https://github.com/chsigg updated https://github.com/llvm/llvm-project/pull/231429
>From f782788f6d306cf3495ef915b6ad3d88a7aa3df3 Mon Sep 17 00:00:00 2001 From: Christian Sigg <[email protected]> Date: Tue, 13 Oct 2026 11:42:50 +0200 Subject: [PATCH] [NVPTX] Set default value of nvptx-allow-ftz-atomics to false This restores the strict FTZ default originally introduced in #200732 and temporarily relaxed in #206154. Frontends can now explicitly opt in to native FTZ-mismatched atomics on a per-instruction basis using !atomic.ignore.denormal.mode metadata (#217585, #217586), which is supported in Clang, MLIR, and XLA. Users who need the previous behavior globally can still pass -nvptx-allow-ftz-atomics=true. Update test cases to verify the new default disallow behavior. --- llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp | 11 +-- .../NVPTX/atomicrmw-allow-ftz-atomics.ll | 8 +- .../NVPTX/atomicrmw-elementwise-sm60.ll | 4 +- .../NVPTX/atomicrmw-elementwise-sm70.ll | 4 +- .../NVPTX/atomicrmw-elementwise-sm90.ll | 4 +- .../NVPTX/atomicrmw-ignore-denormal-mode.ll | 73 +++++++++---------- llvm/test/CodeGen/NVPTX/atomicrmw-sm60.ll | 4 +- llvm/test/CodeGen/NVPTX/atomicrmw-sm70.ll | 4 +- llvm/test/CodeGen/NVPTX/atomicrmw-sm90.ll | 4 +- llvm/test/CodeGen/NVPTX/atomicrmw.py | 4 +- 10 files changed, 60 insertions(+), 60 deletions(-) diff --git a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp index 58340d62bb6fea..537850b883b9eb 100644 --- a/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp +++ b/llvm/lib/Target/NVPTX/NVPTXISelLowering.cpp @@ -112,15 +112,16 @@ static cl::opt<bool> UsePrecSqrtF32( cl::init(true)); // PTX atom.add.f32 has fixed FTZ behavior that may not match the function's -// (see shouldExpandAtomicRMWInIR), so we'd normally fall back to a CAS loop -// when they disagree. This option (enabled by default) allows using atom.add -// anyway, trading correct denormal handling for the speed of the native -// instruction. +// (see shouldExpandAtomicRMWInIR), so by default we fall back to a CAS loop +// when they disagree (strict mode), unless declared insignificant via +// !atomic.ignore.denormal.mode metadata or this option. This option allows +// using atom.add anyway, trading correct denormal handling for the speed of the +// native instruction. static cl::opt<bool> AllowFTZAtomics( "nvptx-allow-ftz-atomics", cl::Hidden, cl::desc("NVPTX Specific: Lower atomicrmw fadd to atom.add even when its " "FTZ behavior does not match the function's denormal mode."), - cl::init(true)); + cl::init(false)); /// Whereas CUDA's implementation (see libdevice) uses ex2.approx for exp2(), it /// does NOT use lg2.approx for log2, so this is disabled by default. diff --git a/llvm/test/CodeGen/NVPTX/atomicrmw-allow-ftz-atomics.ll b/llvm/test/CodeGen/NVPTX/atomicrmw-allow-ftz-atomics.ll index 78c86bfb08e2fa..e78f2428bbde98 100644 --- a/llvm/test/CodeGen/NVPTX/atomicrmw-allow-ftz-atomics.ll +++ b/llvm/test/CodeGen/NVPTX/atomicrmw-allow-ftz-atomics.ll @@ -2,16 +2,16 @@ ; -nvptx-allow-ftz-atomics forces atomicrmw fadd to lower to atom.add even when ; the hardware's FTZ behavior doesn't match the function's denormal mode, where ; we'd otherwise emit a CAS loop. Compare the lowering with the option disabled -; (DISALLOW) against the option enabled or default (ALLOW). +; or default (DISALLOW) against the option enabled (ALLOW). ; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=CHECK,DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=CHECK,ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 | FileCheck %s --check-prefixes=CHECK,ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 | FileCheck %s --check-prefixes=CHECK,DISALLOW ; RUN: %if ptxas-sm_90 && ptxas-isa-8.7 %{ llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=false | %ptxas-verify -arch=sm_90 %} ; RUN: %if ptxas-sm_90 && ptxas-isa-8.7 %{ llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=true | %ptxas-verify -arch=sm_90 %} ; RUN: %if ptxas-sm_90 && ptxas-isa-8.7 %{ llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 | %ptxas-verify -arch=sm_90 %} ; global f32: atom.add flushes, so we'd normally use a CAS loop (function is not FTZ) -; when the option is disabled; when enabled (default), it lowers to atom.add. +; when the option is disabled (default); when enabled, it lowers to atom.add. define float @fadd_f32_global(ptr addrspace(1) %addr, float %val) { ; ; DISALLOW-LABEL: fadd_f32_global( @@ -51,7 +51,7 @@ define float @fadd_f32_global(ptr addrspace(1) %addr, float %val) { } ; generic f32: we can't know the memory space, so we'd normally use a CAS loop -; when the option is disabled; when enabled (default), it lowers to atom.add. +; when the option is disabled (default); when enabled, it lowers to atom.add. define float @fadd_f32_generic(ptr %addr, float %val) { ; ; DISALLOW-LABEL: fadd_f32_generic( diff --git a/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm60.ll b/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm60.ll index 5b631504b96257..a1d20e8e821f03 100644 --- a/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm60.ll +++ b/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm60.ll @@ -1,10 +1,10 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 ; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM60,SM60-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM60,SM60-NOFTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 | FileCheck %s --check-prefixes=SM60,SM60-NOFTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 | FileCheck %s --check-prefixes=SM60,SM60-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM60,SM60-FTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM60,SM60-FTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM60,SM60-FTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM60,SM60-FTZ-DISALLOW ; RUN: %if ptxas-sm_60 && ptxas-isa-5.0 %{ llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -nvptx-allow-ftz-atomics=false | %ptxas-verify -arch=sm_60 %} ; RUN: %if ptxas-sm_60 && ptxas-isa-5.0 %{ llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -nvptx-allow-ftz-atomics=true | %ptxas-verify -arch=sm_60 %} ; RUN: %if ptxas-sm_60 && ptxas-isa-5.0 %{ llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 | %ptxas-verify -arch=sm_60 %} diff --git a/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm70.ll b/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm70.ll index c5cebdf5946657..f1c34e825077d5 100644 --- a/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm70.ll +++ b/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm70.ll @@ -1,10 +1,10 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 ; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM70,SM70-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM70,SM70-NOFTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s --check-prefixes=SM70,SM70-NOFTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s --check-prefixes=SM70,SM70-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM70,SM70-FTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM70,SM70-FTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM70,SM70-FTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM70,SM70-FTZ-DISALLOW ; RUN: %if ptxas-sm_70 && ptxas-isa-6.3 %{ llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -nvptx-allow-ftz-atomics=false | %ptxas-verify -arch=sm_70 %} ; RUN: %if ptxas-sm_70 && ptxas-isa-6.3 %{ llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -nvptx-allow-ftz-atomics=true | %ptxas-verify -arch=sm_70 %} ; RUN: %if ptxas-sm_70 && ptxas-isa-6.3 %{ llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | %ptxas-verify -arch=sm_70 %} diff --git a/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm90.ll b/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm90.ll index 19bd7eab664888..613169d23e0099 100644 --- a/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm90.ll +++ b/llvm/test/CodeGen/NVPTX/atomicrmw-elementwise-sm90.ll @@ -1,10 +1,10 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 ; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM90,SM90-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM90,SM90-NOFTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 | FileCheck %s --check-prefixes=SM90,SM90-NOFTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 | FileCheck %s --check-prefixes=SM90,SM90-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM90,SM90-FTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM90,SM90-FTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM90,SM90-FTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM90,SM90-FTZ-DISALLOW ; RUN: %if ptxas-sm_90 && ptxas-isa-8.7 %{ llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=false | %ptxas-verify -arch=sm_90 %} ; RUN: %if ptxas-sm_90 && ptxas-isa-8.7 %{ llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=true | %ptxas-verify -arch=sm_90 %} ; RUN: %if ptxas-sm_90 && ptxas-isa-8.7 %{ llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 | %ptxas-verify -arch=sm_90 %} diff --git a/llvm/test/CodeGen/NVPTX/atomicrmw-ignore-denormal-mode.ll b/llvm/test/CodeGen/NVPTX/atomicrmw-ignore-denormal-mode.ll index f3cf67199e8b91..212688e5d8ccf1 100644 --- a/llvm/test/CodeGen/NVPTX/atomicrmw-ignore-denormal-mode.ll +++ b/llvm/test/CodeGen/NVPTX/atomicrmw-ignore-denormal-mode.ll @@ -1,6 +1,7 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 ; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=CHECK,STRICT -; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 | FileCheck %s --check-prefixes=CHECK,DEFAULT +; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=CHECK,ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 | FileCheck %s --check-prefixes=CHECK,STRICT ; RUN: %if ptxas-sm_90 && ptxas-isa-8.7 %{ llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=false | %ptxas-verify -arch=sm_90 %} ; PTX atom.add has a fixed FTZ behavior that need not match the function's @@ -13,9 +14,7 @@ ; atom.add.ftz.f16) ; When the two disagree the atomic is expanded into a CAS loop, unless the ; mismatch is explicitly declared insignificant -- either globally via -; -nvptx-allow-ftz-atomics (on by default, hence the STRICT run lines above, -; without which the metadata would never be the deciding factor) or per -; instruction via !atomic.ignore.denormal.mode. +; -nvptx-allow-ftz-atomics or per instruction via !atomic.ignore.denormal.mode. ; f32, global: the native instruction flushes, so an IEEE function needs the ; metadata to avoid the CAS loop. @@ -42,17 +41,17 @@ define float @fadd_f32_global_ieee_no_md(ptr addrspace(1) %addr, float %val) { ; STRICT-NEXT: st.param::func.b32 [func_retval0], %r1; ; STRICT-NEXT: ret; ; -; DEFAULT-LABEL: fadd_f32_global_ieee_no_md( -; DEFAULT: { -; DEFAULT-NEXT: .reg .b32 %r<3>; -; DEFAULT-NEXT: .reg .b64 %rd<2>; -; DEFAULT-EMPTY: -; DEFAULT-NEXT: // %bb.0: -; DEFAULT-NEXT: ld.param::func.b64 %rd1, [fadd_f32_global_ieee_no_md_param_0]; -; DEFAULT-NEXT: ld.param::func.b32 %r1, [fadd_f32_global_ieee_no_md_param_1]; -; DEFAULT-NEXT: atom.relaxed.sys.global.add.f32 %r2, [%rd1], %r1; -; DEFAULT-NEXT: st.param::func.b32 [func_retval0], %r2; -; DEFAULT-NEXT: ret; +; ALLOW-LABEL: fadd_f32_global_ieee_no_md( +; ALLOW: { +; ALLOW-NEXT: .reg .b32 %r<3>; +; ALLOW-NEXT: .reg .b64 %rd<2>; +; ALLOW-EMPTY: +; ALLOW-NEXT: // %bb.0: +; ALLOW-NEXT: ld.param::func.b64 %rd1, [fadd_f32_global_ieee_no_md_param_0]; +; ALLOW-NEXT: ld.param::func.b32 %r1, [fadd_f32_global_ieee_no_md_param_1]; +; ALLOW-NEXT: atom.relaxed.sys.global.add.f32 %r2, [%rd1], %r1; +; ALLOW-NEXT: st.param::func.b32 [func_retval0], %r2; +; ALLOW-NEXT: ret; %r = atomicrmw fadd ptr addrspace(1) %addr, float %val monotonic ret float %r } @@ -116,17 +115,17 @@ define float @fadd_f32_generic_ieee_no_md(ptr %addr, float %val) { ; STRICT-NEXT: st.param::func.b32 [func_retval0], %r1; ; STRICT-NEXT: ret; ; -; DEFAULT-LABEL: fadd_f32_generic_ieee_no_md( -; DEFAULT: { -; DEFAULT-NEXT: .reg .b32 %r<3>; -; DEFAULT-NEXT: .reg .b64 %rd<2>; -; DEFAULT-EMPTY: -; DEFAULT-NEXT: // %bb.0: -; DEFAULT-NEXT: ld.param::func.b64 %rd1, [fadd_f32_generic_ieee_no_md_param_0]; -; DEFAULT-NEXT: ld.param::func.b32 %r1, [fadd_f32_generic_ieee_no_md_param_1]; -; DEFAULT-NEXT: atom.relaxed.sys.add.f32 %r2, [%rd1], %r1; -; DEFAULT-NEXT: st.param::func.b32 [func_retval0], %r2; -; DEFAULT-NEXT: ret; +; ALLOW-LABEL: fadd_f32_generic_ieee_no_md( +; ALLOW: { +; ALLOW-NEXT: .reg .b32 %r<3>; +; ALLOW-NEXT: .reg .b64 %rd<2>; +; ALLOW-EMPTY: +; ALLOW-NEXT: // %bb.0: +; ALLOW-NEXT: ld.param::func.b64 %rd1, [fadd_f32_generic_ieee_no_md_param_0]; +; ALLOW-NEXT: ld.param::func.b32 %r1, [fadd_f32_generic_ieee_no_md_param_1]; +; ALLOW-NEXT: atom.relaxed.sys.add.f32 %r2, [%rd1], %r1; +; ALLOW-NEXT: st.param::func.b32 [func_retval0], %r2; +; ALLOW-NEXT: ret; %r = atomicrmw fadd ptr %addr, float %val monotonic ret float %r } @@ -203,17 +202,17 @@ define half @fadd_f16_global_ftz_no_md(ptr addrspace(1) %addr, half %val) #1 { ; STRICT-NEXT: st.param::func.b16 [func_retval0], %r12; ; STRICT-NEXT: ret; ; -; DEFAULT-LABEL: fadd_f16_global_ftz_no_md( -; DEFAULT: { -; DEFAULT-NEXT: .reg .b16 %rs<3>; -; DEFAULT-NEXT: .reg .b64 %rd<2>; -; DEFAULT-EMPTY: -; DEFAULT-NEXT: // %bb.0: -; DEFAULT-NEXT: ld.param::func.b64 %rd1, [fadd_f16_global_ftz_no_md_param_0]; -; DEFAULT-NEXT: ld.param::func.b16 %rs1, [fadd_f16_global_ftz_no_md_param_1]; -; DEFAULT-NEXT: atom.relaxed.sys.global.add.noftz.f16 %rs2, [%rd1], %rs1; -; DEFAULT-NEXT: st.param::func.b16 [func_retval0], %rs2; -; DEFAULT-NEXT: ret; +; ALLOW-LABEL: fadd_f16_global_ftz_no_md( +; ALLOW: { +; ALLOW-NEXT: .reg .b16 %rs<3>; +; ALLOW-NEXT: .reg .b64 %rd<2>; +; ALLOW-EMPTY: +; ALLOW-NEXT: // %bb.0: +; ALLOW-NEXT: ld.param::func.b64 %rd1, [fadd_f16_global_ftz_no_md_param_0]; +; ALLOW-NEXT: ld.param::func.b16 %rs1, [fadd_f16_global_ftz_no_md_param_1]; +; ALLOW-NEXT: atom.relaxed.sys.global.add.noftz.f16 %rs2, [%rd1], %rs1; +; ALLOW-NEXT: st.param::func.b16 [func_retval0], %rs2; +; ALLOW-NEXT: ret; %r = atomicrmw fadd ptr addrspace(1) %addr, half %val monotonic ret half %r } diff --git a/llvm/test/CodeGen/NVPTX/atomicrmw-sm60.ll b/llvm/test/CodeGen/NVPTX/atomicrmw-sm60.ll index fedb1ba54f8a8a..85d7f8f2f52433 100644 --- a/llvm/test/CodeGen/NVPTX/atomicrmw-sm60.ll +++ b/llvm/test/CodeGen/NVPTX/atomicrmw-sm60.ll @@ -1,10 +1,10 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 ; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM60,SM60-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM60,SM60-NOFTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 | FileCheck %s --check-prefixes=SM60,SM60-NOFTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 | FileCheck %s --check-prefixes=SM60,SM60-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM60,SM60-FTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM60,SM60-FTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM60,SM60-FTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM60,SM60-FTZ-DISALLOW ; RUN: %if ptxas-sm_60 && ptxas-isa-5.0 %{ llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -nvptx-allow-ftz-atomics=false | %ptxas-verify -arch=sm_60 %} ; RUN: %if ptxas-sm_60 && ptxas-isa-5.0 %{ llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 -nvptx-allow-ftz-atomics=true | %ptxas-verify -arch=sm_60 %} ; RUN: %if ptxas-sm_60 && ptxas-isa-5.0 %{ llc < %s -march=nvptx64 -mcpu=sm_60 -mattr=+ptx50 | %ptxas-verify -arch=sm_60 %} diff --git a/llvm/test/CodeGen/NVPTX/atomicrmw-sm70.ll b/llvm/test/CodeGen/NVPTX/atomicrmw-sm70.ll index 12ba9b5fb4f8ca..e40d78b4effaa9 100644 --- a/llvm/test/CodeGen/NVPTX/atomicrmw-sm70.ll +++ b/llvm/test/CodeGen/NVPTX/atomicrmw-sm70.ll @@ -1,10 +1,10 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 ; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM70,SM70-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM70,SM70-NOFTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s --check-prefixes=SM70,SM70-NOFTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | FileCheck %s --check-prefixes=SM70,SM70-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM70,SM70-FTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM70,SM70-FTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM70,SM70-FTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM70,SM70-FTZ-DISALLOW ; RUN: %if ptxas-sm_70 && ptxas-isa-6.3 %{ llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -nvptx-allow-ftz-atomics=false | %ptxas-verify -arch=sm_70 %} ; RUN: %if ptxas-sm_70 && ptxas-isa-6.3 %{ llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 -nvptx-allow-ftz-atomics=true | %ptxas-verify -arch=sm_70 %} ; RUN: %if ptxas-sm_70 && ptxas-isa-6.3 %{ llc < %s -march=nvptx64 -mcpu=sm_70 -mattr=+ptx63 | %ptxas-verify -arch=sm_70 %} diff --git a/llvm/test/CodeGen/NVPTX/atomicrmw-sm90.ll b/llvm/test/CodeGen/NVPTX/atomicrmw-sm90.ll index ab31e4a297bfad..599d775d45da72 100644 --- a/llvm/test/CodeGen/NVPTX/atomicrmw-sm90.ll +++ b/llvm/test/CodeGen/NVPTX/atomicrmw-sm90.ll @@ -1,10 +1,10 @@ ; NOTE: Assertions have been autogenerated by utils/update_llc_test_checks.py UTC_ARGS: --version 6 ; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM90,SM90-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM90,SM90-NOFTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 | FileCheck %s --check-prefixes=SM90,SM90-NOFTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 | FileCheck %s --check-prefixes=SM90,SM90-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM90,SM90-FTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM90,SM90-FTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM90,SM90-FTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM90,SM90-FTZ-DISALLOW ; RUN: %if ptxas-sm_90 && ptxas-isa-8.7 %{ llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=false | %ptxas-verify -arch=sm_90 %} ; RUN: %if ptxas-sm_90 && ptxas-isa-8.7 %{ llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 -nvptx-allow-ftz-atomics=true | %ptxas-verify -arch=sm_90 %} ; RUN: %if ptxas-sm_90 && ptxas-isa-8.7 %{ llc < %s -march=nvptx64 -mcpu=sm_90 -mattr=+ptx87 | %ptxas-verify -arch=sm_90 %} diff --git a/llvm/test/CodeGen/NVPTX/atomicrmw.py b/llvm/test/CodeGen/NVPTX/atomicrmw.py index 05388afa53578e..832020d0e09795 100644 --- a/llvm/test/CodeGen/NVPTX/atomicrmw.py +++ b/llvm/test/CodeGen/NVPTX/atomicrmw.py @@ -80,10 +80,10 @@ run_statement = Template( """; RUN: llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM${sm},SM${sm}-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM${sm},SM${sm}-NOFTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} | FileCheck %s --check-prefixes=SM${sm},SM${sm}-NOFTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} | FileCheck %s --check-prefixes=SM${sm},SM${sm}-NOFTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=false | FileCheck %s --check-prefixes=SM${sm},SM${sm}-FTZ-DISALLOW ; RUN: llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} -denormal-fp-math-f32=preserve-sign -nvptx-allow-ftz-atomics=true | FileCheck %s --check-prefixes=SM${sm},SM${sm}-FTZ-ALLOW -; RUN: llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM${sm},SM${sm}-FTZ-ALLOW +; RUN: llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} -denormal-fp-math-f32=preserve-sign | FileCheck %s --check-prefixes=SM${sm},SM${sm}-FTZ-DISALLOW ; RUN: %if ptxas-sm_${sm} && ptxas-isa-${ptxfp} %{ llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} -nvptx-allow-ftz-atomics=false | %ptxas-verify -arch=sm_${sm} %} ; RUN: %if ptxas-sm_${sm} && ptxas-isa-${ptxfp} %{ llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} -nvptx-allow-ftz-atomics=true | %ptxas-verify -arch=sm_${sm} %} ; RUN: %if ptxas-sm_${sm} && ptxas-isa-${ptxfp} %{ llc < %s -march=nvptx64 -mcpu=sm_${sm} -mattr=+ptx${ptx} | %ptxas-verify -arch=sm_${sm} %} _______________________________________________ llvm-branch-commits mailing list [email protected] https://lists.llvm.org/cgi-bin/mailman/listinfo/llvm-branch-commits
