to apply an earlier fneg.
       if ((Mods & SISrcMods::ABS) == 0) {
         unsigned ModsTmp;
-        std::tie(PeekSrc, ModsTmp) = selectVOP3ModsImpl(PeekSrc);
+        std::tie(Src, ModsTmp) = selectVOP3ModsImpl(Src);
 
-        if ((ModsTmp & SISrcMods::NEG) != 0) {
+        if ((ModsTmp & SISrcMods::NEG) != 0)
           Mods ^= SISrcMods::NEG;
-          Src = PeekSrc;
-        }
 
-        if ((ModsTmp & SISrcMods::ABS) != 0) {
+        if ((ModsTmp & SISrcMods::ABS) != 0)
           Mods |= SISrcMods::ABS;
-          Src = PeekSrc;
-        }
       }
     };
 
@@ -5344,8 +5338,7 @@ AMDGPUInstructionSelector::selectVOP3PMadMixModsImpl(MachineOperand &Root,
 
     Mods |= SISrcMods::OP_SEL_1;
 
-    if (isExtractHiElt(*MRI, PeekSrc, PeekSrc)) {
-      Src = PeekSrc;
+    if (isExtractHiElt(*MRI, Src, Src)) {
       Mods |= SISrcMods::OP_SEL_0;
       CheckAbsNeg();
     }
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/combine-fma-add-ext-mul.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/combine-fma-add-ext-mul.ll
index 4ebe1c499a1769..4d603f7487754a 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/combine-fma-add-ext-mul.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/combine-fma-add-ext-mul.ll
@@ -73,10 +73,14 @@ define amdgpu_vs <5 x float> @test_5xf16_5xf32_add_ext_mul(<5 x half> inreg %x,
 ; GFX10-FAST-DENORM-NEXT:    v_mov_b32_e32 v2, s8
 ; GFX10-FAST-DENORM-NEXT:    v_mov_b32_e32 v3, s9
 ; GFX10-FAST-DENORM-NEXT:    v_mov_b32_e32 v4, s10
+; GFX10-FAST-DENORM-NEXT:    s_lshr_b32 s11, s0, 16
+; GFX10-FAST-DENORM-NEXT:    s_lshr_b32 s12, s1, 16
+; GFX10-FAST-DENORM-NEXT:    s_lshr_b32 s13, s3, 16
+; GFX10-FAST-DENORM-NEXT:    s_lshr_b32 s14, s4, 16
 ; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v0, s0, s3, v0 op_sel_hi:[1,1,0]
-; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v1, s0, s3, v1 op_sel:[1,1,0] op_sel_hi:[1,1,0]
+; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v1, s11, s13, v1 op_sel_hi:[1,1,0]
 ; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v2, s1, s4, v2 op_sel_hi:[1,1,0]
-; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v3, s1, s4, v3 op_sel:[1,1,0] op_sel_hi:[1,1,0]
+; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v3, s12, s14, v3 op_sel_hi:[1,1,0]
 ; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v4, s2, s5, v4 op_sel_hi:[1,1,0]
 ; GFX10-FAST-DENORM-NEXT:    ; return to shader part epilog
 .entry:
@@ -117,12 +121,18 @@ define amdgpu_vs <6 x float> @test_6xf16_6xf32_add_ext_mul_rhs(<6 x half> inreg
 ; GFX10-FAST-DENORM-NEXT:    v_mov_b32_e32 v3, s9
 ; GFX10-FAST-DENORM-NEXT:    v_mov_b32_e32 v4, s10
 ; GFX10-FAST-DENORM-NEXT:    v_mov_b32_e32 v5, s11
+; GFX10-FAST-DENORM-NEXT:    s_lshr_b32 s12, s0, 16
+; GFX10-FAST-DENORM-NEXT:    s_lshr_b32 s13, s1, 16
+; GFX10-FAST-DENORM-NEXT:    s_lshr_b32 s6, s2, 16
+; GFX10-FAST-DENORM-NEXT:    s_lshr_b32 s14, s3, 16
+; GFX10-FAST-DENORM-NEXT:    s_lshr_b32 s15, s4, 16
+; GFX10-FAST-DENORM-NEXT:    s_lshr_b32 s16, s5, 16
 ; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v0, s0, s3, v0 op_sel_hi:[1,1,0]
-; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v1, s0, s3, v1 op_sel:[1,1,0] op_sel_hi:[1,1,0]
+; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v1, s12, s14, v1 op_sel_hi:[1,1,0]
 ; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v2, s1, s4, v2 op_sel_hi:[1,1,0]
-; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v3, s1, s4, v3 op_sel:[1,1,0] op_sel_hi:[1,1,0]
+; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v3, s13, s15, v3 op_sel_hi:[1,1,0]
 ; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v4, s2, s5, v4 op_sel_hi:[1,1,0]
-; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v5, s2, s5, v5 op_sel:[1,1,0] op_sel_hi:[1,1,0]
+; GFX10-FAST-DENORM-NEXT:    v_fma_mix_f32 v5, s6, s16, v5 op_sel_hi:[1,1,0]
 ; GFX10-FAST-DENORM-NEXT:    ; return to shader part epilog
 .entry:
     %a = fmul fast <6 x half> %x, %y
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/fdiv.f16.ll b/llvm/test/CodeGen/AMDGPU/GlobalISel/fdiv.f16.ll
index 89cd18ad9be70b..1a98285230b2cd 100644
--- a/llvm/test/CodeGen/AMDGPU/GlobalISel/fdiv.f16.ll
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/fdiv.f16.ll
@@ -2555,9 +2555,9 @@ define amdgpu_ps i32 @s_fdiv_v2f16(i32 inreg %a.arg, i32 inreg %b.arg) {
 ; GFX9-FLUSH-NEXT:    v_rcp_f32_e32 v1, v1
 ; GFX9-FLUSH-NEXT:    v_mad_mixlo_f16 v0, s0, v0, 0 op_sel_hi:[1,0,0]
 ; GFX9-FLUSH-NEXT:    v_div_fixup_f16 v0, v0, v2, s0
-; GFX9-FLUSH-NEXT:    v_mad_mixlo_f16 v1, s0, v1, 0 op_sel:[1,0,0] op_sel_hi:[1,0,0]
-; GFX9-FLUSH-NEXT:    v_mov_b32_e32 v2, s3
-; GFX9-FLUSH-NEXT:    v_div_fixup_f16 v1, v1, s2, v2
+; GFX9-FLUSH-NEXT:    v_mov_b32_e32 v2, s2
+; GFX9-FLUSH-NEXT:    v_mad_mixlo_f16 v1, s3, v1, 0 op_sel_hi:[1,0,0]
+; GFX9-FLUSH-NEXT:    v_div_fixup_f16 v1, v1, v2, s3
 ; GFX9-FLUSH-NEXT:    v_pack_b32_f16 v0, v0, v1
 ; GFX9-FLUSH-NEXT:    v_readfirstlane_b32 s0, v0
 ; GFX9-FLUSH-NEXT:    ; return to shader part epilog
@@ -2571,7 +2571,7 @@ define amdgpu_ps i32 @s_fdiv_v2f16(i32 inreg %a.arg, i32 inreg %b.arg) {
 ; GFX10-NEXT:    v_rcp_f32_e32 v0, v0
 ; GFX10-NEXT:    v_rcp_f32_e32 v1, v1
 ; GFX10-NEXT:    v_fma_mixlo_f16 v0, s0, v0, 0 op_sel_hi:[1,0,0]
-; GFX10-NEXT:    v_fma_mixlo_f16 v1, s0, v1, 0 op_sel:[1,0,0] op_sel_hi:[1,0,0]
+; GFX10-NEXT:    v_fma_mixlo_f16 v1, s3, v1, 0 op_sel_hi:[1,0,0]
 ; GFX10-NEXT:    v_div_fixup_f16 v0, v0, s1, s0
 ; GFX10-NEXT:    v_div_fixup_f16 v1, v1, s2, s3
 ; GFX10-NEXT:    v_pack_b32_f16 v0, v0, v1
@@ -2588,7 +2588,7 @@ define amdgpu_ps i32 @s_fdiv_v2f16(i32 inreg %a.arg, i32 inreg %b.arg) {
 ; GFX11-NEXT:    v_rcp_f32_e32 v1, v1
 ; GFX11-NEXT:    s_waitcnt_depctr 0xfff
 ; GFX11-NEXT:    v_fma_mixlo_f16 v0, s0, v0, 0 op_sel_hi:[1,0,0]
-; GFX11-NEXT:    v_fma_mixlo_f16 v1, s0, v1, 0 op_sel:[1,0,0] op_sel_hi:[1,0,0]
+; GFX11-NEXT:    v_fma_mixlo_f16 v1, s3, v1, 0 op_sel_hi:[1,0,0]
 ; GFX11-NEXT:    v_div_fixup_f16 v0, v0, s1, s0
 ; GFX11-NEXT:    v_div_fixup_f16 v1, v1, s2, s3
 ; GFX11-NEXT:    v_pack_b32_f16 v0, v0, v1
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/fmamix-constant-bus-violation.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/fmamix-constant-bus-violation.mir
new file mode 100644
index 00000000000000..34f9edd284f2ce
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/fmamix-constant-bus-violation.mir
@@ -0,0 +1,42 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx908 -run-pass=instruction-select,machineverifier -o - %s | FileCheck -check-prefixes=GFX9 %s
+
+---
+name: foo
+legalized: true
+regBankSelected: true
+machineFunctionInfo:
+  mode:
+    fp32-output-denormals: false
+    fp32-input-denormals: false
+body: |
+  bb.0:
+    ; GFX9-LABEL: name: foo
+    ; GFX9: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+    ; GFX9-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+    ; GFX9-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 16
+    ; GFX9-NEXT: [[S_LSHR_B32_:%[0-9]+]]:sreg_32 = S_LSHR_B32 [[COPY]], [[S_MOV_B32_]], implicit-def dead $scc
+    ; GFX9-NEXT: [[S_LSHR_B32_1:%[0-9]+]]:sreg_32 = S_LSHR_B32 [[COPY1]], [[S_MOV_B32_]], implicit-def dead $scc
+    ; GFX9-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[S_LSHR_B32_]]
+    ; GFX9-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[S_LSHR_B32_1]]
+    ; GFX9-NEXT: [[V_FMA_MIX_F32_:%[0-9]+]]:vgpr_32 = nofpexcept V_FMA_MIX_F32 9, [[COPY3]], 8, [[COPY3]], 8, [[COPY2]], 0, 0, 0, implicit $mode, implicit $exec
+    ; GFX9-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[V_FMA_MIX_F32_]], implicit $exec
+    ; GFX9-NEXT: $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+    ; GFX9-NEXT: SI_RETURN_TO_EPILOG implicit $sgpr0
+    %0:sgpr(s32) = COPY $sgpr0
+    %1:sgpr(s32) = COPY $sgpr1
+    %2:sgpr(s32) = G_CONSTANT i32 16
+    %3:sgpr(s32) = G_LSHR %0:sgpr, %2:sgpr(s32)
+    %4:sgpr(s16) = G_TRUNC %3:sgpr(s32)
+    %5:sgpr(s32) = G_LSHR %1:sgpr, %2:sgpr(s32)
+    %6:sgpr(s16) = G_TRUNC %5:sgpr(s32)
+    %7:vgpr(s16) = COPY %4:sgpr(s16)
+    %8:vgpr(s32) = G_FPEXT %7:vgpr(s16)
+    %9:vgpr(s16) = COPY %6:sgpr(s16)
+    %10:vgpr(s32) = G_FPEXT %9:vgpr(s16)
+    %11:vgpr(s32) = G_FNEG %10:vgpr
+    %12:vgpr(s32) = G_FMA %11:vgpr, %10:vgpr, %8:vgpr
+    %13:sgpr(s32) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane), %12:vgpr(s32)
+    $sgpr0 = COPY %13:sgpr(s32)
+    SI_RETURN_TO_EPILOG implicit $sgpr0
+...
diff --git a/llvm/test/CodeGen/AMDGPU/GlobalISel/madmix-constant-bus-violation.mir b/llvm/test/CodeGen/AMDGPU/GlobalISel/madmix-constant-bus-violation.mir
new file mode 100644
index 00000000000000..035685265e0b61
--- /dev/null
+++ b/llvm/test/CodeGen/AMDGPU/GlobalISel/madmix-constant-bus-violation.mir
@@ -0,0 +1,42 @@
+# NOTE: Assertions have been autogenerated by utils/update_mir_test_checks.py UTC_ARGS: --version 5
+# RUN: llc -global-isel=1 -mtriple=amdgcn -mcpu=gfx900 -run-pass=instruction-select,machineverifier -o - %s | FileCheck -check-prefixes=GFX9 %s
+
+---
+name: foo
+legalized: true
+regBankSelected: true
+machineFunctionInfo:
+  mode:
+    fp32-output-denormals: false
+    fp32-input-denormals: false
+body: |
+  bb.0:
+    ; GFX9-LABEL: name: foo
+    ; GFX9: [[COPY:%[0-9]+]]:sreg_32 = COPY $sgpr0
+    ; GFX9-NEXT: [[COPY1:%[0-9]+]]:sreg_32 = COPY $sgpr1
+    ; GFX9-NEXT: [[S_MOV_B32_:%[0-9]+]]:sreg_32 = S_MOV_B32 16
+    ; GFX9-NEXT: [[S_LSHR_B32_:%[0-9]+]]:sreg_32 = S_LSHR_B32 [[COPY]], [[S_MOV_B32_]], implicit-def dead $scc
+    ; GFX9-NEXT: [[S_LSHR_B32_1:%[0-9]+]]:sreg_32 = S_LSHR_B32 [[COPY1]], [[S_MOV_B32_]], implicit-def dead $scc
+    ; GFX9-NEXT: [[COPY2:%[0-9]+]]:vgpr_32 = COPY [[S_LSHR_B32_]]
+    ; GFX9-NEXT: [[COPY3:%[0-9]+]]:vgpr_32 = COPY [[S_LSHR_B32_1]]
+    ; GFX9-NEXT: [[V_MAD_MIX_F32_:%[0-9]+]]:vgpr_32 = V_MAD_MIX_F32 9, [[COPY3]], 8, [[COPY3]], 8, [[COPY2]], 0, 0, 0, implicit $mode, implicit $exec
+    ; GFX9-NEXT: [[V_READFIRSTLANE_B32_:%[0-9]+]]:sreg_32 = V_READFIRSTLANE_B32 [[V_MAD_MIX_F32_]], implicit $exec
+    ; GFX9-NEXT: $sgpr0 = COPY [[V_READFIRSTLANE_B32_]]
+    ; GFX9-NEXT: SI_RETURN_TO_EPILOG implicit $sgpr0
+    %0:sgpr(s32) = COPY $sgpr0
+    %1:sgpr(s32) = COPY $sgpr1
+    %2:sgpr(s32) = G_CONSTANT i32 16
+    %3:sgpr(s32) = G_LSHR %0:sgpr, %2:sgpr(s32)
+    %4:sgpr(s16) = G_TRUNC %3:sgpr(s32)
+    %5:sgpr(s32) = G_LSHR %1:sgpr, %2:sgpr(s32)
+    %6:sgpr(s16) = G_TRUNC %5:sgpr(s32)
+    %7:vgpr(s16) = COPY %4:sgpr(s16)
+    %8:vgpr(s32) = G_FPEXT %7:vgpr(s16)
+    %9:vgpr(s16) = COPY %6:sgpr(s16)
+    %10:vgpr(s32) = G_FPEXT %9:vgpr(s16)
+    %11:vgpr(s32) = G_FNEG %10:vgpr
+    %12:vgpr(s32) = G_FMAD %11:vgpr, %10:vgpr, %8:vgpr
+    %13:sgpr(s32) = G_INTRINSIC_CONVERGENT intrinsic(@llvm.amdgcn.readfirstlane), %12:vgpr(s32)
+    $sgpr0 = COPY %13:sgpr(s32)
+    SI_RETURN_TO_EPILOG implicit $sgpr0
+...



More information about the llvm-commits mailing list